Larissa Conradt, Christian List - Group Decision Making in Humans and Animals (Philosophical Transactions of The Royal Society Series B)

RSTB_364_1518.
qxp 1/27/09 8:21 PM Page 1
Phil. Trans. R. Soc. B | vol. 364 no. 1518 pp. 717–852 | 27 Mar 2009
ISSN 0962-8436
volume 364
27 March 2009 number 1518

volume 364 . number 1518 . pages 717–852
pages 717–852
Group decision making in humans and
animals In this issue
Papers of a Theme Issue compiled and edited by Larissa Conradt and Christian List Group decision making
Introduction
Group decisions in humans and animals: a survey 719
in humans and animals
L. Conradt & C. List Papers of a Theme Issue compiled and edited by
Articles Larissa Conradt and Christian List
Quorum responses and consensus decision making 743
D. J. T. Sumpter & S. C. Pratt
Independence and interdependence in collective decision making: an agent-based model of
nest-site choice by honeybee swarms 755
C. List, C. Elsholtz & T. D. Seeley
Information aggregation and communication in committees 763
D. Austen-Smith & T. J. Feddersen
Evolution of signalling systems with multiple senders and receivers 771
B. Skyrms
Leadership, consensus decision making and collective behaviour in humans 781
Group decision making in humans and animals

J. R. G. Dyer, A. Johansson, D. Helbing, I. D. Couzin & J. Krause
Reciprocity, culture and human cooperation: previous insights and a new cross-cultural experiment 791
S. Gächter & B. Herrmann
Conflicts of interest and the evolution of decision sharing 807
L. Conradt & T. J. Roper
Voting patterns and alliance formation in the European Parliament 821
S. Hix, A. Noury & G. Roland
Behavioural social choice: a status report 833
M. Regenwetter, B. Grofman, A. Popova, W. Messner, C. P. Davis-Stober & D. R. Cavagnaro
Speed versus accuracy in decision-making ants: expediting politics and policy implementation 845
N. R. Franks, F.-X. Dechaume-Moncharmont, E. Hanmore & J. K. Reynolds
The world’s longest running science journal

Founded in 1660, the Royal Society
is the independent scientific academy
of the UK, dedicated to promoting
excellence in science
rstb.royalsocietypublishing.org
Registered Charity No 207043 Published in Great Britain by the Royal Society, 27 March 2009
6–9 Carlton House Terrace, London SW1Y 5AG
RSTB_364_1518.qxp 1/27/09 8:21 PM Page 2
GUIDANCE FOR AUTHORS
Editor
Professor Georgina Mace Selection criteria to satisfy most non-specialist readers. Supplementary
The criteria for selection are scientific excellence, data up to 10Mb is placed on the Society's website free
Publishing Editor originality and interest across disciplines within biology. of charge. Larger datasets must be deposited in
Claire Rawlinson The Editors are responsible for all editorial decisions and recognised public domain databases by the author.
they make these decisions based on the reports received
Editorial Board from the referees and/or Editorial Board members. Many
more good proposals and articles are submitted to us Conditions of publication
Neuroscience and Cognition Professor Anthony Pawson
than we have space to print, we give preference to Articles must not have been published previously, nor be
Dr Brian Billups Dr Stephen Tucker
those that are of broad interest and of high scientific under consideration for publication elsewhere. The main
Dr Andrew Glennerster Organismal, environmental and evolutionary
Professor Bill Harris biology quality. findings of the article should not have been reported in
Professor Trevor Lamb Dr Will Cresswell the mass media. Like many journals, Phil. Trans. R. Soc. B
Professor Obaid Siddiqi Professor Georgina Mace Publishing format employs a strict embargo policy where the reporting of
Professor Andrew Whiten Professor Yadvinder Malhi
Phil. Trans. R. Soc. B articles are published regularly online a scientific article by the media is embargoed until a
Professor Manfred Milinski
Cell and developmental biology Professor Peter Mumby and in print issues twice a month. Along with all Royal specific time. The Executive Editor has final authority in
Professor Julie Ahringer Professor Peter Raven Society journals, we are committed to archiving and all matters relating to publication.
Dr Buzz Baum Professor Karl Sigmund providing perpetual access. The journal also offers the
Dr Louise Cramer Professor Monkombu Swaminathan
facility for including Electronic Supplementary Material
Dr Anne Donaldson Health and Disease
Professor Richard Gardner (ESM) to papers. Contents of the ESM might include Electronic Submission details
Professor Zhu Chen
Professor John Gray Professor Mark Enright details of methods, derivations of equations, large tables For full submission guidelines and access to all journal
Professor Keith Gull Professor Peter Goodfellow of data, DNA sequences and computer programs. content please visit the Phil. Trans. R. Soc. B website at
Professor Laurence Hurst Professor Michael Malim However, the printed version must include enough detail rstb.royalsocietypublishing.org.
Professor Fotis Kafatos Professor Lap-Chee Tsui
Professor Elliot Meyerowitz Professor Nicholas Wald
Dr Tanja van Mourik Professor Joanne Webster
Publishing Editor: Claire Rawlinson AIMS AND SCOPE
Each issue of Phil. Trans. R. Soc. B is devoted to a specific area of the biological sciences.
(tel: +44 (0)20 7451 2602; fax: +44 (0)20 7976 1837; This area will define a research frontier that is advancing rapidly, often bridging
claire.rawlinson@royalsociety.org) traditional disciplines. Phil. Trans. R. Soc. B is essential reading for scientists working The Royal Society is an independent scientific academy For further information on the Society’s activities, please
across the biological sciences. In particular, the journal is focused on the following four
cluster areas: neuroscience and cognition; organismal and evolutionary biology; founded in 1660 and self-governing under Royal contact the following departments on the extensions
cell and developmental biology; and health and disease. As well as theme issues, Charter. The Society has three roles, as the scientific listed by dialling +44 (0)20 7839 5561, or visit the
the journal publishes papers from the Royal Society’s biological discussion meetings. academy of the United Kingdom, as a learned society,
Production Editor: Jessica Mnatzaganian For information on submitting a proposal for a theme issue, consult the journal‘s
Society’s Web site (www.royalsociety.org).
website at rstb.royalsocietypublishing.org. as a funding body.
Research Support (UK grants and fellowships)
Reviews. The journal also publishes reviews in the broad areas of research listed Research appointments: 2547
above. Review articles will often be commissioned, but the Editor is happy to consider The objectives of the Royal Society are to
suggestions / proposals for unsolicited review articles. Please submit an abstract and a Research grants: 2539
6–9 Carlton House Terrace, London SW1Y 5AG, UK
covering letter to the Editorial Office for approval for consideration. 6000 words is • recognise excellence in science Conference grants: 2540
rstb.royalsocietypublishing.org usually a maximum length for reviews.
• support leading-edge scientific research and its Science Advice
ISBN: 978-0-85403-736-0 applications General enquiries: 2585
Copyright © 2008 & 2009 The Royal Society • stimulate international interaction
Science Communication
Except as otherwise permitted under the Copyright, Designs and Patents Act, 1988, this publication may only be reproduced, stored or • further the role of science, engineering General enquiries: 2572
transmitted, in any form or by any other means, with the prior permission in writing of the publisher, or in the case of reprographic and technology in society
reproduction, in accordance with the terms of a licence issued by the Copyright Licensing Agency. In particular, the Society permits the International Exchanges (for grants enabling
• promote the public’s understanding of science research visits between the UK and most other
making of a single photocopy of an article from this issue (under Sections 29 and 38 of this Act) for an individual for the purposes of
research or private study. • provide independent authoritative advice on countries (except the USA))
matters relating to science, engineering General enquiries: 2550
SUBSCRIPTIONS and technology

Library and Information Services
In 2009 Phil. Trans. R. Soc. B (ISSN 0962-8436) will be • encourage research into the history of science Library/archive enquiries: 2606
published twice a month. Full details of subscriptions and
single issue sales may be obtained either by contacting Subscription prices All other
our journal fulfilment agent, Portland Customer 2009 calendar year Europe USA & Canada countries
Services, Commerce Way, Colchester CO2 8HP; tel: printed version plus £2024/US$3845 £2136/US$4058 £2186/US$4153
+44 (0)1206 796351; fax: +44 (0)1206 799331; email: electronic access /€2631
sales@portland-services.com or by visiting our website Cover image: A colony of the ant Temnothorax curvispinosus in which every worker has been marked with a unique
at publishing.royalsociety.org/subscribers. The Royal pattern of paint drops. Individual marking makes possible a detailed analysis of the emergence of group decisions
Society is a Registered Charity No. 207043. from individual behaviour, as when these colonies collectively choose the best among several possible nest sites.
Typeset in India by the Alden Group, Oxfordshire. Printed by Latimer Trend, Plymouth. (See pages 743–753. Photograph by Stephen Pratt.)
This paper meets the requirements of ISO 9706:1994(E) and ANSI/NISO Z39.48-1992 (Permanence of Paper) effective with volume 335, issue 1273, 1992.
Philosophical Transactions of the Royal Society B (ISSN: 0962-8436) is published twice a morth by the Royal Society and distributed in the USA by DSW, 75
Aberdeen Road, Emigsville PA 17318-0437. Periodicals postage paid at Emigsville PA. POSTMASTER: send address changes to Philosophical Transactions
of the Royal Society B, c/o PO Box 437 Emigsville PA 17318-0437.
Papers of a Theme Issue compiled and edited by Larissa Conradt
and Christian List
Contents
Introduction
L. Conradt and C. List
Articles
D. J. T. Sumpter and S. C. Pratt
Independence and interdependence in collective decision making: an agent-based 755

model of nest-site choice by honeybee swarms
C. List, C. Elsholtz and T. D. Seeley

D. Austen-Smith and T. J. Feddersen

B. Skyrms

J. R. G. Dyer, A. Johansson, D. Helbing, I. D. Couzin and J. Krause
Reciprocity, culture and human cooperation: previous insights and a new 791
cross-cultural experiment
S. Gächter and B. Herrmann

L. Conradt and T. J. Roper

S. Hix, A. Noury and G. Roland

M. Regenwetter, B. Grofman, A. Popova, W. Messner, C. P. Davis-Stober
and D. R. Cavagnaro
Speed versus accuracy in decision-making ants: expediting politics and policy 845
implementation
N. R. Franks, F.-X. Dechaume-Moncharmont, E. Hanmore and J. K. Reynolds
717
Downloaded from rstb.royalsocietypublishing.org on May 16, 2010
Phil. Trans. R. Soc. B (2009) 364, 719–742

doi:10.1098/rstb.2008.0276
Published online 12 December 2008
Introduction
Group decisions in humans and animals: a survey

Larissa Conradt1,* and Christian List2
1
JMS Building, Department of Biology and Environmental Sciences, University of Sussex,
Falmer, Brighton BN1 9QR, UK
2
Department of Government, London School of Economics, London WC2A 2AE, UK
Humans routinely make many decisions collectively, whether they choose a restaurant with friends,
elect political leaders or decide actions to tackle international problems, such as climate change, that
affect the future of the whole planet. We might be less aware of it, but group decisions are just as
important to social animals as they are for us. Animal groups have to collectively decide about
communal movements, activities, nesting sites and enterprises, such as cooperative breeding or
hunting, that crucially affect their survival and reproduction. While human group decisions have
been studied for millennia, the study of animal group decisions is relatively young, but is now
expanding rapidly. It emerges that group decisions in animals pose many similar questions to those in
humans. The purpose of the present issue is to integrate and combine approaches in the social and
natural sciences in an area in which theoretical challenges and research questions are often similar,
and to introduce each discipline to the other’s key ideas, findings and successful methods. In order to
make such an introduction as effective as possible, here, we briefly review conceptual similarities and
differences between the sciences, and provide a guide to the present issue.
Keywords: collective decisions; communal decisions; conflict resolution; cooperation;
information sharing; social behaviour
1. GENERAL BACKGROUND mole rats). Group decision making is just as important

Humans usually live in highly sophisticated societies. for social animals as it is for us (see Conradt & Roper
This implies that many important decisions are made 2005 for a review). Dispersing swarms of bees and ants
not by individuals acting alone, but by groups of collectively choose new nest sites on which their
individuals acting collectively. Group decisions in survival depends (Seeley & Buhrman 1999; Visscher
humans range from small-scale decisions, such as 2007; Visscher & Seeley 2007; Franks et al. 2009).
those taken by groups of relatives, friends or colleagues, Homing and migrating birds collectively decide on
to large-scale decisions, such as nation-wide demo- communal routes that determine their chances of
cratic elections and international agreements. Clearly, survival and successful arrival (Wallraff 1978; Simons
human societies cannot function without group 2004; Biro et al. 2006). Bats collectively select roosting
decisions, and some of the most pressing problems sites that are crucial for survival (Kerth et al. 2006).
facing humanity result from failures to reach a group Swarms of insects (Buhl et al. 2006), shoals of fishes
consensus (e.g. the signing of the Kyoto treaty on (Reebs 2000; Hemelrijk & Hildenbrandt 2008; Ward
controlling greenhouse gas emissions). Group decision et al. 2008), flocks of birds (Selous 1931; Ballerini et al.
making has been a central topic in all of the social 2008), groups of carnivores (Gompper 1996), herds of
sciences for millennia (e.g. Plato: The Republic 360 ungulates (Gueron et al. 1996; Prins 1996; Conradt
BC). Nevertheless, many questions remain open, 1998; Ruckstuhl 1998; Fischhoff et al. 2007; Gautrais
particularly how conflicting interests and the sharing et al. 2007) and troops of primates (Stewart & Harcourt
of dispersed information are actually, and should be, in 1994; Trillmich et al. 2004; Meunier et al. 2006; Sellers
principle, reconciled so as to facilitate cooperation and et al. 2007; Sueur & Petit 2008) collectively decide
to reach outcomes that meet various optimality criteria. group movements and group activities with important
These are some of the fundamental questions of social fitness consequences to all individuals (Conradt &
choice theory (e.g. Arrow 1951/1963; Austen-Smith & Roper 2003; Rands et al. 2003; Dostalkova & Spinka
Banks 1999, 2005; Sen 1999; Dryzek & List 2003). 2007). Cooperative species, such as eusocial insects
A large number of animal species also live in groups and communal breeders, collectively decide job
(Krause & Ruxton 2002), some of which can be very allocation in crucial communal enterprises, such as
complex (e.g. eusocial bees, wasps, ants, termites and supplying forage to the hive (Beshers & Fewell 2001),
rearing young (Clutton-Brock 1998) and hunting prey
* Author for correspondence (l.conradt@sussex.ac.uk). (Courchamp et al. 2002). In contrast to the human
Both authors contributed equally to the study. case, the study of group decisions in social animals is
One contribution of 11 to a Theme Issue ‘Group decision making in relatively young, but is now rapidly expanding in the
humans and animals’. natural sciences (see Conradt & Roper 2005 for the
719 This journal is q 2008 The Royal Society

720 L. Conradt & C. List Introduction
most recent review). It emerges that group decisions in and natural scientists when reading cross-disciplinary
animals pose many similar questions to those in literature in the present issue and beyond.
humans, as discussed below.
3. KEY CONCEPTS FOR THE ANALYSIS OF

2. PURPOSE OF THE PRESENT ISSUE GROUP DECISIONS
The purpose of the present issue is to integrate and (a) Aggregate/consensus versus interactive/
combine approaches in the social and natural sciences combined decisions
in an area in which theoretical challenges and research Group decisions can be roughly divided into two
questions are often similar. Each discipline can benefit categories: (i) those in which the group makes a single
from being introduced to the other’s key ideas, findings collective decision, e.g. between multiple options, that is
and successful methods. Over the centuries, the social ‘binding’ in some way for all members, and (ii) those in
sciences have developed a large body of theory on which there need not be a single collectively binding
human group decisions, including many sophisticated decision, but in which individuals decide interdepen-
modelling tools, which can be modified to study animal dently with one another. In the social sciences, the
group decisions. List et al. (2009) give an example for former are often described as ‘aggregate’ or ‘collective’
this in the present issue, drawing on social-scientific decisions and are the subject of social choice theory; the
methods and ideas to develop a model of nest-site latter are described as ‘interactive’ decisions and are the
choice by honeybee (Apis mellifera) swarms. On the subject of game theory. In the natural sciences, the two
other hand, by focusing on relatively less complex categories have become known as ‘consensus’ and
group decisions, the natural sciences can concentrate ‘combined’ decisions, respectively (Conradt & Roper
on fundamental features that might also be applicable 2005; see table 1 for social- and natural-scientific
to humans but are much harder to detect in the examples and a categorization of the group decisions
sophisticated and complex contexts of human group discussed in this issue). Examples of aggregate/
decisions. A good example is Dyer et al.’s (2009) work consensus decisions are national elections, parliamen-
in the present issue. In addition, natural scientists, by tary decisions on whether to pass a new law, choices of
looking at group decisions from an evolutionary point joint movement directions in cohesive groups (e.g.
of view, can add a different approach to human group Couzin et al. 2005) and nest-site choices in eusocial
decisions from the one which most social scientists insects (e.g. Seeley & Buhrman 1999). Examples of
adopt. For example, group decision outcomes that, in interactive/combined decisions are the processes by
evolutionary terms, are ‘good’ for the individual are which many individual consumer choices lead to market
often ‘not good’ for the group, and vice versa (Conradt prices, sharing of reproductive roles in cooperative
& Roper 2003, 2007, 2009). Game theorists recognize breeders (e.g. Clutton-Brock 1998) and job allocations
such tensions, but usually cast them in terms of in honeybee workers (e.g. Beshers & Fewell 2001).
conflicts between individual rationality and group Within each of these two categories, decision
optimality rather than in evolutionary terms. problems come in many different shapes and sizes.
A natural-scientific perspective suggests that social- The objects of choice in aggregate/consensus decisions
scientific analyses of group decisions might be enriched can be just two options (as in a choice between the
by taking our natural and social evolutionary past into acceptance and rejection of some proposal or policy, or
account too (e.g. Helbing et al. 2000). between leaving or staying in a foraging patch), more
Although cross-referencing of natural science publi- than two, but finitely many, options (as in a choice
cations by social scientists (e.g. List 2004; Hastie & between several electoral candidates, nest sites or food
Kameda 2005), and vice versa (e.g. Conradt & Roper sources), or even infinitely—specifically, continu-
2005), has already begun, indicating the mutual interest ously—many options (as in the choice of a rate of
in interdisciplinary exchange, the language and thinking taxation, which can theoretically take any value
in the social and natural sciences are disparate enough between 0 and 100%, or of a movement orientation,
to hamper communication. This introduction is an which can theoretically be any angle between 08 and
attempt to bridge that gap, render the present issue as 3608, or of a movement speed, which may also take a
useful as possible to both social and natural scientists continuum of values). As we illustrate below, the
and set a common ground for future exchanges. number and structure of options matters.
First, we suggest a categorization of group decisions Similarly, in interactive/combined decisions, the
into aggregate/consensus decisions, on the one hand, choices that individuals face can be of many different
and interactive/combined decisions, on the other and kinds. Sometimes they are binary, as in the choice
introduce some key concepts for the analysis of group between cooperation and defection in a Prisoner’s
decisions in each category. Second, we discuss some Dilemma or similar situation. At other times, individuals
common factors influencing group decisions in both have a choice between more than two possible actions or
humans and animals. Third, we review the contri- strategies, perhaps even a continuum of possibilities, as in
butions in the present issue against this background. the choice between different movement directions.
Finally, we make some remarks on the differences Furthermore, the mechanisms by which individual
between human and non-human group decisions choices lead to certain consequences can vary greatly
and suggest some possible directions for future across different interactive/combined decision problems.
research. We have also compiled a brief and informal Recognizing the large variety of different possible decision
glossary of common social and natural science terms problems is important in so far as different concepts and
(appendix A), which is intended to help social modelling tools are needed for their analysis.
Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 721
Table 1. Examples of different categories of group decisions.
social sciences natural sciences
groups with global overview

decision category: interactive/
combined decisions
group decision democratic/parliamentary votes group activity synchronization
perspective voting strategies who decides and why?
references Austen-Smith & Feddersen (2009), Conradt & Roper (2003, 2007, 2009), Rands et al.
Hix et al. (2009) and many more (2003) and Dostalkova & Spinka (2007)
group decision common goods movements in small groups
perspective individual strategies who decides and how?
references Gächter & Herrmann (2009) and Stewart & Harcourt (1994), Gompper (1996),
many more Conradt & Roper (2005), Amé et al. (2006), Biro
et al. (2006), Meunier et al. (2006), Fischhoff et al.
(2007) and Sueur & Petit (2008)
group decision international agreements cooperative breeding
perspective strategies individual strategies
references a large body of literature Clutton-Brock (1998)
decision category: aggregate/
consensus decisions
group decision international agreements group activity synchronization
perspective patterns segregation
references a large body of literature Conradt (1998), Ruckstuhl (1998, 1999), Conradt &
Roper (2000), Ruckstuhl & Neuhaus (2000,
2002) and List (2004)
group decision democratic/parliamentary votes movements in small groups
perspective systems outcomes, speed, accuracy
references a large body of literature List (2004), Trillmich et al. (2004), Kerth et al.
(2006), Gautrais et al. (2007), Sumpter et al.
(2008) and Ward et al. (2008)
group decision military orders or hierarchical decisions cooperative hunting
perspective any patterns
references Courchamp et al. (2002)
self-organizing systems
decision category: interactive/
combined decisions
group decision consumer choice movements in large groups
perspective how do consumers make choices? mechanisms, individual strategies
references a large body of literature Selous (1931), Gueron et al. (1996), Prins (1996),
Couzin & Krause (2003), Couzin et al. (2005),
Ballerini et al. (2008), Hemelrijk & Hildenbrandt
(2008), Conradt & Roper (2009) and Conradt
et al. (in press)
group decision traffic movements job allocation in eusocial insects
perspective individual behaviours individual decisions
references Couzin & Krause (2003) and Beshers & Fewell (2001)
Ishaque & Noland (2008)
group decision panic behaviour in crowds nest choice in eusocial insects
perspective individual strategies behaviour of individual scouts
references Aube & Shield (2004) Seeley & Buhrman (1999), Franks et al. (2009), List
et al. (2009) and Sumpter & Pratt (2009)
decision category: aggregate/
consensus decisions
group decision consumer choice movements in large groups
perspective market prices speed, accuracy, patterns
references a large body of literature Wallraff (1978), Krause et al. (1992), Reebs (2000),
Couzin & Krause (2003), Simons (2004), Couzin
et al. (2005), Buhl et al. (2006) and Dyer et al.
(2009)
group decision panic behaviour in crowds nest choice in eusocial insects
perspective evacuation time decision speed and accuracy
references Helbing et al. (2000) List et al. (2009) and Sumpter & Pratt (2009)
The central concept for the analysis of aggregate/ discussed in §3b. Formally, an aggregation rule is defined
consensus decisions is that of an ‘aggregation rule’, as as a function which assigns to each combination of

Box 1. Aggregation rules in choices between two options.
1. Basic definitions
Suppose a group of n individuals has to choose between two options, A and B. Each individual, i, can cast a
‘vote’, vi , which can take one of the following three values:
8
> C1 : a vote for option A;
<
vi Z 0: an abstention;
>
:
K1 : a vote for option B:
An ‘aggregation rule’ is a function, f, which assigns to each vector of votes across individuals, (v 1, ., vn ),
a corresponding ‘decision’, vZf (v 1, ., vn ), which can also take one of the following three values:
8
> C1 : a decision for option A;
<
vZ 0: a tie;
>
:
K1 : a decision for option B:
Majority voting, for instance, assigns to each vector (v 1, ., vn ) the value C1 if there are more C1s than K1s
in (v 1, ., vn ), the value K1 if there are more K1s than C1s and the value 0 if the numbers of C1s and K1s
are equal. Thus, ‘majority voting’ is defined as the function f with the property that, for each (v 1, ., vn ),
f ðv1 ; .; vn Þ Z signðv1 C/C vn Þ;
where, for any x,
8
> C1; if xO 0;
<
signðxÞ Z 0; if x Z 0;
>
:
K1; if x! 0:
2. Generalized weighted majority rules
A ‘generalized weighted majority rule’ is a function f with the property that, for each (v 1, ., vn),
f ðv1 ; .; vn Þ Z signðw1 v1 C/C wn vn C mÞ;
where (w1, ., wn) is a vector of ‘weights’ across individuals and m is a ‘decision margin’. In the special case of
equal positive weights w1Z/Zwn and a decision margin of 0, a generalized majority rule reduces to (simple)
majority voting again. If mO0, the rule becomes super-majoritarian for B (meaning that a super-majority of
votes is required for a decision in favour of B), and if m!0, it becomes sub-majoritarian for B (meaning that a
sub-majority of votes is sufficient for a decision in favour of B). In the limiting case in which only one individual
has a positive weight and all other individuals have zero weight, the rule becomes dictatorial.
individual inputs (e.g. votes) a resulting collective output interactive/combined category is often in the eye of the
(e.g. a decision outcome). The classic example is majority beholder. Scientists will characterize a group decision as
voting between two options, under which the group an aggregate/consensus decision if they are interested in
selects the option that receives more votes than the other. the aggregation rule that leads to the decision outcome.
However, a dictatorial decision rule, under which the On the other hand, they will characterize a group decision
group always adopts the choice of a fixed single as an interactive/combined decision if they are interested
individual, the ‘dictator’, is also an aggregation rule. in the individual behaviours and strategies underlying the
The central concept for the analysis of interactive/ collective phenomenon (table 1). For example, social
combined decisions is that of an ‘equilibrium’, also scientists comparing the outcomes of different electoral
discussed in §3c. Formally, an equilibrium is defined as systems for the same configurations of votes may view
a combination of strategies across individuals, which elections primarily as aggregate/consensus decisions, but
satisfies certain ‘best-response’ or ‘stability’ criteria. if they are interested in strategic voting behaviour of
Two classic examples of equilibrium concepts are individuals, they may view elections as interactive/
‘Nash equilibrium’ in the social sciences and combined decisions. Natural scientists studying the
‘evolutionarily stable strategies’ in the natural sciences accuracy of navigational decisions in animal groups may
(Maynard Smith & Price 1973). However, many other view these as aggregate/consensus decisions, but if they are
equilibrium concepts have been proposed in the game- interested in the underlying individual behaviours, they
theoretic literature using a range of different best- may view such decisions as interactive/combined.
response or stability criteria for different contexts. As these examples illustrate, aggregate/consensus
Whether a particular group decision falls into decisions can often be seen as resulting from inter-
the aggregate/consensus category or into the active/combined decisions. The relationship between

Table 2. Some empirical examples for unshared, partially shared and equally shared decision making in social animals.
decision making object of group decision
unshared forage patch choice in primates (King et al. 2008)

travelling start in dolphins (Lusseau & Conradt in press)
at least partially shared cohesive group movements in small groups of birds (Black 1988; Biro et al. 2006), carnivores
(Gompper 1996), ungulates (Conradt & Roper 2003) and primates (Stewart & Harcourt 1994;
Boinski & Campbell 1995; Byrne 2000; Milton 2000)
cohesive group movements in large swarms of insects (Buhl et al. 2006), shoals of fishes ( Ward et al.
2008), flocks of birds ( Wallraff 1978; Simons 2004) or herds of ungulates (Prins 1996)
group activity synchronization (Conradt 1998; Ruckstuhl 1998, 1999)
nest-site choice in eusocial insects (Seeley & Buhrman 1999; Franks et al. 2009) and bats (Kerth et al.
2006)
coordination of cooperative hunts (Courchamp et al. 2002)
coordination of reproduction (Clutton-Brock 1998)
equally shared group flight from potential predators
the two is a key theme in the theory of mechanism practical relevance. One of the aims of social choice
design in economics, which investigates what theory is to identify those aggregation rules that could
mechanisms or systems of incentives would induce be practically relevant.
rational individuals to behave so as to bring about an In order to do so, social choice theorists investigate
outcome that could also result from some aggregation which aggregation rules, if any, satisfy certain proper-
rule. Mechanism design has become an important ties of potential interest. An example of such a property
area of research, as the three Economics Nobel Prizes is ‘universal domain’, which requires the aggregation
in 2007 illustrate (see the survey article by the rule to assign a decision outcome to every possible
Royal Swedish Academy of Sciences 2007). In the combination of individual inputs. Universal domain
natural sciences, some recent developments focus on can be a desirable property because it guarantees a
behavioural mechanisms resulting in the implemen- clear decision outcome in all situations. Another
tation of particular aggregation rules. A key example is ‘anonymity’, which requires that all
mechanism is that of ‘quorum response’ whereby an individual group members have equal weight in
individual’s probability of commitment to a particular determining the outcome. Anonymity is an important
decision option increases sharply once a critical democratic principle. A third example is ‘neutrality’,
number of other individuals (the ‘quorum threshold’) which requires that the different decision options be
have committed to that option (e.g. Sumpter et al. treated symmetrically. Neutrality guarantees that no
2008; Ward et al. 2008; Sumpter & Pratt 2009). bias towards one option is built into the aggregation
Through this positive feedback mechanism, interac- rule itself. A fourth example is ‘positive responsive-
tive/combined decisions among multiple individuals ness’, which requires, roughly speaking, that the
can effectively bring about an aggregate/consensus decision outcome be a positively monotonic function
decision in the group. of individual inputs. Positive responsiveness rules out
the perverse possibility that a winning option becomes
(b) Aggregation rules losing by gaining additional individual support. If we
In aggregate/consensus decisions, a group’s aggrega- restrict our attention to aggregation rules satisfying
tion rule is important as it greatly influences the costs such properties, the set of possibilities shrinks dramati-
and benefits of the group’s decisions to individual cally. In particular, it has been proved that, in group
members and to the group as a whole (e.g. Seeley & decisions between two options, majority voting is the
Buhrman 1999; Conradt & Roper 2003; Rands et al. only aggregation rule simultaneously satisfying the four
2003; Couzin et al. 2005; Hastie & Kameda 2005; properties just introduced (May 1952; for an extension
Austen-Smith & Feddersen 2009). We discuss such and further discussion, see Goodin & List 2006a).
costs and benefits in detail when we address different One particularly important class of aggregation rules
factors influencing group decisions. In this section, we for the case of decisions between two options is that of
briefly review possible aggregation rules. ‘generalized weighted majority rules’ (box 1). The
The set of logically possible aggregation rules for a simplest example of an aggregation rule in that class is
given group decision is enormous. For example, in a majority voting itself. This is the special case in which
group of 10 individuals making a decision between just each individual has one vote, all votes have equal weight,
two options, there are already 210Z1024 possible and the option that gets more votes than the other wins.
combinations of individual votes. Since the aggrega- This could be modified by giving different weights to
tion rule has to assign one of two possible outcomes to different individuals. In this case, the option whose sum
each such combination, there are, in principle, 21024 total of weighted votes exceeds that of the other wins.
possible aggregation rules for this decision. This is For example, in the European Council of Ministers,
more than the estimated number of elementary larger countries have greater voting weight than
particles in the Universe. In more complex decision smaller countries. In animal groups, hungrier group
problems, the combinatorial explosion is even more members can gain more influence on group movement
dramatic. Of course, most of these rules are of no directions than well-fed members (Krause et al. 1992;

Box 2. Equilibria in interactive/combined decisions.
1. A Prisoner’s Dilemma
Suppose two individuals interact. Each of them has a choice between two strategies, cooperation and
defection, and the individual’s pay-off depends on his or her own choice and that of the other individual. In a
Prisoner’s Dilemma, the pay-off structure is as shown in the following table. In each cell, the bottom left entry is
individual 1’s pay-off and the top right entry is individual 2’s pay-off.
individual 2
cooperate defect
individual 1 cooperate 3 4
3 0
defect 0 1
4 1
If both individuals cooperate, each receives a pay-off of 3. If both defect, each receives a pay-off of 1. If one
cooperates and the other defects, the defector receives 4 (sometimes called the pay-off from ‘free-riding’) and
the cooperator receives nothing (sometimes called the ‘sucker’s pay-off’). It is easy to see that the situation in
which both individuals defect is the unique Nash as well as dominant strategy equilibrium: regardless of what
the other individual does, each individual receives a higher pay-off from defecting than from cooperating. It is
also easy to see that defection is the unique evolutionarily stable strategy in this game.
2. Coordination games
Again, two individuals interact, and each of them has a choice between two strategies, A and B, with a pay-off
structure as shown in the following table.
individual 2
A B
individual 1 A 3 0
3 0
B 0 1
0 1
What matters in this game is that they both choose the same strategy (a biological example would be reproductive
synchronization). If they fail to coordinate, they both receive nothing. However, they receive a higher pay-off if
they coordinate on strategy A (namely 3, e.g. the optimal time for reproduction) than if they coordinate on
strategy B (namely 1, e.g. a less optimal time). Here, there exists no dominant strategy equilibrium. For each
individual, the best response depends entirely on the strategy of the other individual. However, both the situation
in which the two individuals coordinate on A and the one in which they coordinate on B constitute Nash
equilibria. Assuming the other individual chooses A, it is best to respond by choosing A too, and similarly for B.
Furthermore, both the strategies are evolutionarily stable in this game, since each gets a higher pay-off from
playing against itself than the other strategy gets from playing against it, and so clause (i) of the definition of
evolutionary stability is met. Nonetheless, it is interesting to note that a population that coordinates on strategy A
will receive higher pay-offs than the one that coordinates on strategy B.
To illustrate the differences between the concept of a Nash equilibrium and that of an evolutionarily stable
strategy, consider the following modification of the coordination game, as given by the pay-off structure in the
following table.
individual 2
A B
individual 1 A 0 3
0 1
B 1 0
3 0
(Continued.)

Box 2. (Continued.)
Here again, it matters that both individuals coordinate their strategy. However, this time, they have to
coordinate on opposite strategies: one individual must play strategy A and the other individual strategy B;
otherwise, neither individual receives any pay-off. Moreover, if they coordinate correctly, the individual who
plays strategy B receives a higher pay-off (namely 3) than the individual who plays strategy A (namely 1).
A biological example would be the allocation of roles in a cooperative hunting expedition: the hunt is only
likely to be successful if different roles are adequately allocated, but different roles will incur different costs
in terms of energy and risks. As in the earlier coordination game, there exists no dominant strategy
equilibrium, but two Nash equilibria: (i) individual 1 plays A and individual 2 plays B, and (ii) individual 1
plays B and individual 2 plays A. However, this time, neither strategy A nor B is evolutionarily stable, since
each gets a lower pay-off from playing against itself than the other strategy gets from playing against it. In an
evolutionary situation, a population of individuals could, instead, reach an evolutionarily stable state, in
which the proportions of individuals playing strategy A or B, respectively, reach a dynamic equilibrium
(here, 25% playing A and 75% playing B). Note that a mixed strategy (play A 25% of the time and B 75%
of the time) would be an evolutionarily stable strategy.
Conradt et al. in press). A limiting case is a dictatorial just for one option, then ‘plurality rule’, which selects
aggregation rule, in which only one individual has a the option with the largest number of votes, has many of
positive weight while all others have zero weight. For the properties of majority voting (List & Goodin 2001;
instance, in many animal groups, decisions are probably Goodin & List 2006a). However, if decisions between
made by a dominant individual (e.g. King et al. 2008). multiple options are decomposed into pairwise choices,
Generally, the assignment of weights can lie any- majority voting and its various generalizations may run
where between an equal weight for all group members into problems. It can then happen that there are
(‘egalitarian’, ‘equally shared’ or ‘dispersed’ decision majorities for option A against option B, for option B
making) and a concentration on one individual against option C and also for option C against option A.
(‘dictatorial’, ‘unshared’ or ‘concentrated’ decision An illustrative situation in which such ‘cyclical’ majority
making). Intermediate cases, in which some group preferences occur is the one in which one-third of the
members (e.g. the highest ranking ones), but not all, group prefers A to B to C, a second third prefers B to C
contribute to the group decision, are particularly to A and the remaining third prefers C to A to B. When
common in practice. Examples in the social sciences majority preferences are cyclical, majority voting yields
are oligarchic or meritocratic decisions. However, even no stable winner—a phenomenon known as ‘Condor-
in democracies, at least some group members are cet’s paradox’ (e.g. Gehrlein 1983). Moreover, this
typically excluded from group decisions (e.g. children, problem is not restricted to majority voting. A classic
adolescents and non-citizens). While truly equally theorem, proved by Nobel laureate Kenneth Arrow
shared decisions are very rare in animals, decisions (1951/1963), shows that, among aggregation rules that
ranging from nearly equally shared to completely preserve the pairwise character of majority voting and
unshared ones have been reported in animals from meet a few other minimal conditions, only dictatorial
insects to primates (table 2; for a review, see also rules generally avoid the occurrence of cyclical collective
Conradt & Roper 2005). preferences (‘Arrow’s impossibility theorem’). Import-
Another way to modify majority voting is to adjust the ant questions in social-choice-theoretic research are
decision threshold, so as to make the aggregation rule therefore (i) how much of a difficulty Arrow’s impossi-
‘super-majoritarian’ (e.g. Goodin & List 2006b) or ‘sub- bility theorem poses for successful aggregate/consensus
majoritarian’ (e.g. Vermeule 2005). The ‘decision decision making over more than two options, and
threshold’ for a given option specifies the vote share (ii) how the problem can be circumvented by either
required for that option to win. For example, one of the giving up the pairwise format of choices between
options might win only if the sum total of weighted votes multiple options—as, for instance, plurality rule
for it is at least twice as large as the sum total of weighted does—or relaxing some of the other conditions of
votes for the other (‘super-majority’), while the other Arrow’s theorem (e.g. Sen 1999; Dryzek & List 2003).
option would win otherwise (‘sub-majority’). An
example is a legislature that agrees to change its (c) Equilibrium concepts
constitution if a super-majority of at least two-thirds When we analyse interactive/combined decisions, the
of its members supports the proposed change aim is to identify combinations of strategies that are
(e.g. Grundgesetz der Bundesrepublik Deutschland ‘equilibria’ (for a good introduction to game theory, see
1949). Another example is a group of foraging animals Osborne & Rubinstein 1994). To explain this idea in
that leaves a patch when a sub-majority of at least one- more detail, consider some interactive situation in
third of group members are in favour of leaving. By which each individual has to choose a certain action
permitting the combination of different individual weight or strategy such that the combination of actions or
assignments with different decision thresholds, the class strategies across individuals leads to a resulting out-
of generalized weighted majority rules is very flexible. come. For example, each individual may have to choose
If there are more than two options, some additional between cooperating and defecting, or between
complications arise. If each individual gets to cast a vote different movement directions, or between different

investment options. A resulting social outcome or model the dynamics of evolutionary replications (for a
pattern—e.g. a collective action, the location of the survey of evolutionary game theory, see Alexander
group or a set of market prices—is then determined by 2003). Particularly relevant to group decision making is
those choices. In game theory, such an interactive the extension of the concept of evolutionarily stable
situation is called a ‘game form’ and is formally defined strategies to multi-player games (Blackwell 1997;
as a specification of a set of possible strategies for each Van Doorn et al. 2003; Bukowski & Miekisz 2004;
individual, together with a mapping from combinations Kaminski et al. 2005; Platkowski & Stachowska-Pietka
of strategies across individuals to resulting outcomes. 2005; Bach et al. 2006; Conradt & Roper 2007, 2009;
To determine when a combination of strategies is an Skyrms 2009).
equilibrium, we need to know what the individuals’
pay-offs or preferences are. A game form together with (d) Global overview versus self-organization
a specification of the individuals’ pay-offs or prefer- In many groups, at least some members can gain a
ences over outcomes is called a ‘game’. A combination global overview of the decision-relevant actions of all
of strategies now constitutes an equilibrium if every other group members (see table 1 for examples in the
individual’s strategy satisfies a certain best-response or social and natural sciences). When there is a global
stability criterion in relation to the other individuals’ overview, group decisions could, at least in principle, be
strategies. Different equilibrium concepts result from reached by general negotiations among all members
different ways of spelling out the notion of best and explicit voting (e.g. Austen-Smith & Feddersen
response or stability. The best-known equilibrium 2009; Hix et al. 2009; for a brief review in animals,
concept in the social sciences is that of Nash see also Conradt & Roper 2003), or by central orders
equilibrium. Here, an individual’s strategy counts as a or coercion (Gavre 1977; Clutton-Brock et al. 1982;
best response to the other individuals’ strategies if King et al. 2008; Lusseau & Conradt in press). In
the individual prefers (or is at least indifferent to) the modern human societies, owing to the sophisticated
outcome of choosing that strategy, compared with the means of mass communication, many group decisions
outcome of deviating from it, assuming that the others fall into the category in which a global overview is at
do not deviate. A stronger notion of best response is that least in principle possible (table 1). In animals, only
of a ‘dominant strategy’. An individual’s strategy is relatively small groups can normally make group
‘dominant’ if the individual prefers (or is at least decisions based on a global overview. For such groups,
indifferent to) the outcome of choosing that strategy, voting has been reported in several mammal and bird
compared with the outcome of deviating from it, species (e.g. Prins 1996; for a review, see also Conradt &
regardless of what the other individuals do. In coordination Roper 2005), and dictatorial or coerced decisions in
games (box 2), for example, there are typically multiple others (Clutton-Brock et al. 1982; King et al. 2008).
Nash equilibria, but there is no dominant strategy Animals employ special postures, vocalizations and/or
equilibrium. In a Prisoner’s Dilemma, by contrast, movements to cast their votes (for a brief review, see
there is a unique Nash equilibrium, which is also a Conradt & Roper 2003).
dominant strategy equilibrium (box 2). There is a huge Sometimes groups are so large that no group
game-theoretic literature on various more refined members can have a global overview of the entire
equilibrium concepts. group. In such cases, individual group members can
In the natural sciences, the best-response or stability only react to local information and communication,
criteria used for defining equilibria are usually and group decisions are made in a self-organizing
evolutionary ones. Here, we illustratively explain the manner. That is, all group members follow their own
approach pioneered by Maynard Smith & Price (1973). local behavioural rules, which rely on local infor-
Consider an interactive situation in which interactions mation (which can be continuously updated), local
take place between pairs of individuals. A strategy, call communication and local reaction to neighbouring
it S, is called evolutionarily stable if it satisfies the group members’ actions. The overall result is a global
following condition: for any alternative (‘mutant’) group behaviour that is not centrally orchestrated, but
strategy T, either (i) S receives a greater pay-off from ‘self-organized’ (Camazine et al. 2003; Couzin &
playing against S than T receives from playing against Krause 2003; Amé et al. 2006; Sumpter 2006; Couzin
S, or (ii) S receives the same pay-off from playing 2007; Hemelrijk & Hildenbrandt 2008; Sumpter &
against S as T receives from playing against S, and S Pratt 2009). A good example is given by the
receives a greater pay-off from playing against T than T movements of large flocks of starlings (Sturnus
receives from playing against T. Formally, S is vulgaris; Ballerini et al. 2008). In flying starling flocks,
evolutionarily stable if, for any T, either E(S, S)O first of all, each individual starling avoids collision with
E(T, S) or [E(S, S)ZE(T, S) and E(S, T)OE(T, T)], direct (local) neighbours by keeping a minimum distance
where E(A, B) is the pay-off of playing a strategy A to them. At the same time, because group cohesion is
against a strategy B. The central consequence of this advantageous for social animals (Krause & Ruxton
definition is that, if sufficiently many individuals in a 2002), each starling does not want to get too far away
population play an evolutionarily stable strategy and from the rest of the group. Thus, when the distance to its
pay-offs represent evolutionary fitness, no mutant direct neighbours gets too large, it moves towards those
strategy can successfully invade the population. neighbours and aligns its direction of movement with
Just as the concept of Nash equilibrium is only one them. Finally, each starling avoids any physical obstacles
of many equilibrium concepts proposed in the social it encounters, and especially predators. The overall
sciences, there are a number of different approaches to result is the fascinatingly synchronized and well-
defining evolutionary stability, some of which explicitly coordinated movement of starling flocks that we observe

in nature, and which does not require anybody ‘in In decisions under uncertainty, the way in which
command’ (Selous 1931). Self-organization also occurs information is aggregated across group members can
in humans (e.g. in the movements of pedestrians, traffic greatly influence the decision pay-offs or accuracy
or panicking/escaping crowds, or even markets; table 1). (Seeley & Buhrman 1999; Conradt & Roper 2003; List
However, we do not usually think of such cases as ‘group 2004; Simons 2004; Couzin et al. 2005; Amé et al.
decisions’, mainly because social cohesion (and, thus, 2006; Biro et al. 2006; Passino & Seeley 2006; Codling
the need for consensus) is not generally their aim. et al. 2007; Lusseau 2007; Ward et al. 2008; Dyer et al.
At first sight, self-organization seems to prohibit 2009; Franks et al. 2009; List et al. 2009; Skyrms 2009;
decisions by general negotiation or voting (but see Sumpter & Pratt 2009). To illustrate this, let us focus
Prins 1996; Seeley & Buhrman 1999), or by central on cases in which the group makes an aggregate/
orders or coercion. Therefore, most natural scientists consensus decision and there exists an unambiguously
studying self-organized group decisions do not ask best outcome, i.e. any ‘disagreements’ between group
questions such as ‘which group members make the members are informational: they may have different
decision?’. It is tempting to assume that all group beliefs, but no conflicts of interest (we discuss such
members contribute equally (via similar local beha- conflicts in §4b). The most classic result on the effects
vioural rules) to the overall outcome, rendering a of the aggregation rule on decision accuracy is
question such as ‘who makes the decision?’ irrelevant. ‘Condorcet’s jury theorem’ (e.g. Grofman et al. 1983;
However, as game theorists know, asymmetric equili- List & Goodin 2001; List 2004), which can be
bria are entirely possible, and recent theoretical work summarized as follows. Suppose a group has to make
has suggested that non-equal contributions of group a choice between two options. Each individual has
members to self-organized group decisions (via dissim- some independent information about which option is
ilar local behavioural rules) are not only possible, but better (the ‘independence’ condition), and each
also likely to evolve under natural conditions. Some individual’s information is correct with an equal
group members could use tactics to influence group probability greater than 1/2 but below 1 (the ‘compe-
decision outcomes disproportionately in their own tence’ condition). Assuming that such independent
interest, even within relatively large groups (Conradt and equally competent individuals vote according to
et al. in press). Some participants’ disproportionate their own information, Condorcet’s jury theorem states
influence on self-organized decisions has also been that the probability that majority voting yields the
observed in the social sciences. For example, the correct outcome (i) is greater than the probability that
evacuation pattern of a crowd in an emergency each group member is individually correct, and
situation could be more influenced by individuals in (ii) converges to 1 (certainty) as the group size increases
certain spatial positions within the crowd than by (see box 3 for a numerical example). This result is a
others (Aube & Shield 2004; Dyer et al. 2009). consequence of the law of large numbers.
Notoriously, some participants in markets, e.g. mono- Condorcet’s jury theorem suggests that shared
polists, have disproportionate influence when decisions are better than unshared ones. The more
compared with others. the group members participate in a group decision, the
more accurate the outcome is likely to be. However, it is
not invariably the case that giving all individuals equal
4. FACTORS INFLUENCING GROUP DECISIONS weight in the decision always leads to the most accurate
At least three central factors influence group decisions: outcome. In particular, if the quality of information—
(a) information, (b) interests, and (c) side constraints the individual accuracy or reliability—differs between
(e.g. time, decision costs, fairness constraints). We group members, an unequal distribution of voting
address them in turn. weights can lead to more accurate decisions, where the
weights are assigned as a function of individual
(a) Information accuracy (see box 3 for an example and a general
When groups make decisions, the pay-offs (costs and result; Ben-Yashar & Nitzan 1997). This could explain
benefits) of the decision outcomes (both for the why in many animal groups adults or more experienced
individuals and for the group as a whole) often group members are the main decision makers (Poole
depend on some state of the environment; for et al. 1988; Stewart & Harcourt 1994; Prins 1996;
example, how the weather will develop, which Conradt & Roper 2003).
location yields the most food, whether there is a An unequal distribution of weights is not the only
predator, which travel route is optimal. Decisions deviation from majority voting which can improve
typically take place under uncertainty, i.e. group the group’s overall decision accuracy. The size
members have only incomplete and noisy information of the decision threshold also matters. If one option
about the state of the environment. An individual’s has a greater prior probability of being best than
decision-relevant information constitutes the individ- any other option, then, other things being equal,
ual’s ‘belief ’, and the probability that a belief is a suitable sub-majority threshold for the high-
correct (i.e. it correctly represents the relevant state probability option results in the most accurate
of the environment) is its ‘accuracy’ or ‘reliability’. If decision outcome (box 3).
there exists an unambiguously ‘best’ decision out- Moreover, the costs and benefits that result from a
come for the group, e.g. an objectively best foraging decision depend not only on the decision accuracy,
patch, nest site or economic policy, we further define which was defined as the probability that the best
the accuracy or reliability of the group decision as the option is selected, but also on the costs of different
probability that the best outcome is selected. types of error. The costs of not choosing one particular

Box 3. Informational differences between group members.
1. Unshared decisions versus equally shared decisions
Assume that a group of five animals has to decide between two foraging patches A and B. Each group member
has a probability of 0.75 of choosing the better foraging patch individually. If the group employs a dictatorial
(unshared) aggregation rule, with the dominant member making the decision, the group has a chance of 0.75
of correctly choosing the better foraging patch (which is the probability that the dominant individual makes a
correct decision). On the other hand, if members share the decision equally and use majority voting as their
aggregation rule, the group will choose the better patch correctly as long as at least three group members ‘vote
correctly’. That is, the group chooses the better patch with a probability of
!
X 5 5
0:75i $0:255Ki z0:90;
i Z3 i
which is a 15 per cent increase in accuracy.
2. Difference in information between group members
Assume that the dominant individual is most experienced and can determine the better foraging patch
correctly with a probability of 0.75, while the other four members can do so only with a probability of 0.6.
A majority decision would lead to the better patch if at least three group members voted for the better patch.
That is, with a probability of
! !
X 4 4 X 4 4
4Ki
0:75$ i
0:6 $0:4 C 0:25$ 0:6i $0:44Ki z0:73;
i Z2 i i Z3 i
the group chooses the better patch. This is a lower accuracy than that of 0.75 for the unshared decision by the
dominant individual alone. In such a case, instead of sharing the decision equally, it would be beneficial for
the animals to give the more knowledgeable individual more weight in the decision. Assume, for example, the
dominant individual is given three times the voting weight of the others. The group would then choose the better
foraging patch correctly if either the dominant individual and at least one other individual voted correctly
(resulting in at least 4 : 3 weighted votes for the correct patch) or if the dominant individual voted incorrectly
but all others voted correctly. The unequally shared decision outcome would have an accuracy of
! !
X 4 4 4Ki
X 4 4
0:75$ i
0:6 $0:4 C 0:25$ 0:6i $0:44Ki z0:76;
i Z1 i i Z4 i
which is better than that of an unshared or an equally shared decision. More generally, assuming differentially
reliable group members but an equal prior probability and equal benefits of each foraging patch being best,
the optimal aggregation rule is a weighted majority rule where each individual’s weight is proportional to
log( p/(1Kp)), with p being the individual’s reliability (e.g. Grofman et al. 1983). The fully general result
(Ben-Yashar & Nitzan 1997) is discussed below.
3. Influence of the decision threshold on decision accuracy
(a) Skewed likelihood of a particular option to be the ‘best’
Assume that foraging patch A has a probability of 0.9 of being better than patch B, and that all five group
members have a probability of 0.75 of detecting the better patch correctly. The accuracy of an equally shared
decision depends on the size of the decision threshold. Suppose that at least two members are required to vote in
favour of patch A in order for the group to choose patch A (i.e. the threshold is a sub-majority of two in favour
of A). Then, the group will choose the better patch correctly if either patch A is the better patch and at least two
members vote correctly, or if patch B is the better patch and at least four members vote correctly. That is, the
expected accuracy of an equally shared group decision with a sub-majority threshold of two for patch A is
! !
X 5 5 X 5 5
5Ki
0:9$ i
0:75 $0:25 C 0:1$ 0:75i $0:255Ki z0:95:
i Z2 i i Z4 i
The respective accuracy for a (simple) majority threshold is 0.90 (as above in subsection 1). Finally, if there is a
super-majority threshold for A (e.g. at least four members are required to vote in favour of patch A for the group
to choose patch A), the expected accuracy of the group decision outcome will be
! !
X 5 5 5Ki
X 5 5
0:9$ i
0:75 $0:25 C 0:1$ 0:75i $0:255Ki z0:67:
i Z4 i i Z2 i
(Continued.)

Box 3. (Continued.)
Thus, the most accurate decision here is a shared group decision with a sub-majority threshold in favour of the
foraging patch that is more likely to be the better patch.
(b) Benefits are also skewed
Assume that we still have the same situation as under (a), but now patch A and patch B yield different benefits
when they are the ‘best’ yielding patch, respectively. Assume that when patch B is best, it yields 50 times as
much as patch A yields when it is best. The expected benefits of an equally shared group decision with a sub-
majority threshold of two for patch A are
! !
X 5 5 X 5 5
5Ki
0:9$1$ i
0:75 $0:25 C 0:1$50$ 0:75i $0:255Ki z4:1:
i Z2 i i Z4 i
The expected benefits with a majority threshold are

! !
X 5 5 X 5 5
5Ki
0:9$1$ i
0:75 $0:25 C 0:1$50$ 0:75i $0:255Ki z5:3:
i Z3 i i Z3 i
The expected benefits with a super-majority threshold of four for patch A are:
! !
X 5 5 5Ki
X 5 5
0:9$1$ i
0:75 $0:25 C 0:1$50$ 0:75i $0:255Ki z5:5:
i Z4 i i Z2 i
Here, the threshold that yields the most benefits is a super-majority threshold for patch A (i.e. a sub-majority
threshold for patch B).
(c) The fully general result
Consider a choice between two options, A and B, without any conflict of interests. Suppose that r is the prior
probability of option A being better and each individual i in an n-member group has an individual accuracy pi of
identifying the better option. There are four possible scenarios: (i) option A is better and is chosen, (ii) option A
is better and is not chosen, (iii) option B is better and is chosen, and (iv) option B is better and is not chosen. Let
us write u C for the pay-off difference between (i) and (ii), and u K for the pay-off difference between (iii) and
(iv). The general result (Ben-Yashar & Nitzan 1997) states that the expected pay-off is maximized by a weighted
generalized majority rule of the form
f ðv1 ; .; vn Þ Z signðw1 v1 C/C wn vn C mÞ;
where
— for each i, individual i ’s weight wi is proportional to log( pi /(1Kpi)) and
— the decision margin m is the sum of two parameters:
— the base-rate bias, log(r/(1Kr)), and
— the pay-off-asymmetry bias, log(u C/u K).
It is easy to see that this rule becomes super-majoritarian for B (i.e. a super-majority is required for a decision in
favour of B) if option A has a higher prior probability than option B or the cost of erroneously deciding against A
is higher than the cost of erroneously deciding against B. In the opposite case, it becomes super-majoritarian for
option when it is best need not be the same as the costs from a potential predator. Fleeing is only the best option
of not choosing another when that option is the best. if there really is a predator, which might be less likely
Asymmetries in the costs of different errors can even than there being none. However, if there is a predator
make a decision threshold optimal that fails to and fleeing is best, then the potential costs of not fleeing
maximize overall decision accuracy. For example, one are extremely high (e.g. losing one’s life). On the other
decision option might rarely be the best option, but hand, if not fleeing is the best option (because there is no
when it is the best, it might yield much higher benefits predator), the potential costs of fleeing nonetheless (e.g.
than alternative options. In such a case, a sub-majority to miss the opportunity of some additional foraging) are
threshold for the low-probability option could result in relatively modest in comparison. In such cases, theory
the highest expected benefits, despite the fact that a predicts that a shared group decision with a sub-majority
super-majority threshold would maximize overall threshold for fleeing is optimal (box 3; Ben-Yashar &
decision accuracy (box 3). A good animal example is Nitzan 1997; List 2004). Indeed, this is what we usually
given by group decisions about fleeing or not fleeing observe in nature: a relatively small number of group

members (i.e. a sub-majority) can trigger the flight of disasters, find good foraging sites or avoid predators,
a whole group (e.g. Krause & Ruxton 2002; Lingle & for instance. However, there are also frequent cases in
Pellis 2002; Boland 2003; Stankowich & Blumstein which the members’ interests come apart. What is good
2005; Carter et al. 2008). A social science example is for one individual may be bad for another. Consider,
given by decisions about constitutional changes. Owing for example, different configurations of market prices:
to the central importance of a state’s constitution, some favour consumers and others producers; or
changing it involves great risks and thus much higher different tax laws or redistributive policies: some are
potential costs (i.e. negative benefits) than keeping the better for big companies and others for low-income
status quo. In recognition of this, as we have already individuals. In social animals, group members of
noted, constitutional changes often require a super- different size, sex, age or physiological state are likely
majority of two-thirds or more of legislators, rather than to have different requirements, which often lead to
just a simple majority (e.g. Grundgesetz der Bundesre- different interests. For example, larger individuals may
publik Deutschland 1949). Similarly, in criminal trials, favour longer activity durations than smaller individ-
juries are usually required to make decisions by super- uals (e.g. Clutton-Brock et al. 1982; Conradt 1998;
majority rules, which implement a presumption of Ruckstuhl 1998, 1999), females with vulnerable
innocence, because it is considered to be far more costly dependent offspring may favour safer sites than males
in moral terms to convict the innocent than to acquit the (Ruckstuhl & Neuhaus 2000, 2002), older or larger
guilty. The idea is captured by the famous principle, in individuals may favour sites with different forage than
the words of the English legal scholar Blackstone younger or smaller individuals (Clutton-Brock et al.
(1765–1769), that it is ‘better that ten guilty persons 1982; Gompper 1996; Prins 1996), non-starving
escape than that one innocent suffer’. individuals may favour less exposed sites than starving
The standard results building on Condorcet’s jury individuals (e.g. Krause et al. 1992; Rands et al. 2003),
theorem are based on the assumption that the votes or and so on (for the most recent review, see Conradt &
information of different decision makers are indepen- Roper 2005). When pay-offs are not only different
dent. Crucially, the filtering of errors that is ensured by across individuals but lead to different relative rankings
the pooling of a large number of signals requires that of decision options, we speak of ‘conflicting interests’.
errors are uncorrelated. In the limiting case in which Social scientists also describe the ranking of decision
different individuals’ votes, and thereby their errors, are options from an individual’s perspective as this
perfectly correlated with each other, aggregation yields individual’s ‘preference’ over options.
no gains in accuracy. The benefits of information Assuming a conflict of interests within a group, it is
pooling in the presence of less extreme interdependen- obvious that the way in which different individuals’
cies depend on the nature of these interdependencies interests or preferences are aggregated can make a great
(Boland 1989; Ladha 1992; Dietrich & List 2004; difference to the group’s overall pay-offs, and also to the
Berend & Sapir 2007). Among the kinds of inter- individual pay-offs received by each group member.
dependencies that can significantly compromise Social choice theory has studied the aggregation of
decision accuracy are ‘informational cascades’, in conflicting interests or preferences in great depth,
which a plurality or majority that accidentally emerges beginning with Arrow’s (1951/1963) seminal work.
in support of some option is mistakenly interpreted by While Arrow’s classic theorem, as we have already
others as evidence for the optimality of that option and mentioned, uncovers some of the difficulties of
thereby attracts further support, although few or any aggregation in decisions between more than two
individuals originally had any information in support of options, we can say something positive about majority
that option (e.g. Bikhchandani et al. 1992). Market voting in two-option choices. Just as majority voting is
bubbles or instances of groupthink in committees are good at pooling dispersed information in such choices,
phenomena of this kind in human contexts (e.g. so it also has some appealing properties with regard to
Sunstein 2006), and they could, in principle, also the aggregation of conflicting interests or preferences
occur in animals (Giraldeau et al. 2002; Dugatkin (see box 4 for more details). Suppose that some group
2005; List et al. 2009; Sumpter & Pratt 2009), and members prefer option A to option B, while others have
sometimes do so in practice (Seeley & Buhrman 2001; the reverse preference. It is easy to see that majority
Dyer 2008; Ward et al. 2008). voting, uniquely among aggregation rules, maximizes
In summary, it is obvious that the problem of the number of group members whose preference is
pooling dispersed information across a group of respected. Indeed, this property can be seen as a
individuals is a complex one, and so far we have even defining characteristic of majority voting. Further-
ignored an important additional complicating factor, more, if we assume that each individual receives a
namely the influence of conflicts of interests between pay-off of 1 from having his or her preference respected
group members. We now turn to this issue. and a pay-off of K1 otherwise, then majority voting
maximizes the sum total of pay-offs across the group
(b) Interests (this is the key insight underlying a theorem by Rae
The pay-offs of a decision outcome for a group of 1969 and Taylor 1969).
individuals obviously depend on whether the outcome However, in many real-world cases, different group
promotes, or is at least consistent with, the members’ members have different ‘stakes’ in a decision. Formally,
interests. In §4a, we have made the simplifying an individual’s ‘stake’ in a decision between two options
assumption that all group members share the same is defined as the pay-off difference between the better
interests. In many cases, this assumption is warranted: option from the individual’s perspective and the worse
all group members want to prevent decision-induced one (in the natural sciences, this is also called the

Box 4. Interest differences between group members.
Suppose, as before, that a group of n individuals has to choose between two options, A and B. Suppose that
each individual i receives pay-offs of u i(A) and u i(B) from options A and B, respectively. Assume that when a
vote is taken between A and B, each individual votes for the option with the higher pay-off, i.e. each individual
i’s vote is
8
> C1 ða vote for AÞ; if ui ðAÞO ui ðBÞ;
<
vi Z 0 ðan abstentionÞ; if ui ðAÞ Z ui ðBÞ;
>
:
K1 ða vote for BÞ; if ui ðAÞ! ui ðBÞ:
Recall that if each individual i gets a voting weight of wi , the outcome of a weighted (simple) majority vote is
f ðv1 ; .; vn Þ Z signðw1 v1 C/C wn vn Þ:
Note that option A yields a higher sum total pay-off than option B if and only if
u 1 ðAÞ C u 2 ðAÞ C/C un ðAÞO u 1 ðBÞ C u 2 ðBÞ C/C un ðBÞ;
i.e. if and only if
ðu 1 ðAÞK u 1 ðBÞÞ C ðu 2 ðAÞK u 2 ðBÞÞ C/C ðun ðAÞK un ðBÞÞO 0;
which, in turn, is equivalent to
s1 !v1 C s2 !v2 C/C sn !vn O 0;
where, for each i, si is individual i’s ‘stake’ si Z jui ðAÞK ui ðBÞj, with jxj defined as the absolute value of x.
Similarly, option B yields a higher sum total pay-off than option A if and only if the reverse inequalities hold.
Rewriting this observation in slightly more general terms, we obtain
8
> C1; if u 1 ðAÞ C u 2 ðAÞ C/C un ðAÞO u 1 ðBÞ C u 2 ðBÞ C/C un ðBÞ;
<
signðs1 !v1 C s2 !v2 C/C sn !vn Þ Z 0; if u 1 ðAÞ C u 2 ðAÞ C/C un ðAÞ Z u 1 ðBÞ C u 2 ðBÞ C/C un ðBÞ;
>
:
K1; if u 1 ðAÞ C u 2 ðAÞ C/C un ðAÞ! u 1 ðBÞ C u 2 ðBÞ C/C un ðBÞ:
From this, we can immediately infer that weighted majority rule produces as its winner the option that
maximizes the sum total pay-off, provided that each individual i’s voting weight wi is proportional to his or her
stake si ( Fleurbaey 2008).
individual’s ‘potential consensus cost’). For example, decision stakes or something equivalent. We briefly
a civil servant with a high level of job security has a return to these issues towards the end of the paper.
lower stake in a decision about unemployment benefits However, while maximizing the sum total of pay-offs
than someone on a short-term contract; a resident of across the group is sometimes an explicitly intended
Central London has a higher stake in a decision about outcome in a human context, it is of less direct
inner-city congestion charging than an infrequent visitor relevance to the natural sciences (e.g. Conradt &
to the city. In animals, for example, starving or hungrier Roper 2007; see also §6c). Animals do not ordinarily
individuals might have higher stakes in foraging have an incentive to try to maximize the pay-offs for the
decisions than do well-fed ones (e.g. Prins 1996; Rands group as a whole (Smith 1964, 1976, 1998). Rather,
et al. 2003; Conradt et al. in press); small, vulnerable each individual is likely to try to maximize its own pay-
individuals have higher stakes in decisions about offs, possibly at the expense of other group members.
predator avoidance than do large, less vulnerable Nevertheless, it is not unlikely that constraints on each
ones (e.g. Ruckstuhl & Neuhaus 2000, 2002; Lingle & individual group member often act in such a manner
Pellis 2002). that the decision outcome takes individual stakes into
Generalizing the earlier result about majority voting, account, and might even approach the outcome that
one can show that when different individuals have gives the maximal group pay-offs (Conradt & Roper
different stakes in a decision, and the decision is 2003; Rands et al. 2003; Conradt et al. in press). The
between two options, a weighted majority rule with reason is that in many aggregate/consensus decisions by
weights assigned to the individuals in proportion to social animals, one aspect that is likely to be important
their stakes maximizes the sum total of pay-offs across to all group members is to maintain group cohesion
the group (Fleurbaey 2008). In decisions between (Krause & Ruxton 2002). As a consequence, individ-
more than two options, the picture is more compli- uals have to trade-off, on the one hand, influencing a
cated, but it is widely agreed among social choice decision outcome assertively in their own interest and
theorists that successful preference or interest aggrega- thereby risking group fragmentation against, on the
tion with respect to certain social optimality criteria in other hand, maintaining group cohesion by being less
such cases requires taking into account the individuals’ assertive (e.g. Conradt 1998; Couzin et al. 2005).

The likely result is that those individuals whose stakes responses in animals; for the most recent review, see
are higher will be more assertive than those whose Sumpter & Pratt 2009). Third, considerations of
stakes are lower, so that the former have more influence fairness, legitimacy or preservation of future good
on the decision outcome than the latter (Conradt et al. relations (Brosnan & de Waal 2003; Brosnan et al.
in press). There is empirical evidence which supports 2005) may rule out certain decision-making arrange-
this argument (e.g. Krause et al. 1992). However, the ments that might be optimal from the perspective of
rationale also implies that those individuals within a accuracy or benefit maximization alone. For instance,
group for which group cohesion is least important (as in many democratic settings, weighted majority rules—
opposed to those for which the stakes are highest) even if they might occasionally be accuracy maximizing
might gain the most weight in a group decision in cases of differential individual reliability—are
(Conradt et al. in press). Again, there is empirical considered democratically unacceptable as well as
evidence that this can occur (Prins 1996). potentially open to abuse due to the inbuilt power
So far, we have discussed information aggregation in asymmetries. Similarly, considerations such as respect
the case of no conflict of interests, and interest for certain rights may trump the maximization of
aggregation without considering the possibility of accuracy or benefits alone.
unreliability of information. Under natural conditions,
of course, there may be both, unreliability of information
and conflicts of interest, at the same time. Thus, 5. PRESENT ISSUE
the most difficult question remains: what happens (a) Sharing information
when group members differ in their quality of infor- The first contribution to this issue, by Sumpter & Pratt
mation and have differing interests? While this is a (2009), gives a concise and comprehensive review of
frequently discussed scenario in the social sciences (e.g. the empirical and theoretical literature on quorum
Austen-Smith & Feddersen 2009), there has been very responses in animal group decision making. In many
little natural-scientific work done in this direction. social animals, quorum responses are a likely and
Couzin et al. (2005) suggest that if the numbers of plausible mechanism of reaching aggregate/consensus
individuals within a group which prefer either of two decisions, which can ultimately be interpreted in terms
options are fairly balanced, the differences in infor- of social science’s aggregation rules, but do not require
mation reliability can topple the decision in favour of the complex cognitive abilities. In addition to the review,
better informed individuals. Intuitively, when stakes are the authors present an elegant and effective model of
relatively low and information unreliable, information how animals could optimize decision accuracy (in the
should be the dominant factor influencing group form of information sharing) or decision speed, or solve
decisions. For example, it might be better to follow the trade-off between speed and accuracy, by adjusting
others reliably to a slightly less optimal foraging patch simple parameters in their quorum response.
than to seek, but not to find, a more optimal patch. As we have noted, Condorcet’s jury theorem
Similarly, if stakes are high and information relatively requires the independence of individual judgements.
reliable, interests might be expected to be the dominant On the other hand, without any interdependencies
factor. However, it is less clear what will happen between individuals, real-world groups may often find
in situations in which either (i) stakes are high and it difficult to reach a consensus. This raises some
information is unreliable, or (ii) stakes are low and important questions about how independence and
information is relatively reliable. There is much scope interdependence interact in determining aggregate/
for further research into these questions. consensus decision outcomes and their accuracy. In
response to these questions, the contribution by List
(c) Side constraints et al. (2009) brings together social- and natural-
Group decisions are often subject to important side scientific insights by applying a social-choice-theoretic
constraints such as time constraints (e.g. Passino & model to an animal system: swarming honeybees
Seeley 2006; Franks et al. 2009; Sumpter & Pratt choosing a new nest site. The authors show that both
2009), decision costs, computational limitations (e.g. a sufficient degree of independence and a sufficient
Gigerenzer & Selten 2002) and fairness constraints degree of interdependence between individual bees are
(e.g. Brosnan & de Waal 2003; Brosnan et al. 2005; needed to predict the high accuracy of nest-site choice
Dawes et al. 2007; Fehr et al. 2008). Although a fully observed empirically. Specifically, bees have to be
optimal solution to a given decision problem may exist relatively independent in assessing the quality of
in theory, it can often be difficult or costly in practice to prospective nest sites once they visit them, while they
find it. First, owing to search costs or time constraints, also have to be relatively interdependent in signalling to
not all theoretically possible decision options can be each other which sites are worth inspecting. The
considered by the decision makers. Instead, the interplay between independence and interdependence
decision makers may be restricted to the consideration allows the bees to reach a consensus with high accuracy
of some practically salient or easily identifiable ones within a realistic time frame. List et al.’s (2009) model
(e.g. Seeley & Buhrman 2001; Franks et al. 2009). can be seen as complementary to the quorum response
Second, owing to time constraints or other compu- model by Sumpter & Pratt (2009), the crucial
tational limitations, the full calculation to solve a parameters of which could also be interpreted in
particular optimization problem may often be infeas- terms of ‘independence’ and ‘interdependence’.
ible, and certain shortcuts, which may lead to less Austen-Smith & Feddersen (2009) examine indi-
optimal decisions, may have to be taken in practice (e.g. vidual deliberation and voting strategies underlying
Gigerenzer & Selten 2002; consider also quorum aggregate/consensus decisions in small groups of

humans. The authors illustrate that, surprisingly, members. As in the models by List et al. (2009) and
informative voting—truthfully revealing private infor- Sumpter & Pratt (2009), the balance between indepen-
mation—need not be individually rational even when dence and interdependence plays a significant role here.
all group members share the same interests (the case of Individual group members are attracted towards, and
‘common values’). On the other hand, strategic voting align with, neighbouring group members within a local
can lead to suboptimal decision outcomes. When there interaction range in order to maintain social cohesion
are no conflicts of interests, such problems can be (interdependence). Additionally, informed members
overcome by communicating private information prior balance this interdependence of social attraction against
to voting. Moreover, the incentives for informativeness moving in the direction of a known resource (indepen-
depend on the voting rule used. However, when dence). If interdependence is too low, the group splits. If
individuals can differ not only in their information independence is too low, the group does not move
but also in their interests, these positive results break efficiently towards the resource.
down. The authors point out that, even when there is
only a small degree of uncertainty about whether or not (b) Resolving conflicts
group members share common values, there may not Often groups have to make decisions in situations with
exist a voting rule that leads all individuals to vote considerable conflicts of interests between members
informatively, and individuals may also have incentives with respect to the optimal decision outcome. Resol-
not to reveal their private information truthfully in ving such conflicts requires cooperation. Gächter &
deliberation prior to voting. The authors give some Herrmann (2009) investigate the basis of cooperative
illuminating insights into the complexity that human behaviour in humans in ‘common goods’ experiments,
strategic considerations add to the more straightfor- in which the best interest of the individual is different
ward processes of information sharing described by from the best interest of other group members. Direct
List et al. (2009) and Sumpter & Pratt (2009). They and indirect reciprocity, and peer punishment, are the
also show how group decision making is greatly most important determinants of successful cooperation
complicated when there are informational and interest in such situations. However, a large number of
differences between group members. individuals cooperate (or punish free riders) altruisti-
As illustrated by the first three contributions to this cally even when there is no opportunity for either direct
issue, coordinated action requires the transmission and or indirect reciprocity. Culture has a strong influence
processing of information among group members. on such behaviour. Surprisingly, the authors also find
Information transmission in groups usually involves that antisocial punishment, where cooperators rather
signalling between multiple senders and receivers. As than free riders are punished, is much more widespread
discussed in Skyrms’ (2009) contribution, this can be than previously assumed. Understanding antisocial
modelled in terms of ‘sender–receiver games’, in which punishment is an important task for future research,
senders observe certain states of the world, transmit because antisocial punishment is a strong inhibitor
particular signals—which may or may not accurately of cooperation.
convey their information—and elicit resulting acts in Conflicts can be resolved, for example, by sharing
the receivers. Sender–receiver games can have multiple decisions equally between members. Conradt & Roper
Nash equilibria, but the only evolutionarily stable ones (2009) explore which conditions favour the evolution
are so-called ‘signalling systems’, in which information of equally shared decisions. Interestingly, these con-
transmission is accurate. Despite their evolutionary ditions depend crucially on whether the modality about
stability, Skyrms reports that, surprisingly, signalling which the group decides is ‘continuous’ or ‘disjunct’. A
systems need not generally evolve. Other equilibria, continuous modality is, for example, timing of commu-
which are not evolutionarily stable as defined by nal activities if the mean of all the timings preferred by
Maynard Smith & Price (1973), can still be ‘dynami- individual group members could constitute a sensible
cally stable’ in a sense defined in the paper, and Skyrms compromise. On the other hand, an example for a
discusses the properties of such equilibria. With respect disjunct modality is communal spatial destination if the
to group decision making, this implies that suboptimal mean of all preferred destinations (e.g. the space in the
information transmission in animal groups can, in middle between two foraging patches) is not a sensible
principle, persist over evolutionary time scales, even compromise. In decisions on continuous modalities,
when the aim of information sharing is not hampered the higher the potential consensus costs are, the more
by conflicts of interest between senders and receivers. likely it is that an equally shared decision evolves.
Although signalling systems might not be guaranteed By contrast, in decisions on disjunct modalities,
to evolve, Dyer et al. (2009) show that decision-relevant the higher the potential consensus costs are, the more
information can nevertheless be shared efficiently within likely it is that an unshared decision evolves, or a
groups and without having to employ any intentional decision that is only shared between certain like-
signalling at all. The authors review theoretical and minded group members.
empirical studies on leadership in social animals and In humans, important decisions are often about
humans. They report that, in self-organizing groups, a disjunct modalities, and potential consensus costs can
relatively small minority of informed group members be high. As Conradt & Roper’s (2009) work suggests,
can already lead a large majority of uninformed in such cases, it could be important for like-minded
members in a preferred direction with high accuracy. individuals to try to form an alliance to influence
This can happen without any intentional signalling by decision outcomes in their joint interest. Hix et al.
the informed members, and when the informed (2009) investigate alliance formation in the European
members are not even identifiable to uninformed Parliament. Although cohesion is neither enforced, nor

directly rewarded or punished, group association to researchers construct summary statistics of individual
cross-national ‘political groups’ with similar political preferences (e.g. psychologists aggregating the
views can explain 90 per cent of an MEP’s voting responses given by several experimental participants
behaviour (by contrast, nationality only explains 10%). to preference questions).
Reasons for forming such voluntary alliances are
division of labour (e.g. with respect to information (c) Respecting side constraints
gathering), reciprocal altruism and voting cooperation While the pooling of information and the resolution of
within political groups. Cohesive voting within alli- conflicts play important roles in group decision
ances is also maintained by the possibility of punish- making, we have also pointed out that group decisions
ment of individuals through the prospect of are often subject to important side constraints, such as
withholding future influential positions within the time, costs, computational and fairness constraints.
alliance. These observations are in good agreement Several of the issue’s contributions either implicitly or
with Gächter & Herrmann’s (2009) results about the explicitly discuss such constraints. As already noted,
basis of cooperation in humans. However, Hix et al. Sumpter & Pratt (2009) address some of the trade-offs
(2009) also report that political group association can between speed and accuracy in group decisions, and
break down when there is a chance that a vote is pivotal List et al. (2009) suggest that the interplay between
and there are strong national interests at stake. The independence and interdependence is one of the
authors further illustrate the power that is conveyed by factors contributing to solving such a trade-off. Fair-
setting the agenda: which issues are put up for a vote, ness constraints feature in Gächter & Herrmann’s
and how this is done, can significantly influence (2009) analysis of human cooperative behaviour in
decision outcomes and policies. Again, cooperation common goods experiments, and at least implicitly in
within alliances with respect to agenda setting can offer Hix et al.’s (2009) discussion of some of the properties
great advantages to individual MEPs. of political alliances.
In parliamentary and electoral decisions, there are The paper by Franks et al. (2009) puts its central
often more than two alternatives. While many countries focus on temporal side constraints on decisions. In the
and other political units use plurality rule as their case of emergencies, urgency constrains the time
electoral method—under which each voter casts a vote available for group decision making: groups may have
for only one option—many legislatures decompose to make quick decisions, often at the expense of
many-option choices into multiple pairwise choices. decision accuracy. Franks et al. describe an empirical
For instance, a parliament that ultimately seeks to example of the speed–accuracy trade-off in nest-site
decide between the status quo, a particular policy choices by emigrating ants. The authors highlight the
proposal and an amended version of that proposal different stages at which such group decisions can be
may first take a pairwise vote between the original sped up and traded against accuracy, and the role
proposal and its amended version, and next between played by the ants’ quorum response mechanism. An
the winner of that first vote and the status quo. As noted important side constraint in the ants’ decision making
earlier, Condorcet’s classic paradox highlights the is the number of active scouts available at each stage of
possibility that majority voting over multiple pairs of the process, not only during the decision making itself,
options may produce cyclical majority preferences, but also during the implementation of its outcome.
meaning, for instance, that an amended proposal may Since high scout numbers at one stage can lead to a low
be majority preferred to the original proposal, the availability of scouts at another stage, recruiting scouts
original proposal to the status quo and the status quo, in optimally to different stages is crucial in order to avoid
turn, to the amended proposal. In such cases, there decisions that are neither accurate nor fast.
exists no stable majority winner, and the decision
outcome may arbitrarily depend on the order in which
pairwise votes are taken. Although a large body of work 6. DIFFERENCES BETWEEN HUMAN AND
in social choice theory suggests that this phenomenon NON-HUMAN GROUP DECISIONS
should be ubiquitous (e.g. Gehrlein 1983), there is So far, we have emphasized concepts for the analysis of
strikingly little empirical evidence for it. Regenwetter group decisions and factors influencing such decisions,
et al. (2009) survey some recent developments in which are common to humans and non-humans. In
behavioural social choice theory that seek to account conclusion, it is also worth looking at some of the
for the discrepancy between the standard theoretical differences between human and non-human group
predictions and the lack of empirical support for them. decisions. We focus on three central areas in which such
In particular, they show that the predicted ubiquity of differences arise: first, the kind of ‘rationality’ at work;
majority cycles is based on some statistical assumptions second, the role of language, which affects both the sort
about the distribution of voter preferences (so-called of communication that can take place prior to a group
‘cultures of indifference’) that are not empirically decision and the possible content of the decision itself;
supported. Once the theory is revised by taking into and third, the kinds of optimality concepts that are
account the kinds of preference distributions that we relevant for the assessment of group decisions.
find in many real-world political settings, its new
prediction is that majority cycles should be much less (a) Rationality
frequent than commonly assumed. The authors discuss An important difference between the social-scientific
a number of implications of this finding and consider and natural-scientific analysis of group decisions lies in
the application of their insights about aggregation the kind of rationality that is attributed to the agents
paradoxes to other, non-voting settings in which under investigation (humans versus non-humans).

In the social sciences, human individuals are usually decision in the two cases is very different. In the
modelled as being ‘rational’ in some appropriate sense. non-human case, communication takes the form of the
On standard game- and decision-theoretic approaches, exchange of relatively simple signals and the sub-
this means, roughly, that individuals act in such a way as sequent decision consists in the support for one
to maximize the utility or pay-offs they expect to attain, in particular option or in the choice of a concrete
the light of their beliefs about the environment. This behavioural strategy. In the human case, by contrast,
makes them very flexible. If they are presented with a the expressive resources of language can make both
new situation and a new pay-off structure, they will stages—the ‘communication stage’ and the ‘decision
adjust their strategies or actions so as to maximize their stage’—much more complex.
individual expected utility or pay-offs in the new At the communication stage, language allows
situation, so long as they have enough information to humans to exchange not only simple informational
update their beliefs accordingly. In the recent, more signals, but also complex arguments, hypothetical
psychologically informed approaches in the areas of considerations, analogies and anecdotes and entire
behavioural game and decision theory, this picture is beliefs systems or theories. The theory of deliberative
somewhat refined. It is acknowledged, in particular, that democracy addresses the ways in which linguistic
humans exhibit a number of cognitive constraints, some communication can affect—sometimes positively and
of which may be traced back to our evolutionary history. at other times adversely—successful group decision
Instead of explicitly maximizing expected utility, for making (e.g. Cohen 1989; Gutman & Thompson
example, individuals often use simple rules of thumb 1996; Elster 1998; Dryzek & List 2003; Sunstein
(‘heuristics’), which may lead to systematic errors 2006). In so-called ‘deliberative polling’ experiments,
(‘biases’) (e.g. Gigerenzer & Selten 2002). Nonetheless, for example, it has been shown that a period of group
behavioural game and decision theory retain at least the deliberation among randomly chosen participants,
assumption that humans choose their strategies or before and after which they are individually interviewed,
actions relatively flexibly—albeit under some psycho- can significantly change their opinions on political issues.
logical constraints—in response to their beliefs and In the best-case scenario, group deliberation not only
preferences about the environment. increases the participants’ factual information, but also
In the natural sciences, by contrast, the way in which makes them more other-regarding and leads them to
rationality comes into play is quite different. While develop a shared understanding of their decision
some variant of the standard game- and decision- problem (Luskin et al. 2002; Fishkin & Luskin 2005;
theoretic understanding of rationality may still be List et al. 2000/2006). Under less benign circumstances,
applicable, with further constraints, to animals with for instance when groups are too homogeneous and
relatively sophisticated cognitive systems (e.g. share an initial bias towards certain opinions (e.g. pro-
primates, mammals or birds; Dennett 1987), it seems war), group deliberation can further reinforce this bias, a
clear that many insects or fishes, for example, cannot be phenomenon sometimes described as ‘group polar-
usefully understood in this way. Instead, the notion of ization’ and related to the phenomenon of informational
rationality is thought to apply at an evolutionary level. cascades mentioned earlier (Bikhchandani et al. 1992;
Thus, it is no longer the case that individuals Sunstein 2002, 2006).
themselves make rational choices between different Similarly, at the decision stage, human language
possible strategies, but the selection of strategies takes allows the expression of much more complex decision
place through an evolutionary process. In this picture, ‘contents’ than we find among animals. The goal of a
seemingly rational strategies can be found in individual human group decision need not merely be to select one
animals not because these individuals explicitly chose option from a given set of options, but, instead, it can
them, but because their ancestors who happened to be to generate an explicit ranking of all the options in an
play these strategies received sufficient fitness benefits. order of collective preference (Arrow 1951/1963). In
At the risk of oversimplification, the difference such cases, the decision makers often do not merely
between a social-scientific, non-evolutionary under- cast a ‘vote’ for one option each, but express an entire
standing of rationality and a natural-scientific, evolution- ranking over options.
ary one lies in the place at which the rational choice or The options under consideration can also be
selection of strategies is located. In the non-evolutionary extremely complex. Committees, expert panels,
picture, it is the individual itself that makes rational multi-member courts and boards of organizations
choices. In the evolutionary picture, the seemingly frequently make choices between entire belief systems
rational selection of strategies takes place as a by-product or theories, with a complex internal structure. The
of an evolutionary process. More optimal strategies lead Intergovernmental Panel on Climate Change or the
to greater reproductive fitness. Individuals themselves, United Nations Development Programme, for
however, cannot be described as rational choosers. example, regularly produce extensive reports on some
complex natural or socio-economic phenomena and
(b) Language arrive at these reports through the interaction of a large
One of the most significant differences between human number of experts. The theory of judgement aggrega-
and non-human group decisions lies in the role that tion seeks to develop a general theoretical framework
language can, or cannot, play in such decisions. While for modelling how groups of individuals can make
humans and non-humans share the capacity both to consistent collective judgements on several, often
communicate prior to making a decision and to decide, logically connected propositions on the basis of the
by ‘voting’ or acting, to bring about a particular group members’ individual judgements on these
outcome, the nature of the communication and propositions (e.g. List & Pettit 2002; for an

introductory survey, see List in press). Familiar There are examples of both weighted majority
aggregation rules such as majority voting are not rule and explicit social-welfare-oriented decision-
generally satisfactory in such cases. For example, making arrangements in human societies. As already
assume that a group of city councillors has to decide mentioned, weighted majority rule is used in the
about three propositions: European Union Council of Ministers, where larger
p: ‘we will have £10 000 left over at the end countries, which may have a higher stake in many
decisions, have a greater voting weight than smaller
of the year’; countries. Implicit applications of weighted majority
rule can also take the form of the inclusion (a weight
p/ q : ‘if we have £10 000 left over at the of 1) or exclusion (a weight of 0) of certain individuals
end of the year; then we should renovate within or from the franchise (Fleurbaey 2008). For
the hospital’ example, in most democratic countries, citizens are
allowed to vote while visitors and temporary residents
and are not. Although many criticisms of this arrangement
could be raised, one rationale behind it might be that
q: ‘we should renovate the hospital:’ citizens have a higher stake in national decisions than
visitors or temporary residents.
If a third of the councillors believes p, p/q and q, a An explicit social-welfare-oriented decision-making
second third believes p/q, but also not p and not q, and arrangement might involve a ‘social planner’—e.g. a
a final third believes p, but not p/q, and not q, then government official or organization—who is instructed
there are majorities for p, for p/q and yet also for not q, to assess the welfare consequence of different policy
a logically inconsistent set of propositions. An import- options for the affected individuals and to make a
ant theoretical challenge is to provide good models of recommendation as to which policy maximizes overall
how real-world groups and committees avoid such welfare, according to the appropriate welfare standard
collective inconsistencies. In conclusion, it should be (Sen 1999). Often, the implementation of such a
apparent that the availability of language both enriches recommendation is further expected to be ‘incentive
and complicates group decision making. compatible’, meaning that whenever the policy
implementation involves situations of interactive/
(c) Optimality concepts combined decisions, the intended outcomes should
In the natural sciences, the principle of natural constitute equilibria (see, again, the survey article by
selection automatically introduces optimality concepts the Royal Swedish Academy of Sciences 2007).
(in a metaphor: ‘the survival of the fittest’) to group Finding good solutions to the kinds of decision
decision making (see also §6a). Natural selection is the problems such a social planner is faced with is a
process by which certain heritable units become significant challenge and an important topic within
relatively more common in successive generations of a welfare economics and the theory of mechanism
population of reproducing organisms due to differential design. However, such a topic is unlikely to find room
reproduction. However, the kinds of optimality con- for exploration in the natural sciences as such. As we
straints these processes adhere to are very different have seen, here, there is no human planner consciously
from those that play a role in human contexts. In the seeking to realize a previously defined goal. Instead, the
social sciences, moral criteria (e.g. fairness, justice or pursuit of some optimum is a by-product of the process
the achievement of the greatest ‘social welfare’) play an of natural selection.
important part in defining what the ‘optimum’ is.
These moral criteria are, as such, irrelevant to
evolutionary concepts. 7. CONCLUDING REMARKS
While natural selection can, in principle, occur on Perhaps the most striking observation that both the
different levels (e.g. genes, individuals, groups), social and the natural scientist have made while
genuine group-level selection (i.e. selection which preparing the present introduction is that, in current
cannot be explained equally well, or better, on a work on group decision making, the natural sciences
lower level) is likely to be rare (Smith 1964, 1976, are to some extent ‘reinventing the wheel’. Many
1998). Maximization of group-level pay-offs is there- concepts and mathematical tools that have been
fore unlikely to be a driving factor in the biological available in an advanced and sophisticated form in
evolution of group decision making. In many human the social sciences for some time are being redis-
contexts, by contrast, the maximization of group- covered, sometimes in a slightly different form, by
level pay-offs—or, less crudely, group-level or social natural scientists. This suggests that communication
‘welfare’—is a desired outcome. One way of achieving between the two fields could save natural scientists a
this outcome in decisions between two options, as we considerable amount of time. However, the social-
have seen, is to use weighted majority voting with scientific literature on group decisions is so vast that it is
weights proportional to stakes (at least when group- difficult for a natural scientist to digest this literature
level welfare is defined as the sum total of individual and to see the forest for all the trees. We hope that the
pay-offs). More generally, a substantial literature in present issue will help to open the door.
welfare economics and political philosophy addresses L.C. is supported by a Royal Society University Research
the question of how social welfare can be defined Fellowship, and would like to thank the Royal Society,
and measured (e.g. Arrow 1951/1963; Rawls 1971; particularly, for their support with respect to part-time work
Sen 1999). and maternity leave. C.L. is supported by a Philip

Leverhulme Prize and would like to thank the Leverhulme Boland, P. J. 1989 Majority systems and the Condorcet jury
Trust for its support. We are also grateful to Jason Alexander, theorem. Statistician 38, 181–189. (doi:10.2307/2348873)
Marc Fleurbaey, Mathias Koenig-Archibugi, Tim Roper and Boland, C. R. J. 2003 An experimental test of predator
Laura Valentini for their very helpful comments and advice. detection rates using groups of free-living emus. Ethology
Finally, we would like to thank all our anonymous referees for 109, 209–222. (doi:10.1046/j.1439-0310.2003.00860.x)
their great help in assessing and commenting on the Brosnan, S. F. & de Waal, F. B. M. 2003 Monkeys reject
contributions in this issue. unequal pay. Nature 425, 297–299. (doi:10.1038/
nature01963)
Brosnan, S. F., Schiff, H. C. & de Waal, F. B. M. 2005
Tolerance for inequity may increase with social closeness
REFERENCES in chimpanzees. Proc. R. Soc. B 272, 253–258. (doi:10.
Alexander, J. M. 2003 Evolutionary game theory. Stanford 1098/rspb.2004.2947)
encyclopedia of philosophy. See http://plato.stanford.edu/ Buhl, J., Sumpter, D. J. T., Couzin, I. D., Hale, J. J.,
entries/game-evolutionary/. Despland, E., Miller, E. R. & Simpson, S. J. 2006 From
Amé, J.-M., Halloy, J., Rivault, C., Detrain, C. & disorder to order in marching locusts. Science 312,
Deneubourg, J. L. 2006 Collegial decision making based 1402–1406. (doi:10.1126/science.1125142)
on social amplification leads to optimal group formation. Bukowski, M. & Miekisz, J. 2004 Evolutionary and
Proc. Natl Acad. Sci. USA 103, 5835–5840. (doi:10.1073/ asymptotic stability in symmetric multi-player games.
pnas.0507877103) Int. J. Game Theory 33, 41–54. (doi:10.1007/
Arrow, A. K. 1951/1963 Social choice and individual values. s001820400183)
New York, NY: Wiley. Byrne, R. W. 2000 How monkeys find their way: leadership,
Aube, F. & Shield, R. 2004 Modeling the effect of leadership coordination and cognitive maps of African baboons. In
on crowd flow dynamics. Cell. Automat. Proc. 3305, On the move: how and why animals travel in groups (eds
601–611. S. Boinski & P. A. Garber), pp. 491–518. Chicago, IL:
Austen-Smith, D. & Banks, J. S. 1999 Positive political theory I: University of Chicago Press.
collective preference. Ann Arbor, MI: University of Michigan Camazine, S., Deneubourg, J. L., Franks, N. R., Sneyd, J.,
Press. Theraulaz, G. & Bonabeau, E. 2003 Self-organization in
Austen-Smith, D. & Banks, J. S. 2005 Positive political theory biological systems. Princeton, NJ: Princeton University
II: strategy and structure. Ann Arbor, MI: University of Press.
Michigan Press. Carter, J., Lyons, N. J., Cole, H. L. & Goldsmith, A. R. 2008
Austen-Smith, D. & Feddersen, T. J. 2009 Information Subtle cues of predation risk: starlings respond to a
aggregation and communication in committees. Phil. predator’s direction of eye-gaze. Proc. R. Soc. B 275,
Trans. R. Soc. B 364, 763–769. (doi:10.1098/rstb.2008. 1709–1715. (doi:10.1098/rspb.2008.0095)
0256) Clutton-Brock, T. H. 1998 Reproductive skew, concessions
Bach, L. A., Helvik, T. & Christiansen, F. B. 2006 The and limited control. Trends Ecol. Evol. 13, 288–292.
evolution of n-player cooperation—threshold games and (doi:10.1016/S0169-5347(98)01402-5)
ESS bifurcations. J. Theor. Biol. 238, 426–434. Clutton-Brock, T. H., Guinness, F. E. & Albon, S. D. 1982
Ballerini, M. et al. 2008 Empirical investigation of starling Red deer: behaviour and ecology of two sexes. Chicago, IL:
flocks: a benchmark study in collective animal behaviour. University of Chicago Press.
Anim. Behav. 76, 201–215. (doi:10.1016/j.anbehav.2008. Codling, E. A., Pitchford, J. W. & Simpson, S. D. 2007 Group
02.004) navigation and the “many-wrongs principle” in models of
Ben-Yashar, R. & Nitzan, S. 1997 The optimal decision rule animal movement. Ecology 88, 1864–1870. (doi:10.1890/
for fixed-size committees in dichotomous choice situ- 06-0854.1)
ations: the general result. Int. Econ. Rev. 38, 175–186. Cohen, J. 1989 Deliberation and democratic legitimacy. In
(doi:10.2307/2527413) The good polity: normative analysis of the state (eds
Berend, D. & Sapir, L. 2007 Monotonicity in Condorcet’s A. Hamlin & P. Pettit), pp. 17–34. Oxford, UK: Blackwell.
jury theorem with dependent voters. Soc. Choice Welfare Conradt, L. 1998 Could asynchrony in activity between the
28, 507–528. (doi:10.1007/s00355-006-0179-y) sexes cause intersexual social segregation in ruminants?
Beshers, S. N. & Fewell, J. H. 2001 Models of division of Proc. R. Soc. B 265, 1359–1363. (doi:10.1098/rspb.1998.
labor in social insects. Annu. Rev. Entomol. 46, 413–440. 0442)
(doi:10.1146/annurev.ento.46.1.413) Conradt, L. & Roper, T. J. 2000 Activity synchrony and social
Bikhchandani, S., Hirshleifer, D. & Welch, I. 1992 A theory cohesion: a fission–fusion model. Proc. R. Soc. B 267,
of fads, fashions, custom, and cultural change as 2213–2218. (doi:10.1098/rspb.2000.1171)
informational cascades. J. Polit. Econ. 100, 992–1026. Conradt, L. & Roper, T. J. 2003 Group decision-making
(doi:10.1086/261849) in animals. Nature 421, 155–158. (doi:10.1038/nature01294)
Biro, D., Sumpter, D. J. T., Meade, J. & Guilford, T. 2006 Conradt, L. & Roper, T. J. 2005 Consensus decision making
From compromise to leadership in pigeon homing. Curr. in animals. Trends Ecol. Evol. 20, 449–456. (doi:10.1016/
Biol. 16, 2123–2128. (doi:10.1016/j.cub.2006.08.087) j.tree.2005.05.008)
Black, J. M. 1988 Preflight signaling in swans—a mechanism Conradt, L. & Roper, T. J. 2007 Democracy in animals: the
for group cohesion and flock formation. Ethology 79, evolution of shared group decisions. Proc. R. Soc. B 274,
143–157. 2317–2326. (doi:10.1098/rspb.2007.0186)
Blackstone, W. 1765–1769 Commentaries on the laws of Conradt, L. & Roper, T. J. 2009 Conflicts of interest and the
England. Oxford, UK: Clarendon Press. evolution of decision sharing. Phil. Trans. R. Soc. B 364,
Blackwell, P. G. 1997 The n-player war of attrition and 807–819. (doi:10.1098/rstb.2008.0257)
territorial groups. J. Theor. Biol. 189, 175–181. (doi:10. Conradt, L., Krause, J., Couzin I. D. & Roper, T. J. In press.
1006/jtbi.1997.0504) Leading according to need in animal groups. Am. Nat. 173.
Boinski, S. & Campbell, A. F. 1995 Use of trill vocalizations Courchamp, F., Rasmussen, G. S. A. & Macdonald, D. W.
to coordinate troop movement among white-faced 2002 Small pack size imposes a trade-off between hunting
capuchins—a 2nd field-test. Behaviour 132, 875–901. and pup-guarding in the painted hunting dog Lycaon pictus.
(doi:10.1163/156853995X00054) Behav. Ecol. 13, 20–27. (doi:10.1093/beheco/13.1.20)

Couzin, I. D. 2007 Collective minds. Nature 445, 715. Giraldeau, L. A., Valone, T. J. & Templeton, J. J. 2002
(doi:10.1038/445715a) Potential disadvantages of using socially acquired infor-
Couzin, I. D. & Krause, J. 2003 Self-organization and mation. Phil. Trans. R. Soc. B 357, 1559–1566. (doi:10.
collective behavior in vertebrates. Adv. Study Behav. 32, 1098/rstb.2002.1065)
1–75. (doi:10.1016/S0065-3454(03)01001-5) Gompper, M. E. 1996 Sociality and asociality in white-nosed
Couzin, I. D., Krause, J., Franks, N. R. & Levin, S. A. 2005 coatis (Nasua narica): foraging costs and benefits. Behav.
Effective leadership and decision-making in animal groups Ecol. 7, 254–263. (doi:10.1093/beheco/7.3.254)
on the move. Nature 433, 513–516. (doi:10.1038/ Goodin, R. E. & List, C. 2006a A conditional defense of
nature03236) plurality rule: generalizing May’s theorem in a restricted
Dawes, C. T., Fowler, J. H., Johnson, T., McElreath, R. & informational environment. Am. J. Polit. Sci. 50, 940–949.
Smirnov, O. 2007 Egalitarian motives in humans. Nature (doi:10.1111/j.1540-5907.2006.00225.x)
446, 794–796. (doi:10.1038/nature05651) Goodin, R. E. & List, C. 2006b Special majorities
Dennett, D. 1987 The intentional stance. Cambridge, MA: rationalized. Br. J. Polit. Sci. 36, 213–241. (doi:10.1017/
MIT Press. S0007123406000135)
Dietrich, F. & List, C. 2004 A model of jury decisions where Grofman, B., Owen, G. & Feld, S. L. 1983 Thirteen
all jurors have the same evidence. Synthese 142, 175–202.
theorems in search of the truth. Theory Decis. 15,
(doi:10.1007/s11229-004-1276-z)
261–278. (doi:10.1007/BF00125672)
Dostalkova, I. & Spinka, M. 2007 Synchronisation of
Gueron, S., Levin, S. A. & Rubenstein, D. I. 1996 The
behaviour in pairs: the role of communication and
dynamics of herds: from individuals to aggregations.
consequences in timing. Anim. Behav. 74, 1735–1742.
J. Theor. Biol. 182, 85–98. (doi:10.1006/jtbi.1996.0144)
(doi:10.1016/j.anbehav.2007.04.014)
Dryzek, J. S. & List, C. 2003 Social choice theory and Gutman, A. & Thompson, D. 1996 Democracy and disagree-
deliberative democracy: a reconciliation. Br. J. Polit. Sci. ment. Cambridge, MA: Harvard University Press.
33, 1–28. (doi:10.1017/S0007123403000012) Hastie, R. & Kameda, T. 2005 The robust beauty of majority
Dugatkin, L. A. 2005 Mistakes and the evolution of copying. rules in group decisions. Psychol. Rev. 112, 494–508.
Ethol. Ecol. Evol. 17, 327–333. (doi:10.1037/0033-295X.112.2.494)
Dyer, J. R. G. 2008 Leadership, decision making and Helbing, D., Farkas, I. & Vicsek, T. 2000 Simulating
collective behaviour in animal groups. PhD thesis, dynamical features of escape panic. Nature 407,
University of Leeds, UK. 487–490. (doi:10.1038/35035023)
Dyer, J. R. G., Johansson, A., Helbing, D., Couzin, I. D. & Hemelrijk, C. K. & Hildenbrandt, H. 2008 Self-organized
Krause, J. 2009 Leadership, consensus decision making shape and frontal density of fish schools. Ethology 114,
and collective behaviour in humans. Phil. Trans. R. Soc. B 245–254. (doi:10.1111/j.1439-0310.2007.01459.x)
364, 781–789. (doi:10.1098/rstb.2008.0233) Hix, S., Noury, A. & Roland, G. 2009 Voting patterns and
Elster, J. (ed.) 1998 Deliberative democracy, Cambridge, UK: alliance formation in the European Parliament. Phil.
Cambridge University Press. Trans. R. Soc. B 364, 821–831. (doi:10.1098/rstb.2008.
Fehr, E., Bernhard, H. & Rockenbach, B. 2008 Egalitarian- 0263)
ism in young children. Nature 454, 1079–1084. (doi:10. Ishaque, M. M. & Noland, R. B. 2008 Behavioural issues in
1038/nature07155) pedestrian speed choice and street crossing behaviour: a
Fischhoff, I. R., Sundaresan, S. R., Cordingley, J., Larkin, review. Trans. Rev. 28, 61–85. (doi:10.1080/014416407
H. M., Sellier, M. J. & Rubenstein, D. I. 2007 Social 01365239)
relationships and reproductive state influence leadership Kaminski, D., Miekisz, J. & Zaborowski, M. 2005 Stochastic
roles in movements of plains zebra, Equus burchellii. Anim. stability in three-player games. Bull. Math. Biol. 67,
Behav. 73, 825–831. (doi:10.1016/j.anbehav.2006.10.012) 1195–1205. (doi:10.1016/j.bulm.2005.01.004)
Fishkin, J. S. & Luskin, R. C. 2005 Experimenting with Kerth, G., Ebert, C. & Schmidtke, C. 2006 Group
a democratic ideal: deliberative polling and public decision making in fission–fusion societies: evidence
opinion. Acta Polit. 40, 284–298. (doi:10.1057/palgrave. from two-field experiments in Bechstein’s bats. Proc.
ap.5500121) R. Soc. B 273, 2785–2790. (doi:10.1098/rspb.2006.
Fleurbaey, M. 2008 Weighted majority and democratic 3647)
theory. Working paper, CNRS-CERSES, University King, A. J., Douglas, C. M. S., Huchard, E., Isaac, N. J. B. &
Paris-Descartes, LSE and IDEP. Cowlishaw, G. 2008 Dominance and affiliation mediate
Franks, N. R., Dechaume-Moncharmont, F.-X., Hanmore,
despotism in a social primate. Curr. Biol 18, 1833 –1838.
E. & Reynolds, J. K. 2009 Speed versus accuracy in
(doi:10.1016/j.cub.2008.10.048)
decision-making ants: expediting politics and policy
Krause, J. & Ruxton, G. D. 2002 Living in groups. Oxford,
implementation. Phil. Trans. R. Soc. B 364, 845–852.
UK: Oxford University Press.
(doi:10.1098/rstb.2008.0224)
Krause, J., Bumann, D. & Todt, D. 1992 Relationship
Gächter, S. & Herrmann, B. 2009 Reciprocity, culture and
human cooperation: previous insights and a new cross- between the position preference and nutritional state of
cultural experiment. Phil. Trans. R. Soc. B 364, 791–806. individuals in schools of juvenile roach (Rutilus rutilus).
(doi:10.1098/rstb.2008.0275) Behav. Ecol. Sociobiol. 30, 177–180. (doi:10.1007/
Gautrais, J., Michelena, P., Sibbald, A., Bon, R. & BF00166700)
Deneubourg, J. L. 2007 Allelomimetic synchronization Ladha, K. 1992 The Condorcet jury theorem, free speech
in merino sheep. Anim. Behav. 74, 1443–1454. (doi:10. and correlated votes. Am. J. Polit. Sci. 36, 617–634.
1016/j.anbehav.2007.02.020) (doi:10.2307/2111584)
Gavre, M. 1977 Essay on theory of enlightened despotism— Lingle, S. & Pellis, S. M. 2002 Fight or flight? Antipredator
Krieger, L. Am. Polit. Sci. Rev. 71, 1141–1142. (doi:10. behavior and the escalation of coyote encounters with
2307/1960129) deer. Oecologia 131, 154–164. (doi:10.1007/s00442-001-
Gehrlein, W. 1983 Condorcet’s paradox. Theory Decis. 15, 0858-4)
161–197. (doi:10.1007/BF00143070) List, C. 2004 Democracy in animal groups: a political science
Gigerenzer, G. & Selten, R. (eds) 2002 Bounded rationality: perspective. Trends Ecol. Evol. 19, 168–169. (doi:10.1016/
the adaptive toolbox. Cambridge, MA: MIT Press. j.tree.2004.02.004)

List, C. In press. Judgment aggregation: a short introduction. Regenwetter, M., Grofman, B., Popova, A., Messner, W.,
In Handbook of the philosophy of economics (ed. U. Maki). Davis-Stober, C. P. & Cavagnaro, D. R. 2009 Behavioural
Amsterdam, The Netherlands: Elsevier. ( http://personal. social choice: a status report. Phil. Trans. R. Soc. B 364,
lse.ac.uk/list/pdf-files/ja-intro.pdf ) 833–843. (doi:10.1098/rstb.2008.0259)
List, C. & Goodin, R. E. 2001 Epistemic democracy: Royal Swedish Academy of Sciences 2007 Mechanism design
generalizing the Condorcet jury theorem. J. Polit. Philos. theory: scientific background on the Sveriges Riksbank
9, 277–306. (doi:10.1111/1467-9760.00128) prize in economic sciences in memory of Alfred Nobel
List, C. & Pettit, P. 2002 Aggregating sets of judgments: an 2007, Stockholm. See http://nobelprize.org/nobel_prizes/
impossibility result. Econ. Philos. 18, 89–110. economics/laureates/2007/ecoadv07.pdf.
List, C., Luskin, R. C., Fishkin, J. & McLean, I. 2000/2006 Ruckstuhl, K. E. 1998 Foraging behaviour and sexual
Deliberation, single-peakedness, and the possibility of segregation in bighorn sheep. Anim. Behav. 56, 99–106.
meaningful democracy: evidence from deliberative polls. (doi:10.1006/anbe.1998.0745)
Working paper, London School of Economics. See http:// Ruckstuhl, K. E. 1999 To synchronise or not to synchronise:
personal.lse.ac.uk/list/PDF-files/DeliberationPaper.pdf. a dilemma for young bighorn males? Behaviour 136,
List, C., Elsholtz, C. & Seeley, T. D. 2009 Independence and 805–818. (doi:10.1163/156853999501577)
interdependence in collective decision making: an agent- Ruckstuhl, K. E. & Neuhaus, P. 2000 Sexual segregation in
based model of nest-site choice by honeybee swarms. Phil. ungulates: a new approach. Behaviour 137, 361–377.
Trans. R. Soc. B 364, 755–762. (doi:10.1098/rstb.2008. (doi:10.1163/156853900502123)
0277) Ruckstuhl, K. E. & Neuhaus, P. 2002 Sexual segregation in
Luskin, R. C., Fishkin, J. S. & Jowell, R. 2002 Considered ungulates: a comparative test of three hypotheses. Biol.
opinions: deliberative polling in Britain. Br. J. Polit. Sci. Rev. 77, 77–96.
32, 455–487. (doi:10.1017/S0007123402000194) Seeley, T. D. & Buhrman, S. C. 1999 Group decision making
Lusseau, D. 2007 Evidence for social role in a dolphin social in swarms of honey bees. Behav. Ecol. Sociobiol. 45, 19–31.
network. Evol. Ecol. 21, 357–366. (doi:10.1007/s10682- (doi:10.1007/s002650050536)
006-9105-0) Seeley, T. D. & Buhrman, S. C. 2001 Nest-site selection in
Lusseau, D. & Conradt, L. In press. The emergence of honey bees: how well do swarms implement the “best-
unshared consensus decisions in bottlenose dolphins. of-N” decision rule? Behav. Ecol. Sociobiol. 49, 416–427.
Behav. Ecol. Sociobiol. (doi:10.1007/s002650000299)
May, K. O. 1952 A set of independent, necessary and Sellers, W. I., Hill, R. A. & Logan, B. S. 2007 An agent-based
sufficient conditions for simple majority decision. Econo- model of group decision making in baboons. Phil. Trans.
metrica 20, 680–684. (doi:10.2307/1907651) R. Soc. B 362, 1699–1710. (doi:10.1098/rstb.2007.2064)
Maynard Smith, J. & Price, G. 1973 The logic of animal Selous, E. 1931 Thought-transference (or what?) in birds.
conflict. Nature 246, 15–18. (doi:10.1038/246015a0) London, UK: Constable & Co.
Meunier, H., Leca, J. B., Deneubourg, J. L. & Petit, O. 2006 Sen, A. K. 1999 The possibility of social choice. Am. Econ.
Group movement decisions in capuchin monkeys: the Rev. 89, 349–378.
utility of an experimental study and a mathematical model Simons, A. M. 2004 Many wrongs: the advantage of group
to explore the relationship between individual and navigation. Trends Ecol. Evol. 19, 453–455. (doi:10.1016/
collective behaviours. Behaviour 143, 1511–1527. j.tree.2004.07.001)
(doi:10.1163/156853906779366982) Skyrms, B. 2009 Evolution of signalling systems with
Milton, K. 2000 Tactics of food search and group movements multiple senders and receivers. Phil. Trans. R. Soc. B
in primates and other animals. In On the move: how and 364, 771–779. (doi:10.1098/rstb.2008.0258)
why animals travel in groups (eds S. Boinski & P. A. Garber), Smith, J. M. 1964 Group selection and kin selection. Nature
pp. 375–418. Chicago, IL: University of Chicago Press. 201, 1145–1147. (doi:10.1038/2011145a0)
Osborne, M. J. & Rubinstein, A. 1994 A course in game theory. Smith, J. M. 1976 Commentary—group selection. Q. Rev.
Cambridge, MA: MIT Press. Biol. 51, 277–283. (doi:10.1086/409311)
Passino, K. M. & Seeley, T. D. 2006 Modeling and analysis of Smith, J. M. 1998 Unto others: the evolution and psychology
nest-site selection by honeybee swarms: the speed and of unselfish behavior. Nature 393, 639–640. (doi:10.1038/
accuracy trade-off. Behav. Ecol. Sociobiol. 59, 427–442. 31383)
(doi:10.1007/s00265-005-0067-y) Stankowich, T. & Blumstein, D. T. 2005 Fear in animals: a
Platkowski, T. & Stachowska-Pietka, J. 2005 ESSs in n-player meta-analysis and review of risk assessment. Proc. R. Soc.
mixed games. Appl. Math. Comput. 167, 592–606. (doi:10. B 272, 2627–2634. (doi:10.1098/rspb.2005.3251)
1016/j.amc.2003.10.063) Stewart, K. J. & Harcourt, A. H. 1994 Gorillas vocalizations
Poole, J. H., Payne, K., Langbauer, W. R. & Moss, C. J. 1988 during rest periods—signals of impending departure.
The social contexts of some very low-frequency calls of Behaviour 130, 29–40. (doi:10.1163/156853994X00127)
African elephants. Behav. Ecol. Sociobiol. 22, 385–392. Sueur, C. & Petit, O. 2008 Shared or unshared consensus
(doi:10.1007/BF00294975) decision in macaques? Behav. Process. 78, 84–92. (doi:10.
Prins, H. H. T. 1996 Ecology and behaviour of the African 1016/j.beproc.2008.01.004)
buffalo. London, UK: Chapman and Hall. Sumpter, D. J. T. 2006 The principles of collective animal
Rae, D. W. 1969 Decision-rules and individual values in behaviour. Phil. Trans. R. Soc. B 361, 5–22. (doi:10.1098/
constitutional choice. Am. Polit. Sci. Rev. 63, 40–56. rstb.2005.1733)
(doi:10.2307/1954283) Sumpter, D. J. T. & Pratt, S. C. 2009 Quorum responses and
Rands, S. A., Cowlishaw, G., Pettifor, R. A., Rowcliffe, J. M. consensus decision making. Phil. Trans. R. Soc. B 364,
& Johnstone, R. A. 2003 Spontaneous emergence of 743–753. (doi:10.1098/rstb.2008.0204)
leaders and followers in foraging pairs. Nature 423, Sumpter, D., Buhl, J., Biro, D. & Couzin, I. 2008 Information
432–434. (doi:10.1038/nature01630) transfer in moving animal groups. Theory Biosci. 127,
Rawls, J. 1971 A theory of justice. Cambridge, MA: Harvard 177–186. (doi:10.1007/s12064-008-0040-1)
University Press. Sunstein, C. R. 2002 The law of group polarization. J. Polit.
Reebs, S. G. 2000 Can a minority of informed leaders Philos. 10, 175–195. (doi:10.1111/1467-9760.00148)
determine the foraging movements of a fish shoal? Anim. Sunstein, C. R. 2006 Infotopia: how many minds produce
Behav. 59, 403–409. (doi:10.1006/anbe.1999.1314) knowledge. New York, NY: Oxford University Press.

Taylor, M. 1969 Proof of a theorem on majority rule. Behav. Behavioural economics: An area of economics that seeks to
Sci. 14, 228–231. (doi:10.1002/bs.3830140307) explain economic phenomena by taking into account
Trillmich, J., Fichtel, C. & Kappeler, P. M. 2004 Coordi- empirical findings on the psychology of human decision
nation of group movements in wild Verreaux’s sifakas making; one of its key questions is whether, and to what
(Propithecus verreauxi ). Behaviour 141, 1103–1120. extent, economic agents satisfy, or violate, various classical
(doi:10.1163/1568539042664579) conditions of rationality.
Van Doorn, G. S., Hengeveld, G. M. & Weissing, F. J. 2003
Behavioural finance: An area of behavioural economics that
The evolution of social dominance—II: multi-player
studies how financial market behaviour is affected by the
models. Behaviour 140, 1333–1358. (doi:10.1163/
156853903771980611) psychology of human decision making.
Vermeule, A. 2005 Submajority rules: forcing accountability Behavioural game theory: An area of game theory that models
upon majorities. J. Polit. Philos. 13, 74–98. (doi:10.1111/ strategic behaviour on the basis of empirically informed
j.1467-9760.2005.00214.x) assumptions about human rationality; see also behavioural
Visscher, P. K. 2007 Group decision making in nest-site economics.
selection among social insects. Annu. Rev. Entomol. 52, Behavioural social choice theory: An area of social choice theory
255–275. (doi:10.1146/annurev.ento.51.110104.151025) that empirically tests social-choice-theoretic results and
Visscher, P. K. & Seeley, T. D. 2007 Coordinating a group their underlying assumptions.
departure: who produces the piping signals on honeybee Borda efficiency: The probability that the winning outcome
swarms? Behav. Ecol. Sociobiol. 61, 1615–1621. (doi:10. (top-ranked option) of a given aggregation rule coincides
1007/s00265-007-0393-3)
with the winner under the Borda rule, assuming that all
Wallraff, H. G. 1978 Social interrelations involved in
possible combinations of individual preferences are
migratory orientation of birds—possible contribution of
field studies. Oikos 30, 401–404. (doi:10.2307/3543490) equally probable.
Ward, A. J. W., Sumpter, D. J. T., Couzin, J. D., Hart, P. J. B. Borda rule/Borda count: An aggregation rule whose input is a
& Krause, J. 2008 Quorum decision making facilitates combination of individual preference orderings over some
information transfer in fish schools. Proc. Natl Acad. Sci. options and whose output is either a collective preference
USA 105, 6948–6953. (doi:10.1073/pnas.0710344105) ordering over these options or a top-ranked option,
defined as follows; each option gets a score from each
voter: if the option is ranked top among k options, it gets a
score of k; if it is ranked second from top, it gets a score of
APPENDIX A. GLOSSARY kK1, and so on; collectively, the option with the highest
Accuracy/reliability: The probability that a belief is correct; in sum total score comes top, the option with the second
the context of a decision in which there exists an highest comes second, and so on; for example, if 10
independently best outcome, the probability that this individuals have the preferences AOBOC and 25
outcome is reached. individuals have the preferences BOCOA, then A gets a
Acyclic binary relation: A binary relation, R, with the property score of 10!3C25!1Z55, B gets a score of 10!2C
that it is never the case that x 1Rx 2, x 2Rx 3, ., xk Rx 1; for 25!3Z95 and C gets a score of 10!1C25!2Z60;
example, ‘!’ is an acyclic binary relation, while ‘Z’ is not; consequently, the social preference is BOCOA.
preference relations are typically required to be acyclic; Complete/connected binary relation: A binary relation, R, with
Condorcet’s paradox, however, shows that majority the property that, for any x 1 and x 2, either x 1Rx 2 or x 2Rx 1
preferences may violate this requirement. (or both).
Aggregate/consensus decision: A single collective decision, e.g. Common values: The case in which different group members
between multiple options, that is ‘binding’ in some way for have identical interests and their differences are at most
all group members. informational.
Aggregation rule (sometimes also called voting rule): A function Condorcet’s jury theorem: If all members of a group have an
which assigns to each combination of individual inputs independent and equal accuracy/reliability better than
(e.g. votes) a resulting collective output (e.g. a decision random but less than perfect of making a correct judge-
outcome); different aggregation rules differ in what the
ment on some binary issue, then the majority judgement is
admissible inputs and outputs are; see, for example, social
more likely to be correct than any individual judgement
welfare functions.
and the probability of a correct majority judgement
Anonymity: A requirement that all individual group
converges to 1 as the group size increases.
members should be given equal weight in determining
Condorcet’s paradox: The phenomenon that majority prefer-
the outcome of an aggregate/consensus decision;
ences may be cyclic even when all individual preferences
anonymity is frequently imposed as a condition on
democratic aggregation rules, e.g. in May’s theorem. are acyclic; for example, if one-third of a group prefers A to
Antisymmetric binary relation: A binary relation, R, with B to C, a second third prefers B to C to A and the
the property that if x1Rx2 and x2Rx1, then x1Zx2; for remaining third prefers C to A to B, there are majorities for
example, ‘%’ is an antisymmetric binary relation. A against B, for B against C and for C against A.
Arrow’s impossibility theorem: A classic result in social choice Condorcet winner: An option which beats, or at least ties with,
theory showing that, in decisions between more than two all other options in pairwise majority voting.
alternatives, the only aggregation rules satisfying some Consensus cost: A difference between the fitness benefits which
minimal conditions (among which is the decomposition of a particular group member would have gained if the group
decisions into pairwise choices) and guaranteeing decision outcome had been the option that is optimal for
complete and transitive collective preferences are dictator- that member and the benefits gained in the realized
ial ones. aggregated/consensus decision outcome.
Behavioural decision theory: An area of decision theory that Coordination game: See box 2 for an example.
seeks to construct empirically informed models of human Culture of indifference: A generic term for probability
decision making. See also behavioural economics. distributions of individual preferences within a population

with particular symmetry properties; examples are the Fads, stock market bubbles: Examples of informational
impartial culture and the impartial anonymous culture. cascades.
Cyclic binary relation: A binary relation, R, with the property Game theory: A mathematical theory of interactive decision
that, for some set of options x 1, ., xk, we have x 1Rx 2, making; focuses on various kinds of strategic situations
x 2Rx 3, ., xkRx 1. (games) and models how rational players would behave in
Cyclic majority preference: See Condorcet’s paradox. them; investigates the existence and properties of different
Decision theory: A mathematical theory of individual decision kinds of equilibria in games.
making; focuses on various, often idealized, properties of Generalized weighted majority rule: See box 1.
individual rationality. Global overview: All members of a group can directly
Dictatorial rule/dictatorship: An aggregation rule whose output communicate with and/or observe all other members of
is always the input of a fixed individual. the group.
Direct reciprocity: An act of altruism towards an individual in Groupthink: The adoption of a particular viewpoint by a
the expectation of later repayment through a reciprocal act group as a result of conformism or the minimization of
of altruism by this individual; see also reciprocal altruism. conflict, without sufficient critical testing; related to
Dominant strategy equilibrium: A combination of strategies informational cascades.
across individuals in a game such that every individual Heuristics and biases: Rules of thumb in decision making
prefers (or is at least indifferent to) the outcome of (heuristics), which may lead to systematic errors (biases);
choosing his or her strategy, compared with the outcome a central topic in behavioural decision theory.
of deviating from it, regardless of what the other Impartial anonymous culture: A probability distribution of
individuals do; the situation in which all individuals defect individual preferences according to which all possible
in a Prisoner’s Dilemma is a dominant strategy frequencies across different preference orderings are
equilibrium. equally likely to occur; statistically, this is subtly different
Equally shared decision: A natural science term for a decision from an impartial culture.
with an aggregation rule in which all group members have Impartial culture: A probability distribution of individual
equal weights (anonymity). preferences according to which all possible preference
Equilibrium: A combination of strategies across individuals orderings are equally likely to occur.
that satisfies certain ‘best-response’ or ‘stability’ criteria; Indirect reciprocity: An act of altruism towards an individual in
different best-response or stability criteria lead to different the expectation of gaining a positive ‘reputation’ resulting
equilibrium concepts. in later repayment through altruistic acts by other
Eusociality/eusocial: Reproduction is confined to one or few individuals.
members of a colony and workers are functionally sterile; Informational cascade: A phenomenon in markets or other
mainly found in bees, wasps, ants and termites. information pooling settings where a plurality or majority
Evolutionarily stable strategy (ESS ): A strategy S such that, for that accidentally emerges in support of some proposition is
any other strategy T, either E(S, S)OE(T, S) or [E(S, S)Z mistakenly interpreted by others as evidence for the truth
E(T, S) and E(S, T)OE(T, T)], where E(A, B) is the of that proposition and thereby attracts further support,
pay-off of playing a strategy A against a strategy B; although few or any individuals originally judged the
the central consequence of this definition is that, if proposition to be true; examples of informational cascades
sufficiently many individuals in a population play an include fads or stock market bubbles.
evolutionarily stable strategy and pay-offs represent Informative voting: Casting a vote (e.g. in a jury decision) that
evolutionary fitness, no mutant strategy can successfully reveals one’s private information (e.g. about the guilt or
invade the population. innocence of the defendant).
Evolutionarily stable state: A dynamic evolutionary equilibrium Interactive/combined decision: A set of interdependent decisions
of a population; every population in which all individuals by group members affecting each other.
use an evolutionarily stable strategy is in an evolutionarily Linear order: A transitive, antisymmetric and complete/con-
stable state, but populations can also be in an evolution- nected binary relation; for example, ‘%’ is a linear order.
arily stable state if nobody uses an evolutionarily stable May’s theorem: In a two-option choice, majority voting is the
strategy; for example, a population with a proportion x of only aggregation rule that simultaneously satisfies uni-
‘hawks’ and 1Kx of ‘doves’ is in an evolutionarily stable versal domain, anonymity, neutrality and positive
state if the expected pay-offs for doves and hawks in responsiveness.
random pairings within the population are equal and Mechanism design theory: Investigates what mechanisms or
increase for hawks (decrease for doves) if x 0 !x, and vice systems of incentives induce rational individuals to behave
versa if x 0 Ox (and if no further alternative strategies to so as to bring about a particular intended outcome (e.g.
hawks and doves are biologically possible); here, neither sincere voting, truthful bidding in auctions).
hawks nor doves play evolutionarily stable strategies; see Mixed strategy: A strategy which can be seen as a lottery/
also box 2 for a further example. randomization over pure strategies; an individual has a
Expected utility theory: An area of decision theory in which mixed strategy if he or she has fixed probabilities p1, ., pk
individual decision making is modelled as the maximiza- (with kR1, piR0 for each i, and p1C/CpkZ1) of playing
tion of the expected value of some utility function. pure strategies S1, ., Sn, respectively.
Experimental economics: An area of economics in which Nash equilibrium: A combination of strategies across individ-
experiments (with real human subjects, e.g. volunteers uals in a game with the property that no individual
or college students) are used to test various hypotheses would prefer the outcome if it unilaterally deviated from
about human economic behaviour; to create realistic its strategy.
incentives, subjects usually receive monetary pay- Neutrality: Requires that all options should be treated equally
offs depending on their performance in the relevant in an aggregate/consensus decision; neutrality is frequently
strategic tasks. imposed as a condition on democratic aggregation rules,

e.g. in May’s theorem; super- or sub-majority rules, for also direct and indirect reciprocity), and (ii) a theory about
example, violate neutrality. how altruism could evolve in unrelated individuals.
Oligarchic rule: An aggregation rule whose output is Self-organization in groups of animals or humans: Emerging
determined by the inputs of a subset of the group group behaviour when individual group members behave
members; the limiting case of an oligarchic rule is a according to individual rules that are based on local
dictatorial rule (here, the subset of decisive group information and/or local communication and have some
members is a singleton). interdependence with the behaviour of neighbouring
Plurality rule: An aggregation rule in which each individual group members, but there is no individual that has a
casts one vote and the option with the largest number of global overview and directs the behaviour of the group as a
votes is selected. whole; a good example is a moving flock of starlings.
Positional voting rule: A class of aggregation rules based on the Social choice theory: A mathematical theory of collective
assignment of scores to options as a function of their decision making; focuses on various kinds of aggregation
position within individual preference orderings; the most problems and studies the properties of different aggrega-
prominent example is the Borda rule. tion rules, often using an axiomatic approach.
Positive responsiveness: A requirement that the output of an Social welfare function: Arrow’s (1951/1963) term for an
aggregate/consensus decision should be a positively aggregation rule whose input is a combination of
monotonic function of the individual inputs (e.g. votes); individual preference orderings over some given options
formally, if some option A wins or is tied with another and whose output is a single collective preference ordering
option B in pairwise voting, then any change of votes in
over these options; the term ‘social welfare’ comes from
favour of A should preserve A as the winner or break the tie
the fact that Arrow introduced this concept in the context
in favour of A; positive responsiveness is frequently placed
of welfare economics.
as a desideratum on democratic aggregation rules, e.g. in
Stake: See potential consensus costs.
May’s theorem.
Stake holders: All individuals affected by a particular decision.
Potential consensus costs (also called decision stake): A difference
Sub- or super-majority rule: A special case of a generalized
between the (fitness) benefits which a particular group
weighted majority rule in which the decision threshold is
member would gain if the group decision outcome were
tilted in favour of one and against the other option (and
the optimal option for that member and the benefits which
this member would gain otherwise. group members typically have equal weight).
Prisoner’s Dilemma: See box 2 for an example. Transitive binary relation: A binary relation, R, with the
Prospect theory/cumulative prospect theory: Prominent psycho- property that if x 1Rx 2 and x 2Rx 3, then x 1Rx 3. For
logically informed theories of human decision making example, ‘!’ is a transitive binary relation.
under risk. Universal domain: A requirement that any possible com-
Pure strategy: A strategy which involves no lottery/ binations of individual inputs (e.g. votes) should be
randomization. admissible in an aggregate/consensus decision; universal
Quorum response: A feedback mechanism in group decisions domain is frequently imposed as a desideratum on
whereby an individual’s probability of commitment to a democratic aggregation rules, e.g. in May’s theorem and
particular decision option increases sharply once a critical Arrow’s theorem.
number of other individuals (the ‘quorum threshold’) have Unshared decision: A natural science term for a decision with a
committed to that option. dictatorial aggregation rule.
Quorum threshold: See quorum response. Utility function: A function which assigns to each option a real
Reciprocal altruism: (i) An altruistic behaviour of one number, interpreted as the utility, a measure of desir-
individual towards another in the expectation of later ability, of that option; while a probability function
repayment through acts of altruism by the same, or other, represents beliefs or information, a utility function
individuals (‘if you scratch my back, I scratch yours’; see represents desires or interests.

Phil. Trans. R. Soc. B (2009) 364, 743–753

doi:10.1098/rstb.2008.0204
Quorum responses and consensus

decision making
David J. T. Sumpter1,* and Stephen C. Pratt2
1
Department of Mathematics, Uppsala University, PO Box 480, 75106 Uppsala, Sweden
2
School of Life Sciences, Arizona State University, PO Box 874501, Tempe, AZ 85287, USA
Animal groups are said to make consensus decisions when group members come to agree on the same
option. Consensus decisions are taxonomically widespread and potentially offer three key benefits:
maintenance of group cohesion, enhancement of decision accuracy compared with lone individuals
and improvement in decision speed. In the absence of centralized control, arriving at a consensus
depends on local interactions in which each individual’s likelihood of choosing an option increases
with the number of others already committed to that option. The resulting positive feedback can
effectively direct most or all group members to the best available choice. In this paper, we examine the
functional form of the individual response to others’ behaviour that lies at the heart of this process.
We review recent theoretical and empirical work on consensus decisions, and we develop a simple
mathematical model to show the central importance to speedy and accurate decisions of quorum
responses, in which an animal’s probability of exhibiting a behaviour is a sharply nonlinear function of
the number of other individuals already performing this behaviour. We argue that systems relying on
such quorum rules can achieve cohesive choice of the best option while also permitting adaptive
tuning of the trade-off between decision speed and accuracy.
Keywords: quorum responses; collective animal behaviour; Condorcet’s theorem;
social insect migration; decision making
1. INTRODUCTION everyone to choose an option measures decision

Group decision-making is characterized by individuals speed; and the proportion of individuals choosing the
making choices that rely on the decisions of others. One ‘best’ option gives the decision accuracy.
benefit of this interdependency is the maintenance of How does consensus arise from interactions
cohesion. Choosing the same destination taken by among group members, and how does individual
others, for example, can make an animal less likely to behaviour influence the cohesion, speed and accuracy
be picked out by a predator. Other potential benefits of decision making? In recent years, these questions
are in the speed and accuracy of an individual’s have been addressed by the theoretical and experi-
decisions, both of which can be improved by copying mental study of self-organization (Deneubourg &
the choice of a better-informed neighbour. This paper Goss 1989; Bonabeau et al. 1997; Camazine et al.
concerns group decisions in which cohesion, speed and 2001; Deneubourg et al. 2002; Sumpter 2006). In
accuracy are important factors. We will refer to these as general, self-organization explains how positive feed-
consensus decisions, defined as cases when all back created by imitative behaviour can generate
members of a group come to agree on the same option heterogeneous social patterns in uniform environ-
(Britton et al. 2002; Conradt & Roper 2005). ments. In the context of decision making, this implies
Consensus decisions are well illustrated by the that a group faced with a choice between two or more
choice of a shelter or nest site, and many experi- identical options can spontaneously and cohesively
mental studies have addressed this phenomenon choose only one of them. Self-organization can also
( Visscher & Camazine 1999; Pratt et al. 2002; address decision making when options clearly differ
Jeanson et al. 2004a; Seeley & Visscher 2004a; Ame in quality. For example, positive feedback provided
et al. 2006; Seeley et al. 2006; Visscher 2007). by pheromone trail recruitment allows ants to choose
Experimenters typically offer a group of animals a the shorter of two routes to a food source (Goss et al.
choice between two or more alternative shelters and 1989). Colonies of ants and honeybees (Apis
observe the process by which they make their choice. mellifera) can also direct their foragers to the better
A decision is assumed to have been made once all of two or more food sources, because recruitment
individuals have settled at a shelter. The degree to effectiveness is graded according to source quality
which individuals are aggregated at a single choice (Seeley et al. 1991; Sumpter & Beekman 2003).
gives a measure of their cohesion; the time taken for Quality-dependent recruitment differences similarly
underlie nest site selection in social insects (Mallon
et al. 2001; Franks et al. 2003b; Seeley 2003). These
* Author for correspondence (david@math.uu.se). studies show that positive feedback mediated by
One contribution of 11 to a Theme Issue ‘Group decision making in relatively simple interactions can allow social groups
humans and animals’. to make accurate consensus decisions.

744 D. J. T. Sumpter & S. C. Pratt Quorum responses and decision making
In this paper, we examine in detail a key feature of

1.0
probability that the majority is correct

consensus decisions, namely the functional form of an
individual’s response to others’ behaviour. We argue for
the central importance of quorum responses, in which an
animal’s probability of exhibiting a behaviour is a
sharply nonlinear function of the number of other
individuals already performing this behaviour. We first
0.8
review the theory for why and how consensus can yield
more accurate decisions than those of lone individuals.
We then describe a taxonomically diverse array of cases
in which quorum-like responses have been found to
underlie group decision-making. Next, we present a
simple mathematical model to investigate how the
0.6
functional form of the response to the behaviour of
others affects cohesion, accuracy and speed of decision 0 10 20 30 40 50 60 70 80 90 100
making. We show that the sharply nonlinear nature of a group size
quorum response allows cohesive choice of the best Figure 1. Condorcet’s theory. The probability that the
option while also permitting adaptive tuning of the majority of individuals are correct (for odd numbers of
inevitable trade-off between decision speed and individuals) when each is correct with probability pZ0.6.
accuracy. Finally, we investigate these ideas and
compare them with data using a more detailed model
of nest choice by Temnothorax ants. must be made to settle ties, but the overall shape of
the curve is unchanged. For groups of size 100, the
majority is almost never wrong, showing that majority
2. THE WISDOM OF CROWDS decisions are good way to pool information and
In his popular science book ‘The wisdom of crowds’, improve decision accuracy ( List 2004; King &
James Surowiecki gives a number of powerful examples Cowlishaw 2007).
of how a large group of poorly informed individuals Although, Condorcet’s theorem seems to provide a
can make better decisions than a small number of powerful method for groups to make correct decisions,
informed ‘experts’. A telling example is provided by it relies on two key assumptions—that individuals are
Galton (1907), who examined 800 entries in a ‘guess unbiased, and that they are independent. Both these
the weight of the ox competition’, where a crowd of assumptions must be treated with care. For example, if
fairgoers competed to guess how much a large ox would a group of navigating birds each follow an internal
weigh after slaughter. Although, the estimates varied compass with a consistent clockwise bias, then no
widely, their average value was only 1 pound (450 g) matter how many individual headings are averaged,
less than the true weight of 1197 pounds (544.5 kg). each will be similarly misled and the group decision will
Acting independently, the crowd ‘knew’ the weight of be inaccurate. Distinguishing variation due to random
the ox. There are many such examples of heightened error from that due to consistent bias can therefore
collective accuracy in humans, including the reliability pose a difficult problem.
of audience opinions on ‘Who wants to be a million- The second assumption of independent individual
aire’; the accurate prediction of American presidential choices presents a larger challenge. Indeed, this
elections by betting; and Google’s successful ranking assumption contradicts the very definition of group
of World Wide Web search results by the number of decision-making given in the first sentence of this
links to each website (Surowiecki 2004). paper—that individuals condition their own choices on
The collective wisdom argument was first formal- those of others. How can collective decisions preserve
ized by a French intellectual of the 18th century, the
independence but still come to a final consensus? In
Marquis de Condorcet (Borland 1989; List 2004;
human decision-making, this paradox lies at the basis
Austen-Smith & Feddersen 2009). He considered
of ‘groupthink’ ( Janis 1972, 1982). Groupthink occurs
binary choices between two options, in which each
when the pressures of group members on one another
individual has a probability p of making a correct
narrow down the range of opinions. It is most likely
decision in the absence of others with which to confer.
In this situation, one can apply the binomial theorem to when group members have similar backgrounds and
find the probability that the majority of the individuals interests. Janis (1972) proposed that groupthink can be
are correct. Assuming that an odd number of prevented by allowing a large number of individuals to
individuals n must each make a decision independently first collect information independently before present-
of one another, then the probability that the majority ing their recommended course of action to a smaller
make the correct choice is number of centralized evaluators. By correctly weight-
! ing these independent recommendations, itself no easy
Xn n i task, the evaluators can arrive at an average of the
mðn; pÞ Z p ð1K pÞi :
i opinions presented. While effective for humans, this
i Z nC1
2
solution demands complex information-processing
Figure 1 plots this function for pZ0.6. As the number mechanisms that may not be available to animal
of individuals goes to infinity, m(n,p)/1 and the societies. We now turn our attention to how these
majority decision is always correct. If n is even a rule groups can solve the problem of groupthink.

Quorum responses and decision making D. J. T. Sumpter & S. C. Pratt 745
3. POSITIVE FEEDBACK AND QUORUM (a) 8

RESPONSES
The collective behaviour of animal groups is often 7
individual per second (×10–3)

probability of leaving per
decentralized, with no leader integrating different 6
sources of information or telling the others what to
do (Seeley 1995, 2002). Instead, a pattern emerges 5
from a large number of strictly local interactions that 4
carry information throughout the group. A key feature
of these interactions is positive feedback, in which an 3
animal’s probability of exhibiting a particular behaviour 2
is an increasing function of the number of conspecifics
already performing this behaviour (Deneubourg & 1
Goss 1989; Bonabeau et al. 1997). In the context of
collective decision-making, positive feedback allows 0 2 4 6 8 10 12
the selection of a particular option to cascade through no. of non-self cockroaches under shelter
the group, as the growing number of adherents to an
option increases its attractiveness to undecided (b) 1.0
animals. Moreover, this imitative behaviour often
takes a step-like form, with an individual’s probability
probability of transport
0.8
of selecting an option changing sharply when the
number of like-minded conspecifics crosses a
threshold. Here we refer to this functional form as a 0.6
quorum response, following well-studied cases in
which threshold group sizes trigger key changes in 0.4
behaviour (Pratt et al. 2002; Seeley & Visscher 2004b).
0.2
(a) Cockroach aggregation
Various species of cockroach benefit from increased
growth rates when in aggregations ( Prokopy & 0 10 20 30 40 50 60 70
Roitberg 2001). German cockroaches (Blattella germa- no. of ants in nest
nica) can reduce water loss in dry conditions by
clustering together (Dambach & Goehlen 1999) and Figure 2. Examples of empirical quorum responses in the
typically gather in dark shelters during the daytime decisions of migrating insects. (a) Cockroaches. Crosses
indicate measured leaving times, dashed line is fit given by
(Ishii & Kuwahara 1968; Rivault 1989). Ame et al.
Ame et al. (2006) of
(2004) tested the contribution of social interaction to
these aggregations. They presented a group of cock- q
a :
roaches with two identical shelters, each with sufficient 1 C r xK1
S
capacity to shelter all the insects. In the majority of
with parameter values SZ40, qZ0.01, rZ1667 and aZ2
trials over 80 per cent of the insects chose the same and solid line is the best fit of the equation
shelter. Thus even in the absence of a difference
between the two options a consensus is reached for only q
fC a ;
one of them. 1 C r xK1
S
Consensus is reached through a very simple rule: an with parameter values SZ40, 4Z0.00051, qZ0.0067,
individual’s probability of leaving a shelter decreases as rZ1667 and aZ1.73. This second fitted line allows for the
the shelter’s population increases. The probability fact that the probability of leaving does not go to zero with
drops quite sharply with population, giving rise to a the number under the shelter. (b) A quorum rule governs the
step like quorum response (figure 2a). By incorporating probability of a Temnothorax scout switching from tandem run
this quorum rule into models of cockroach behaviour, recruitment of fellow scouts to faster transport of the bulk of
Ame et al. (2004) showed that it could explain the colony. Crosses show proportions of scouts choosing
consensus shelter choice. A disproportional response transport over tandem runs at different populations under
high urgency. Open circles show corresponding data under
to the presence of other cockroaches was the key
low urgency. Solid and dashed lines, respectively, show a
element. Ame et al. (2006) fitted the function Hill function fit to these data: probability of transportZ
q xk/(xkCTk), where x is the new site population.
a ;
1 C r Sx
to the probability per second per cockroach of leaving a shelter increases more than linearly with the number of
shelter, where x is the number of cockroaches under the cockroaches under the shelter. This prediction accorded
shelter (figure 2a). The parameters determine the shape with the value of az2 measured from the experiments.
of the response: q is the rate at which cockroaches leave Further investigation of the model shows that provided
an unoccupied shelter; r and S determine the density at that aO1, even a relatively weak positive response to the
which cockroaches respond to conspecifics and a presence of conspecifics is sufficient to generate a
determines the steepness of this response. The model consensus (Millor et al. 2006). It was thus the sharply
predicted that a consensus will be reached for one of the nonlinear reaction to others—the quorum response—
shelters as long as aO1, that is, the time spent in the that generated a collective decision.

(b) Nest site selection by social insects called tandem running. These recruited ants in turn
For many social insects, the survival of the colony make their own independent assessments and may also
depends crucially upon remaining together and making begin to recruit, a process that gradually increases the
a good decision about where to live. This is especially population of ants visiting the site. Once the scouts
true when colonies live in preformed cavities, such as perceive their site’s population to have reached a
honeybees nesting in tree cavities and Temnothorax ants threshold, they enter the final phase of full commitment
in rock crevices or hollow nuts. These colonies have (Pratt et al. 2002) (figure 2b). They abandon tandem
limited opportunities to repair a poor initial choice, but runs from the old nest in favour of speedier transports,
must instead live with the consequences or emigrate to by which the passive majority of the colony’s workers,
a new home. Emigration is especially costly for as well as the queens and brood, are brought to the new
honeybees, because they have to abandon their site (Pratt et al. 2005).
investment in comb construction, brood-rearing and Despite the many differences between honeybee and
food storage. A poor initial choice can therefore greatly ant emigration, their nest site selection relies on a
reduce a colony’s reproductive success. fundamentally similar strategy. There is no require-
Honeybee emigration usually occurs in spring, when ment for direct comparison of multiple sites by well-
the queen and a swarm of roughly 10 000 worker bees informed insects. Instead, scouts aware of only a single
leave their old nest and temporarily settle in a densely- candidate site recruit to it with a strength that depends
packed swarm. Several hundred scout bees then fly out on their independent assessment of its quality. Because
to search for a new home. Successful scouts use the the recruited scouts themselves recruit, this generates
waggle dance to recruit fellow scouts to the sites they positive feedback on site populations that is stronger for
have found. Recruited bees may in turn dance for a site, better sites. This advantage is then amplified by a
creating a positive feedback loop that drives up the quorum rule that accelerates movement to the site with
population of scouts visiting a site. Bees tune their the fastest early population growth. Owing to the
dancing to the quality of the site they are advertising, quality-dependent recruitment advantage, this will
hence better sites enjoy more effective recruitment and usually be a superior site.
faster population growth (Seeley & Buhrman 1999;
Seeley & Visscher 2004a). Scouts periodically return to (c) Other insects and spiders
the site they are advertising and somehow assess its Together with various colleagues, Jean-Louis Deneu-
population. Once this exceeds a threshold value, or bourg has shown that a variety of gregarious arthropods
quorum, they return to the swarm to perform a respond to a choice between two identical options by
behaviour called piping (Seeley & Visscher 2003, randomly selecting one of them (Deneubourg et al.
2004b). Piping induces the thousands of non-scout 2002). Repeated over many experimental trials, this
bees to warm their flight muscles in preparation for the leads to a U-shaped distribution of outcomes, with
swarm to fly to the new nest site, guided by the minority roughly half of the groups unanimously choosing each
of knowledgeable scouts (Seeley et al. 2003). This option, and very few splitting between them. Examples
process unfolds over one to several days, during which a include selection between feeders by foraging ants
large number of sites are found and advertized by at (Goss et al. 1989; Beckers et al. 1993; Jeanson et al.
least a few bees. Usually, only one site reaches quorum 2004a), between settlement locations by social spiders
and induces swarm lift off, but rare split decisions have (Saffre et al. 2000; Jeanson et al. 2004b) and between
been observed, in which the bees engage in an aerial escape routes for ants fleeing a disturbance (Altshuler
tug-of-war as rival groups of scouts attempt to lead the et al. 2005). Positive feedback is seen in each of these
swarm in different directions. In these cases, the bees cases: ants grow more likely to join a foraging trail as its
are forced to re-settle and begin the process again concentration of recruitment pheromone increases;
(Lindauer 1955, 1961). spiders are more likely to follow a route to a settlement
Ants of the genus Temnothorax form much smaller location as it is reinforced with the silk strands of other
colonies than honeybees, typically with no more than spiders; and escaping ants are more likely to take an exit
100–200 individuals. Colonies can be easily kept in chosen by many nest-mates. All of these cues increase
artificial nests and induced to emigrate in the in strength with the number of other individuals that
laboratory. They typically move within a few hours, have already selected that option. Moreover, the
reliably choosing the best site from as many as five function relating joining probability to cue strength is
alternatives that they discriminate according to cavity sharply nonlinear, or quorum-like. These empirical
area, ceiling height, entrance size, light level and other observations demonstrate a basic property of all
features (Pratt & Pierce 2001; Franks et al. 2003b). collective decision-making: positive feedback together
Approximately 30 per cent of a colony’s workers with nonlinear quorum responses lead to U-shaped
actively partake in the selection process. These active choice distributions and consensus decisions.
ants go through four phases of graded commitment to
any potential new home (Pratt et al. 2005). Each ant (d) Birds, fish and primates
begins in an exploration phase during which she For vertebrate groups migrating over long distances,
searches for nest sites. After finding one, she enters consensus building may improve navigational accuracy.
an assessment phase in which she evaluates its quality. The analogue to Condorcet’s theorem in this case is the
The length of this phase is inversely related to the theory of many wrongs (Wallraff 1978; Simons 2004).
quality of the site (Mallon et al. 2001), and is followed This theory assumes that each animal has imprecise
by a canvassing phase during which the ant leads fellow information about the route to its target, and shows
scouts to the site, using a slow recruitment method that averaging these estimates allows the group to reach

consensus on a more accurate path. Biro et al. (2006) individuals can observe the choices of others in order
showed that interactions between a pair of homing to improve their decision-making accuracy.
pigeons (Columba livia) were important in determining We begin with a group of n individuals not
their navigational route. When conflict between routes committed to either option. Each of these finds one
was small the birds followed an average of the two, but of the two options with a constant probability r per time
when conflict was large one bird led and the other step. This probability is independent of the actions of
followed. Pairs of pigeons flew more direct routes home others. If an individual arrives at an option and no one
than did solo birds. This result is consistent with the else is there, then she commits to it with the probability
many wrong hypothesis, but it could also be explained apx for option X and apy for option Y. If an individual
by birds flying more ‘confidently’ when in pairs. arrives at an option and other individuals are present,
Experiments on larger groups would be needed to say the probability of her committing and remaining at the
whether quorums play a role, but in other contexts option is an increasing function of the number already
birds do make choices based on threshold responses to commited. Specifically, if x is the committed number at
conspecific numbers (Collins & Sumpter 2007). the option then the probability that the arriving
Ward et al. (2008) showed clear use of quorum-like individual commits is
rules by fish making binary movement decisions in the
presence of replica ‘leader’ fish. They found that fish xk
px a C ðmKaÞ k ; ð4:1Þ
chose a movement direction as a function of group size T Cxk
and the number of fish (or replicas) going left and right.
The probability of following in a particular direction where a and m are, respectively the minimum and
was a steeply increasing function of the number already maximum probability of committing; T is the quorum
moving in that direction. Ward et al. (2008) further threshold at which this probability is halfway between a
showed that if two or three replica fish swam past a and m; and k determines the steepness of the function.
replica predator then the group of fish could be induced A similar function determines the probability of
to follow, despite the fact that lone fish would seldom selecting option Y and by setting pxOpy , we assume
pass the same replica predator. that individuals prefer X to Y.
Despite their relatively high cognitive abilities, the Equation (4.1) includes a range of possible
movement decisions of capuchin monkeys (Cebus responses to conspecifics. If kZ1 then the probability
capucinus) have also proven consistent with simple of an individual choosing an option is proportional to
copying of the decisions of others (Meunier et al. the number that have already made that choice. If kO1
2007). Their response is not quorum-like: the prob- then equation (4.1) has a point of inflection and the
ability of following increased in proportion to the function is sigmoidal. As k increases the response
number taking a particular direction. In general, the approaches a step-like switch at the threshold T.
movement decisions of primate groups may depend on In order to define a quorum response, we first
dominance hierarchies, past experience and complex consider a purely linear response function
social structure (Boinski & Garber 2000). However, the x
interactions of these monkeys provide evidence that px a C ðmKaÞ ; ð4:2Þ
2T
simple copying should not be ruled out as an
explanation of complex movement decisions. which shares with equation (4.1) the property that
when xZT the probability of committing is half way
between m and a. We define a quorum response to be
4. ACCURACY THROUGH QUORUM RESPONSES one in which the probability of committing is always
Why are quorum responses such a ubiquitous feature of less than the linear response whenever the number of
group decision-making? In particular, why do individ- conspecifics is less than T and is greater or equal to that
ual response probabilities change sharply when a of the linear response for some number of conspecifics
threshold is exceeded rather than varying in proportion greater or equal to T. This definition captures the
to the stimulus? A first answer to these questions is concept of a less than linear response to numbers below
given by several theoretical models that show how the threshold and a greater than linear response above
quorum responses generate cohesion ( Nicolis & the threshold. By identifying conditions under which
Deneubourg 1999; Millor et al. 2006). This effect is our linear equation is equal to equation (4.1), we find
seen empirically in the U-shaped distributions of that a quorum response occurs if only if kR2 (figure 3).
groups choosing between two identical options. We note, however, that it may be equally valid to argue
However, cohesion is just one of the three desirable that the existence of a point of inflection defines a
properties of consensus decision-making. The others quorum response, so that quorum responses occur for
we quoted in the introduction are accuracy and speed, kO1. The important biological point is that quorum
to which we can add the ability to adjust the trade-off responses involve a sharply increasing nonlinear
between these two properties. Here we investigate all response to the conspecifics.
these aspects within the framework of a simple quorum The above model demands very limited cognitive
response model. powers on the part of individuals. In particular, they have
no way of directly comparing the two options. We assume
(a) Quorum response model that rejecting one option does not increase an individual’s
We developed a simple model of how a population of probability of accepting the other. The population
partially informed individuals chooses between two already committed gives individuals an indirect method
options. This model is designed to look at how to gather information about available options.

0.9 investigate how different values for k, T and a affect

k=9 speed and accuracy, we systematically varied these
0.8 parameters and measured their affect on the time
probability of commiting
0.7
needed for all individuals to make a choice and the
k=2 proportion choosing the better option (figure 5). The
0.6 results show that speed is maximized by setting a to its
k=1
maximum value of 1 (assuming that mZ1 as well).
0.5
Greater speed, however, comes at the expense of more
0.4 individuals choosing the worse option. Accuracy is
maximized with low a, high k and T of approximately
0.3 10, but these values also produce relatively slow
0.2 decisions. Thus, for a given quorum threshold, the
trade-off between speed and accuracy can be tuned by
0.1 altering the base acceptance probability, a.
0 2 4 6 8 10 12 14 16 18 20 The quorum threshold, T, has more complex effects
no. of committed individuals: x than does a. For large k, T can be also be used to tune
speed and accuracy. For example, when kZ4 or 9,
Figure 3. Commitment to an option as a function of the
number of conspecifics that have already chosen it (x). The decision speed is maximized for TZ0, but accuracy is
dashed line shows the purely linear response given by maximized when Tz10. However, for a wide range of
equation (4.2). The solid lines show nonlinear responses threshold values (T between approximately 5 and 15),
given by equation (4.1), for different values of k. For kO2 relatively small differences in choice quality produce
equation (4.1) gives a quorum response: that is, the high levels of commitment to the better option.
probability of committing is less than the linear response for There is also an important difference between a and
x!Tand greater than or equal to the linear response for xRT. T in how speed and accuracy change when one
Other parameters are pxZ1, TZ10, aZ0.1 and mZ0.9. parameter is fixed and the other varied. If T is chosen
to maximize accuracy (e.g. Tz10 when kZ9) a can be
(b) Model simulation
Figure 4a,b give examples of the choices over time of tuned to achieve either the maximum possible accuracy
(over all tested combinations of T and a values) or the
nZ40 individuals for shallow proportional responses
maximum possible speed (i.e. by choosing aZ1). The
(TZ10 and kZ1) and steep quorum responses (TZ10
same is not the case for fixed a and varying T. If a
and kZ9), respectively. For both types of responses,
is large then tuning T can do little to improve the
the proportion of committed individuals grows
resulting low accuracy; if a is small then setting TZ0
slowly for the two options, but slightly faster for the
improves speed but not as much as would setting a to
preferred option X. After the number of adherents to X
a value of 1. Thus by choosing appropriate values of
reaches the threshold T, commitment to X significantly
T and k, and adjusting a as needed, individuals can
outpaces commitment to Y. Averaged over 1000
tune the speed and accuracy of their decisions to
simulations, 75.5 per cent of individuals choose X for
particular circumstances.
a shallow response, while 83.3 per cent do so for the
The simulations also showed that tuning speed
steep quorum response. In both cases the proportion
and accuracy with a works best with an intermediate
choosing the better option is higher than that were
threshold value and a steep quorum response (high k).
each to make an independent decision, in which case
For fixed k, we determined the parameter values of a
px/( pxCpy)Z66.7 per cent would be expected to
and T that give the fastest possible average time until a
choose X. Thus, in these simulations choices based
decision is made given a minimum requirement for
on copying others reduce individual errors and make
accuracy (figure 6). When the requirement for accuracy
group decision-making more accurate than indepen-
is low, a similarly high speed can be achieved for any
dent assessment alone.
value of k, by choosing appropriate values for T and a.
While a steep quorum response led on average to
For higher accuracy requirements, however, a k value
more accurate decisions, the distribution of decision-
of 1 leads to distinctly slower attainable speeds. Thus
making accuracy is wider for kZ9 than for kZ1
steep thresholds not only give more accurate decisions,
(figure 4c,d ). This observation reflects the amplifi-
they also allow them to be made more rapidly.
cation of small initial errors for steep responses. If,
through random fluctuations, the least favourable
option happens to be chosen by more than a threshold (d) Comparison to Condorcet’s theorem
number of individuals, then the quorum rule amplifies Given 40 individuals, each with a 1/3 probability of
these early errors and nearly all individuals make the making the wrong choice, then by Condorcet’s
same incorrect choice. theorem, the probability of a majority error is just
3.33 per cent. This is notably lower than even the most
accurate decisions made using quorum responses: for
(c) Speed-accuracy trade-off steep thresholds between 5 and 15 and low spon-
Decision makers typically face a trade-off between taneous accept rates, approximately 10 per cent of
speed and accuracy. In the simulations, a steep quorum individuals take the least favourable option. This result
function (kZ9) yielded a more accurate decision, but is not particularly surprising. Condorcet’s theorem
the time taken for all individuals to choose was longer provides an upper bound for the accuracy of collective
on average (307.8G71.0 time steps, meanGs.d.) than decision-making. What is striking is that a simple
when kZ1 (253.7G64.0 time steps). In order to copying rule based on threshold responses can

(a) 40 (b)
no. of individuals
30
20
10
0 100 200 300 400 500 0 100 200 300 400 500
time time
(c) (d)
0.30
proportion of groups
0.25
0.20
0.15
0.10
0.05
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
proportion at X proportion at X
Figure 4. Simulations of a simple quorum response model, for (a,c) shallow (kZ1) and (b,d ) steep (kZ9) thresholds. (a,b) plot
the change in the number of individuals committed to options X, solid line; and Y, dotted line for one simulation with kZ1 and
kZ9, respectively. (c,d ) show the distribution taken over 1000 simulation runs of the proportion of individuals choosing X after
everyone has decided. Other parameters are rZ0.02, pxZ1, pyZ0.5, TZ10, aZ0.1 and mZ0.9.
substantially reduce errors compared with purely these circumstances, colonies moved much faster but
independent decision-making. often made poor choices, splitting their population
between the two candidate nests or even moving
entirely into the inferior one.
5. SPEED VERSUS ACCURACY TRADE-OFFS IN We have previously developed a detailed agent-
ANT MIGRATION based model of Temnothorax emigration (Pratt et al.
The decision making of animal groups can be 2005). This agent-based model is more complex than
considerably more complicated than a simple threshold the general quorum model described earlier, but both
response to the decisions of others. We described earlier include the same fundamental mechanisms: an intrin-
the complex, multistage algorithm used by Temnothorax sic rate of accepting an option that depends on that
ants to evaluate candidate nest sites during colony option’s quality, and a quorum function described by
emigration. Progress through four stages of increasing parameters for threshold value (T ) and steepness (k).
commitment to a site is governed both by each scout’s Furthermore, both models make similar predictions for
independent assessment of site quality and by the the effects of T and the acceptance rate on speed and
indirect influence of her nest-mates, via a quorum rule accuracy: for a wide range of T values, the acceptance
(figure 2b). Complicating this basic structure are a host rate provides a sensitive mechanism for adjusting speed
of behavioural nuances, including ‘reverse’ recruitment and accuracy. The model predicted that ants achieve a
of scouts from the new to the old nest, direct speed/accuracy trade-off by quantitative tuning the
comparison of multiple sites by individual ants, acceptance rate and, to a lesser degree, the quorum
changes in the efficiency of recruitment with time and threshold (Pratt & Sumpter 2006). The small effect of
many others (Pratt et al. 2005; Pratt & Sumpter 2006). the quorum threshold is at first surprising, because one
Experiments have shown that this complex algo- might suppose that the reaching of a threshold marks
rithm allows colonies to tune the trade-off between the point at which transportation can commence and
decision speed and accuracy (Pratt & Sumpter 2006). the emigration can be completed. However, as Franks
When choosing between a good and a mediocre nest, et al. (2009) rightly point out in another paper in this
colonies showed dramatically different behaviour issue, reaching the threshold too soon can result in an
depending on the urgency of their need to move. In insufficient number of committed ants to complete the
the low-urgency situation colonies in an intact but transportation of ants from the old nest.
poor-quality nest had an opportunity to improve their Our agent-based model was not previously
housing. They took a long time to emigrate, but they examined for effects of k, so we systematically varied
generally made very accurate decisions, moving their this parameter and monitored its effect on emigration
entire population directly to the better candidate nest. speed and accuracy. The results match those for the
Greater urgency was created by destroying the colony’s simpler model, with greater accuracy as quorum
old nest, leaving them completely exposed. Under steepness increases, and little cost in speed (figure 7).

(a) (b) (c) (d )

20 500
quorum threshold: T
15 400
300
10
200
5
100
(e) (f ) (g) (h)

20
quorum threshold: T
0.3
15
10 0.2
5
0.1
–1.0 – 0.5 0 –1.0 – 0.5 0 –1.0 –0.5 0 –1.0 –0.5 0
spontaneous accept probability: log (a)
Figure 5. Speed and accuracy of decision making for the simple quorum response model. Predicted effects of the parameters a,
T and k on (a–d ) the time until all individuals have made a decision and (e–h) the accuracy of that decision. In each image, a
and T are varied for different threshold steepness, k. The plots show mean duration (time steps of the model) and accuracy
(proportion of individuals choosing the less attractive option Y) over 1000 simulations for each parameter combination.
(a) kZ1; (b) kZ2; (c) kZ4; (d ) kZ9; (e) kZ1; ( f ) kZ2; (g) kZ4; (h) kZ9.
8 6. DISCUSSION
Quorum responses are a ubiquitous feature of con-
sensus decision-making. While previous work has
speed: 1/(×10 –3 time taken for all
7
emphasized the importance of these responses in
generating aggregation and cohesion, here we have
individuals to decide)
6 emphasized that they also improve decision accuracy.

The shape of the response curve is particularly
5 important in this context. Individuals can make more
accurate decisions if they sharply increase their
4 probability of committing to an option at a threshold
number of individuals already committed. Interest-
ingly, these steep threshold responses can sometimes
3
amplify random fluctuations and lead to mass adoption
of incorrect choices. This sort of process may account
2 for observations of mass copying (Laland & Williams
0.65 0.70 0.75 0.80 0.85 1998; Dall et al. 2005) or peer pressure in humans
accuracy: proportion of individuals choosing better option (Milgram et al. 1969; Milgram 1992) and may lead
Figure 6. Speed–accuracy trade-off for the simple quorum animals to make decisions in groups they would not
response model. For fixed k and for a fixed minimum have made by themselves. Although, quorum responses
requirement for decision accuracy, we searched over all lead to poor decisions in some notable cases, on average
parameter values of a and T which give the fastest possible they allow greater accuracy than do complete indepen-
average time until all individuals have made a decision. This dence or weak responses to the behaviour of others.
was done repeatedly for different minimum accuracy Another important property of quorum responses is
requirements to give a speed versus accuracy trade-off. that they can be used to tune speed and accuracy. By
Solid line, kZ9; dashed line, kZ4; dashed-dotted line, fixing a steep threshold and then tuning the baseline
kZ2; dotted line, kZ1.
rate at which an option is accepted, decisions can be
In accordance with these predictions, our experiments made either more accurately or more quickly. The
showed that ants made dramatic increases in accep- same is not true in the absence of a threshold, where
tance rate, and smaller decreases in T, in response to reducing baseline acceptance slows decision making
increased urgency of emigration (Pratt & Sumpter but does little to increase accuracy. Temnothorax ants
2006). Re-analysis of this data further shows that ants take advantage of this property to tune their decision
also used a significantly steeper quorum function when making for speed or accuracy (Pratt & Sumpter 2006).
accuracy was emphasized under low urgency (ANOVA: Our simple model suggests that many other animals
1
k lowUrgencyZ3.7, k highUrgencyZ1.7, F415 Z 10, exhibiting quorum responses may also be able to tune
p!0.01). These experiments provide strong evidence their decisions in this way.
of the ants tune their responses to their speed versus Other studies have emphasized the precise tuning of
accuracy requirements without changing their under- quorum size itself for the balancing of decision
lying behavioural algorithm. speed and accuracy, either over evolutionary time

(a) (b) (c) (d)

12
quorum threshold: T 10
7500
8 5000
6
4 2500
2 0
(e) (f) (g) (h)

12
quorum threshold: T
0.3
10
8 0.2
6 0.1
4
0
2
–2 –1 0 1 –2 –1 0 1 –2 –1 0 1 –2 –1 0 1
accept rate: log (accept)
Figure 7. Trade-off of decision speed and accuracy for an agent-based model of Temnothorax emigrations. Predicted effects of the
parameters accept, quorum threshold and k on the (a–d ) duration and (e–h) accuracy of emigrations. In each image, the intrinsic
accept rate and quorum threshold are varied for different threshold steepness, k. The plots show mean duration and accuracy
over 32 simulations for each parameter combination. All other parameters are set to values estimated as described in Pratt
(2005). Accept gives the recruitment initiation rate at good nests; the rate for mediocre nests was obtained by multiplying by the
factor 0.52 (the ratio of observed values of accept for mediocre and good nests). (a) kZ1; (b) kZ2; (c) kZ4; (d ) kZ8; (e) kZ1;
( f ) kZ2; ( g) kZ4; (h) kZ8.
(Passino & Seeley 2006), or dynamically in response to accurate. Because quorum responses are clearly used
the changing conditions experienced by a society by animals with conflicting interests, the effect of this
(Franks et al. 2003a). Our results suggest instead that conflict on quorum parameter values remains as an
the quorum size may not require tight regulation or exciting theoretical and experimental challenge.
have a particularly large direct influence on speed and
accuracy. As long as individuals employ a quorum rule,
the threshold can vary quite widely with little effect,
REFERENCES
and the group can achieve both accuracy and Altshuler, E., Ramos, O., Nunez, Y., Fernandez, J., Batista-
tunability, by adjusting the more sensitive acceptance Leyva, A. J. & Noda, C. 2005 Symmetry breaking in
parameter. Nonetheless, as discussed above for escaping ants. Am. Nat. 166, 643–649. (doi:10.1086/
Temnothorax, there is empirical evidence that individ- 498139)
uals change their quorum size and steepness according Ame, J. M., Rivault, C. & Deneubourg, J. L. 2004 Cockroach
to circumstances ( Franks et al. 2003a; Dornhaus et al. aggregation based on strain odour recognition. Anim.
2004; Pratt & Sumpter 2006). Thus a functional role Behav. 68, 793–801. (doi:10.1016/j.anbehav.2004.01.009)
for tuning the quorum cannot be ruled out. Ame, J. M., Halloy, J., Rivault, C., Detrain, C. &
An important question that we have not addressed in Deneubourg, J. L. 2006 Collegial decision making based
this paper is conflict in consensus decision-making on social amplification leads to optimal group formation.
Proc. Natl Acad. Sci. USA 103, 5835–5840. (doi:10.1073/
(Conradt & Roper 2005, 2009; Wood & Acland 2007;
pnas.0507877103)
Sumpter et al. 2008). The models presented here
Austen-Smith, D. & Feddersen, T. J. 2009 Information
assume no conflict of interest between group members aggregation and communication in committees. Phil.
and that the inherent tendency to lead or follow others Trans. R. Soc. B 364, 763–769. (doi:10.1098/rstb.2008.
does not vary between individuals. These are reason- 0256)
able assumptions for many insect societies, but are less Beckers, R., Deneubourg, J. L. & Goss, S. 1993 Modulation
likely to hold for the movements of more loosely of trail laying in the ant Lasius niger (Hymenoptera,
associated vertebrate groups. A first step to incorporat- Formicidae) and its role in the collective selection of a food
ing conflict would be to test the evolutionary stability of source. J. Insect Behav. 6, 751–759. (doi:10.1007/
quorum responses; that is, to determine whether selfish BF01201674)
individuals could exploit the quorum parameter values Biro, D., Sumpter, D. J. T., Meade, J. & Guilford, T. 2006
that optimize group accuracy to improve their own From compromise to leadership in pigeon homing. Curr.
Biol. 16, 2123–2128. (doi:10.1016/j.cub.2006.08.087)
accuracy. For example, by waiting until everyone else
Boinski, S. & Garber, P. A. 2000 On the move: how and why
has made a decision, an individual might be able to animals travel in groups. Chicago, IL: University of Chicago
maximize its own probability of making an accurate Press.
choice. This strategy that should evolve when each Bonabeau, E., Theraulaz, G., Deneubourg, J. L., Aron, S. &
individual aims to increase its own performance Camazine, S. 1997 Self-organization in social insects.
without regard to the outcome for others might Trends Ecol. Evol. 12, 188–193. (doi:10.1016/S0169-
produce group decisions that are neither fast nor 5347(97)01048-3)

Borland, P. J. 1989 Majority systems and the Condorcet jury Jeanson, R., Deneubourg, J. L. & Theraulaz, G. 2004b
theorem. Statistician 38, 181–189. (doi:10.2307/2348873) Discrete dragline attachment induces aggregation in
Britton, N. F., Franks, N. R., Pratt, S. C. & Seeley, T. D. 2002 spiderlings of a solitary species. Anim. Behav. 67,
Deciding on a new home: how do honeybees agree? Proc. R. 531–537. (doi:10.1016/j.anbehav.2003.06.013)
Soc. B 269, 1383–1388. (doi:10.1098/rspb.2002.2001) King, A. J. & Cowlishaw, G. 2007 When to use social
Camazine, S., Deneubourg, J. L., Franks, N. R., Sneyd, J., information: the advantage of large group size in
Theraulaz, G. & Bonabeau, E. 2001 Self-organization in individual decision making. Biol. Lett. 3, 137–139.
biological systems. Princeton studies in complexity. Prince- (doi:10.1098/rsbl.2007.0017)
ton, NJ: Princeton University Press. Laland, K. N. & Williams, K. 1998 Social transmission of
Collins, L. M. & Sumpter, D. J. T. 2007 The feeding maladaptive information in the guppy. Behav. Ecol. 9,
dynamics of broiler chickens. J. R. Soc. Interface 4, 65–72. 493–499. (doi:10.1093/beheco/9.5.493)
(doi:10.1098/rsif.2006.0157) Lindauer, M. 1955 Schwarmbienen auf Wohnungssuche.
Conradt, L. & Roper, T. J. 2005 Consensus decision making Z. Vgl. Physiol. 37, 263–324. (doi:10.1007/BF00303153)
in animals. Trends Ecol. Evol. 20, 449–456. (doi:10.1016/ Lindauer, M. 1961 Communication among social bees.
j.tree.2005.05.008) New York, NY: Atheneum.
Conradt, L. & Roper, T. J. 2009 Conflicts of interest and the List, C. 2004 Democracy in animal groups: a political science
evolution of decision sharing. Phil. Trans. R. Soc. B 364, perspective. Trends Ecol. Evol. 19, 168–169. (doi:10.1016/
807–819. (doi:10.1098/rstb.2008.0257) j.tree.2004.02.004)
Dall, S. R. X., Giraldeau, L. A., Olsson, O., McNamara, Mallon, E. B., Pratt, S. C. & Franks, N. R. 2001 Individual
J. M. & Stephens, D. W. 2005 Information and its use by and collective decision-making during nest site selection
animals in evolutionary ecology. Trends Ecol. Evol. 20, by the ant Leptothorax albipennis. Behav. Ecol. Sociobiol. 50,
187–193. (doi:10.1016/j.tree.2005.01.010) 352–359. (doi:10.1007/s002650100377)
Dambach, M. & Goehlen, B. 1999 Aggregation density and Meunier, H., Leca, J. B., Deneubourg, J. L. & Petit, O. 2007
longevity correlate with humidity in first-instar nymphs of Group movement decisions in capuchin monkeys: the
the cockroach Blattella germanica. J. Insect Physiol. 45, utility of an experimental study and a mathematical
423–429. (doi:10.1016/S0022-1910(98)00141-3) model to explore the relationship between individual and
Deneubourg, J. L. & Goss, S. 1989 Collective patterns and collective behaviours. Behaviour 143, 1511–1527. (doi:10.
decision-making. Ethol. Ecol. Evol. 1, 295–311. 1163/156853906779366982)
Deneubourg, J. L., Lioni, A. & Detrain, C. 2002 Dynamics of Milgram, S. 1992 The individual in the social world. New York,
aggregation and emergence of cooperation. Biol. Bull. 202, NY: McGraw-Hill.
Milgram, S., Bickman, L. & Berkowitz, L. 1969 Note on the
262–267. (doi:10.2307/1543477)
drawing power of crowds of different size. J. Pers. Soc.
Dornhaus, A., Franks, N. R., Hawkins, R. M. & Shere,
Psychol. 13, 79–82. (doi:10.1037/h0028070)
H. N. S. 2004 Ants move to improve: colonies of
Millor, J., Ame, J. M., Halloy, J. & Deneubourg, J. L. 2006
Leptothorax albipennis emigrate whenever they find a
Individual discrimination capability and collective
superior nest site. Anim. Behav. 67, 959–963. (doi:10.
decision-making. J. Theor. Biol. 239, 313–323. (doi:10.
1016/j.anbehav.2003.09.004)
1016/j.jtbi.2005.07.020)
Franks, N. R., Dornhaus, A., Fitzsimmons, J. P. & Stevens,
Nicolis, S. C. & Deneubourg, J. L. 1999 Emerging patterns
M. 2003a Speed versus accuracy in collective decision
and food recruitment in ants: an analytical study. J. Theor.
making. Proc. R. Soc. B 270, 2457–2463. (doi:10.1098/
Biol. 198, 575–592. (doi:10.1006/jtbi.1999.0934)
rspb.2003.2527) Passino, K. M. & Seeley, T. D. 2006 Modeling and analysis of
Franks, N. R., Mallon, E. B., Bray, H. E., Hamilton, M. J. & nest-site selection by honeybee swarms: the speed and
Mischler, T. C. 2003b Strategies for choosing between accuracy trade-off. Behav. Ecol. Sociobiol. 59, 427–442.
alternatives with different attributes: exemplified by (doi:10.1007/s00265-005-0067-y)
house-hunting ants. Anim. Behav. 65, 215–223. (doi:10. Pratt, S. C. 2005 Behavioral mechanisms of collective nest-
1006/anbe.2002.2032) site choice by the ant Temnothorax curvispinosus. Insect. Soc.
Franks, N. R., Dechaume-Moncharmont, F.-X., Hanmore, 52, 383–392. (doi:10.1007/s00040-005-0823-z)
E. & Reynolds, J. K. 2009 Speed versus accuracy in Pratt, S. C. & Pierce, N. E. 2001 The cavity-dwelling ant
decision-making ants: expediting politics and policy Leptothorax curvispinosus uses nest geometry to discrimi-
implementation. Phil. Trans. R. Soc. B 364, 845–852. nate between potential homes. Anim. Behav. 62, 281–287.
(doi:10.1098/rstb.2008.0224) (doi:10.1006/anbe.2001.1777)
Galton, F. 1907 Vox populi. Nature 75, 450–451. (doi:10. Pratt, S. C. & Sumpter, D. J. T. 2006 A tunable algorithm for
1038/075450a0) collective decision-making. Proc. Natl Acad. Sci. USA 103,
Goss, S., Aron, S., Deneubourg, J. L. & Pasteels, J. M. 1989 15 906–15 910. (doi:10.1073/pnas.0604801103)
Self-organized shortcuts in the Argentine ant. Naturwis- Pratt, S. C., Mallon, E. B., Sumpter, D. J. T. & Franks, N. R.
senschaften 76, 579–581. (doi:10.1007/BF00462870) 2002 Quorum sensing, recruitment, and collective
Ishii, S. & Kuwahara, Y. 1968 Aggregation of German decision-making during colony emigration by the ant
cockroach Blattella germanica nymphs. Experientia 24, Leptothorax albipennis. Behav. Ecol. Sociobiol. 52, 117–127.
88–89. (doi:10.1007/BF02136814) (doi:10.1007/s00265-002-0487-x)
Janis, I. L. 1972 Victims of groupthink. New York, NY: Pratt, S. C., Sumpter, D. J. T., Mallon, E. B. & Franks, N. R.
Houghton Mifflin. 2005 An agent-based model of collective nest choice by
Janis, I. L. 1982 Groupthink: psychological studies of policy the ant Temnothorax albipennis. Anim. Behav. 70,
decisions and fiascoes, 2nd edn. New York, NY: Houghton 1023–1036. (doi:10.1016/j.anbehav.2005.01.022)
Mifflin. Prokopy, R. J. & Roitberg, B. D. 2001 Joining and avoidance
Jeanson, R., Deneubourg, J. L., Grimal, A. & Theraulaz, G. behaviour in nonsocial insects. Annu. Rev. Entomol. 46,
2004a Modulation of individual behavior and collective 631–665. (doi:10.1146/annurev.ento.46.1.631)
decision-making during aggregation site selection by the Rivault, C. 1989 Spatial distribution of the cockroach,
ant Messor barbarus. Behav. Ecol. Sociobiol. 55, 388–394. Blattella germanica, in a swimming-bath facility. Entomol.
(doi:10.1007/s00265-003-0716-y) Exp. Appl. 53, 247–255. (doi:10.1007/BF00162856)

Saffre, F., Mailleux, A. C. & Deneubourg, J. L. 2000 Seeley, T. D., Visscher, P. K. & Passino, K. M. 2006 Group
Exploratory recruitment plasticity in a social spider decision making in honey bee swarms. Am. Sci. 94,
(Anelosimus eximius). J. Theor. Biol. 205, 37–46. (doi:10. 220–229.
1006/jtbi.2000.2042) Simons, A. M. 2004 Many wrongs: the advantage of group
Seeley, T. D. 1995 The wisdom of the hive. Cambridge, MA: navigation. Trends Ecol. Evol. 19, 453–455. (doi:10.1016/
Belknap Press of Harvard University Press. j.tree.2004.07.001)
Seeley, T. D. 2002 When is self-organization used in Sumpter, D. J. T. 2006 The principles of collective animal
biological systems? Biol. Bull. 202, 314–318. (doi:10. behaviour. Phil. Trans. R. Soc. B 361, 5–22. (doi:10.1098/
2307/1543484) rstb.2005.1733)
Seeley, T. D. 2003 Consensus building during nest-site Sumpter, D. J. T. & Beekman, M. 2003 From nonlinearity to
selection in honey bee swarms: the expiration of dissent. optimality: pheromone trail foraging by ants. Anim. Behav.
Behav. Ecol. Sociobiol. 53, 417–424. 66, 273–280. (doi:10.1006/anbe.2003.2224)
Seeley, T. D. & Buhrman, S. C. 1999 Group decision making Sumpter, D. J. T., Buhl, J., Biro, D. & Couzin, I. D. 2008
in swarms of honey bees. Behav. Ecol. Sociobiol. 45, 19–31. Information transfer in moving animal groups. Theory
(doi:10.1007/s002650050536) Biosci. 127, 177–186. (doi:10.1007/s12064-008-0040-1)
Surowiecki, J. 2004 The wisdom of crowds. London, UK:
Seeley, T. D. & Visscher, P. K. 2003 Choosing a home: how
Little, Brown.
the scouts in a honey bee swarm perceive the completion
Visscher, P. K. 2007 Group decision making in nest-site
of their group decision making. Behav. Ecol. Sociobiol. 54,
selection among social insects. Annu. Rev. Entomol. 52,
511–520. (doi:10.1007/s00265-003-0664-6)
255–275. (doi:10.1146/annurev.ento.51.110104.151025)
Seeley, T. D. & Visscher, P. K. 2004a Group decision making
Visscher, P. K. & Camazine, S. 1999 Collective decisions
in nest-site selection by honey bees. Apidologie 35, and cognition in bees. Nature 397, 400. (doi:10.1038/
101–116. (doi:10.1051/apido:2004004) 17047)
Seeley, T. D. & Visscher, P. K. 2004b Quorum sensing Wallraff, H. G. 1978 Social interrelations involved in
during nest-site selection by honeybee swarms. Behav. migratory orientation of birds—possible contribution of
Ecol. Sociobiol. 56, 594–601. (doi:10.1007/s00265-004- field studies. Oikos 30, 401–404. (doi:10.2307/3543490)
0814-5) Ward, A., Sumpter, D. J. T., Couzin, I. D., Hart, P. J. B. &
Seeley, T. D., Camazine, S. & Sneyd, J. 1991 Collective Krause, J. 2008 Quorum decision-making facilitates
decision-making in honey bees: how colonies choose information transfer in fish shoals. Proc. Natl Acad.
among nectar sources. Behav. Ecol. Sociobiol. 28, Sci. USA 105, 6948–6953. (doi:10.1073/pnas.
277–290. (doi:10.1007/BF00175101) 0710344105)
Seeley, T. D., Kleinhenz, M., Bujok, B. & Tautz, J. 2003 Wood, A. J. & Acland, G. J. 2007 Evolving the selfish herd:
Thorough warm-up before take-off in honey bee swarms. emergence of distinct aggregating strategies in an individ-
Naturwissenschaften 90, 256–260. (doi:10.1007/s00114- ual-based model. Proc. R. Soc. B 274, 1637–1642. (doi:10.
003-0425-4) 1098/rspb.2007.0306)

Phil. Trans. R. Soc. B (2009) 364, 755–762

doi:10.1098/rstb.2008.0277
Independence and interdependence in collective

decision making: an agent-based model of
nest-site choice by honeybee swarms
Christian List1,*, Christian Elsholtz2 and Thomas D. Seeley3
1
Department of Government, London School of Economics, London WC2A 2AE, UK
2
Department of Mathematics, Royal Holloway, University of London, Egham, Surrey TW20 0EX, UK
3
Department of Neurobiology and Behavior, Cornell University, Ithaca, NY 14853, USA
Condorcet’s jury theorem shows that when the members of a group have noisy but independent
information about what is best for the group as a whole, majority decisions tend to outperform
dictatorial ones. When voting is supplemented by communication, however, the resulting
interdependencies between decision makers can strengthen or undermine this effect: they can
facilitate information pooling, but also amplify errors. We consider an intriguing non-human case of
independent information pooling combined with communication: the case of nest-site choice by
honeybee (Apis mellifera) swarms. It is empirically well documented that when there are different nest
sites that vary in quality, the bees usually choose the best one. We develop a new agent-based model of
the bees’ decision process and show that its remarkable reliability stems from a particular interplay of
independence and interdependence between the bees.
Keywords: group decision making; honeybees; nest-site choice; Condorcet’s jury theorem;
information pooling; agent-based model
1. INTRODUCTION decision makers. On the one hand, these may facilitate

Since the Marquis de Condorcet’s work in the eight- information pooling and filtering (e.g. Luskin et al. 2002;
eenth century it is known that, when the members of a Farrar et al. in press), but on the other hand, they may also
group have only noisy and partially reliable information lead to the amplification of certain errors, such as in fads
about what is best for the group as whole, democratic and informational cascades, as briefly discussed at the end
decisions tend to outperform dictatorial ones. Con- of this paper (e.g. Bikhchandani et al. 1992; Zuber et al.
dorcet showed that, if each member of a jury has an 1992; Sunstein 2002, 2006).
equal and independent chance better than random, but In this paper, we consider an intriguing non-human
worse than perfect, of making a correct judgement on case of information pooling combined with communi-
whether a defendant is guilty, the majority of jurors is cation: the case of nest-site choice by honeybee (Apis
more likely to be correct than each individual juror. mellifera) swarms. We present a new theoretical model
Moreover, the probability of a correct majority judge- of the honeybees’ collective decision process and
ment approaches certainty as the jury size increases. This investigate the role played by both information pooling
result is a consequence of the law of large numbers: from and communication in it.
many independent but noisy signals, majority voting can It is a long-standing empirical fact that, in late spring
extract the information while filtering out the noise. or early summer, a colony of honeybees that has reached a
This insight, which has become known as ‘Condorcet’s certain size tends to divide itself: the queen leaves with
jury theorem’, has sparked a large body of social scientific roughly two-thirds of the worker bees, and a daughter
work on the reliability of various decision procedures in queen stays behind in the parental nest with the rest of the
juries, committees, legislatures, electorates and other worker bees. How does the swarm that has left the colony
settings (e.g. Grofman et al. 1983; Borland 1989; find a new home? Empirical work by Lindauer (1955)
Austen-Smith & Banks 1996; List & Goodin 2001; List and Seeley et al. (2006) have revealed a mechanism
2004). While the original theorem highlights the benefits involving a ‘search committee’ of several hundred bees,
of pooling independent information held by multiple the scouts, who fly out to inspect potential nest sites and
individuals, a complexity arises in collective decisions then come back and perform waggle dances to advertise
when voting is supplemented by communication, as any good sites they have discovered. Initially, the scouts
visit and inspect sites randomly, but once the dancing
investigated by the theory of deliberative democracy
activity has built-up, they are more likely to visit and
(e.g. Elster 1986; Miller 1992; Knight & Johnson 1994;
inspect sites advertised by others. Back at the swarm,
Dryzek & List 2003; Austen-Smith & Feddersen 2009).
each bee dances for the site she has inspected, with the
Communication can create interdependencies between
duration of the dance depending on her perception of the
* Author for correspondence (c.list@lse.ac.uk). site’s quality: the better the site, the longer the dance.
One contribution of 11 to a Theme Issue ‘Group decision making in Thus, high-quality sites receive more advertisement and
humans and animals’. are visited by more scout bees, which in turn generate

756 C. List et al. Independence and interdependence
even more dance activity for these sites. The process are k potential nest sites, labelled 1, 2, ., k, where each
eventually leads to a ‘consensus’: the dancing and visiting nest site j has an objective quality qjR0. We assume
concentrates on one popular site, and the swarm moves discrete time periods, labelled 1, 2, 3, ., and explicitly
there. (The bees’ final decision to move appears to model the behaviour of all n individual scout bees in
involve ‘quorum sensing’, as discussed by Seeley & each period.
Visscher 2003.) The striking empirical fact is that, when At each time, a scout bee can be in one of two states:
different possible nest sites vary in quality, the bees either she dances for one of the k potential nest sites or
usually choose the best one (Seeley & Buhrman 2001). she does not dance for any site, which can mean that
While the empirical details of this process are well she has not yet found a site, she has flown out to search
understood, the mechanisms underlying its striking for sites, she is observing other bees, or she is resting.
reliability still lack a full explanation. Our model of the Formally, the state of bee i at time t is represented by a
bees’ decision making is innovative in combining two two-component vector xi,tZ(si,t ,di,t), where
features: first, it is agent based, in the sense that we
explicitly model the individual behaviour of each scout — si,t2{0, 1, 2, ., k} is the site for which bee i dances
as a simple stochastic process, and second and more at time t, with si,tZ0 meaning that she does not
importantly, it integrates insights from Condorcet’s dance at time t, and
jury theorem with those from the theory of deliberative — di,tR0 is the remaining duration of bee i’s dance at
democracy. Using computer simulations based on this time t.
model, we are able to predict that, under a wide range
of parameter conditions, a consensus among the bees We initialize the model by assuming there is no
emerges for the best nest site with a high probability, dancing activity at time 1, i.e. for all i, xi,1Z(0,0).
even when the quality differences between sites are
relatively small. Furthermore, we show that the remark- (b) How each bee changes her state from one time
able reliability of the bees’ decision process stems from period to the next
a particular interplay of independence and interdependence As in any agent-based model, the state of bee i at time
between the bees, as defined formally below. tC1 depends on her own state at time t and the state of
Other mathematical models of nest-site choice by other bees at time t. We need to distinguish between
honeybees are a differential equation model by Britton two cases: either bee i does not dance for any site at
et al. (2002), a matrix model by Myerscough (2003), time t, in which case she may or may not fly out and find
another agent-based model by Passino & Seeley (2006), a site to dance for at time tC1. Or she already dances
and a density-dependent Markov process model by for one of the sites at time t, in which case she continues
Perdriau & Myerscough (2007). While shedding light her dance at time tC1 unless its duration is over.
on several important aspects of the bees’decision process, We now discuss each case in turn.
none of these models exhibits both main characteristics of Case 2.1. Bee i does not dance for any site at time t
ours, i.e. being agent based and explicitly modelling the (i.e. si,tZ0).
interplay between independence and interdependence.
Moreover, our model is particularly parsimonious and In this case, she has a certain probability of flying to
makes very robust predictions. Since our computational one of the k sites and inspecting it. For each site j, we write
results appear to be consistent with existing empirical pj,tC1 to denote the probability that the bee finds site j and
findings about the bees, we suggest that our model dances for it at time tC1. Further, p0,tC1 denotes the
adequately captures some key elements of the bees’ probability that the bee remains at rest or finds no site, so
decision-making process. that she does not dance at time tC1. Thus, the first
The paper is structured as follows. After a formal component of the bee’s state at time tC1, namely si,tC1,
exposition of our model, we state our two main takes the values 1, 2, ., k (one of the sites) or 0 (no site)
hypotheses about the bees’ decision process. The first with probabilities p1,tC1, p2,tC1, ., pk,tC1 and p0,tC1,
is, roughly, that this process is robustly reliable for a large respectively. By definition, these probabilities add up to 1.
class of parameter conditions; and the second that the How are the probabilities determined? The prob-
presence of both independence and interdependence ability that a bee finds a given site depends on two
between individual bees is necessary and sufficient for the factors: first, an a priori probability of how likely she is
overall reliability. Methodologically, both hypotheses are to find that site without any advertisement by other
formulated as hypotheses about our model of the bees bees (this may depend on the site’s location, distance
rather than as hypotheses about the real world bees from the swarm, etc.) and second, the proportion of
themselves; but to the extent that the model behaviour bees dancing for it. Formally, for each j (including the
is consistent with the empirically observed behaviour of case jZ0 of no site), we define
the bees, our hypotheses can be considered empirically pj;tC1 Z ð1KlÞpj C lfj;t ;
adequate as well. To provide a computational test of our
hypotheses, we finally report our computer simulations, where pj is the a priori probability of the jth site; fj,t is the
followed by a brief concluding discussion. proportion of bees dancing for site j at time t; and l is the
relevant weight, ranging between 0 and 1. The weight l
captures the amount of interdependence between the bees.
2. MODEL If lZ0, each bee’s probabilities of finding the various
(a) Basic ingredients of the model sites remain the a priori probabilities, regardless of how
There are n scout bees, labelled 1, 2, ., n, who many bees dance for them: this is the limiting case in
participate in the decision-making process, and there which the bees do not influence each other at all through

Independence and interdependence C. List et al. 757
communication. If lZ1, by contrast, each bee’s the states of the n bees at each time, we can determine
probabilities of finding the various sites are perfectly the total number of bees dancing for each site at that
proportional to the numbers of bees dancing for them: time. Specifically, for each j (including the case jZ0 of
this is the opposite limiting case in which the bees’ no site), the number of bees dancing for site j at time t is
dancing completely determines all bees’ decisions to
inspect the various sites. nj;t Z jfi : si;t Z jgj:
It remains to define the second component of bee i’s Now different criteria of consensus are conceivable.
state at time tC1, her dance duration di,tC1 if she has Generally, a consensus criterion can be understood as a
begun a dance for one of the k sites, say site j. (If she mapping from the individual-level pattern of dance
has found no site, di,tC1 is set to zero.) Initially, we activity to a chosen nest site. According to the one such
assume that di,tC1 is always stochastically determined criterion, a consensus for a site j at a given time t would
by the bee’s independent assessment of site j ’s true require that all the bees engaged in dance activity at
quality qj. Later, we allow that there is a probability mR0 time t support site j, i.e. nj,tO0 while nh,tZ0 for all hsj
that di,tC1 is unrelated to qj, so as to capture the possibility with hs0. However, as shown by Seeley & Visscher
that the bee’s dance is prompted by mimicking other bees (2003), a consensus in this unanimitarian sense is not
rather than by an inspection of site j. A value of mZ0 necessary, nor even generally sufficient, for the bees’
represents the original case in which bees always selection of a nest site. Instead, the bees appear to make
independently assess a site’s quality before dancing for decisions by ‘quorum’, requiring merely ‘sufficient’
it, while a value of mZ1 represents the opposite case in support for a site to be chosen.
which bees join dances solely based on the probability This can be modelled in a number of different ways.
distribution ( p1,tC1, p2,tC1, ., pk,tC1) over the k sites and For the present purposes, we focus on two illustrative
thus—given a sufficiently large value of l—based on criteria. According to the first and less demanding
mimicking the dances of others, without paying attention criterion—which is arguably too weak to capture the
to any site’s quality. Generally, quorum requirement fully—site j is the winner at time t
8 if it receives more support than any other site at t, i.e.
> q expðTs Þ with probability 1Km nj,tOnh,t for any hsj with hs0. According to the
< j
di;tC1 Z ð‘independent assessment’Þ second and more demanding criterion, site j is the
>
: winner at time t if it meets a two-part condition: (i) it
K expðTs Þ with probability m ð‘mimicking’Þ; receives more than twice the amount of support
received by the second most popular site (i.e. nj,tO
where Ts is a normally distributed random variable with
2nh,t for any hsj with hs0) and (ii) more than 20 per
mean 0 and standard deviation sR0, and K (relevant
cent of the scout bees are engaged in dance activity at t
only in the case of mimicking) is some strictly positive
(i.e. n 0,t!0.8n).
constant. The parameter s specifies the bee’s reliability; a
Our model would also allow the use of other
small value of s corresponds to a high reliability, a large
consensus criteria, but these would yield broadly
value to a low one. Under our definition, the bee’s dance
similar results about the bees’ overall reliability.
duration for any given site fluctuates around the
Differences between such criteria would become more
numerical value of the site’s true quality (setting aside
significant in relation to speed-accuracy trade-offs,
the case of mimicking, where the dance duration
which are not the focus of this paper.
fluctuates around the quality-independent constant K ).
The bee’s error is multiplicative, i.e. an overestimation of
the site’s quality by a factor of cO0, i.e. the bee 3. HYPOTHESES
erroneously takes the site’s quality to be c times its true Since our model is designed to represent the bees’
quality—is as likely as an underestimation by the same empirically observed decision-making behaviour, the
factor, i.e. she takes the site’s quality to be 1/c times its model should predict the reliability of the bees’ decision
true quality. Our results are robust to changes in the process under empirically realistic assumptions. What
functional form of the error, e.g. we obtain similar results do we mean by ‘realistic’? It is reasonable to assume,
when the error takes an additive rather than multi- first, that individual bees are neither very reliable nor
plicative form. completely unreliable, and second, that the bees’
Case 2.2. Bee i dances for one of the sites, say site j, at waggle dances have a significant but not exclusive
time t (i.e. si,tO0). influence on other bees’ decisions to investigate
potential nest sites. The first assumption corresponds
In this case, she continues to dance for the same site at to a non-extremal value of the bees’ reliability
time tC1 with the remaining dance time reduced by one parameter s, and the second to a non-extremal value
period, unless that dance time is over; in the latter case, of the interdependence parameter l. Initially, we
her state is reset to the state of no dancing. Formally, assume no mimicking between the bees (i.e. mZ0).
(
ðsi;t ; di;t K1Þ if di;t O 1 We expect the following:
xi;tC1 Z
ð0; 0Þ otherwise: Hypothesis 3.1. Under a wide range of non-extremal
parameter values of s and l (and mZ0), the bees
choose the best nest site.
(c) When is a consensus reached?
So far we have only modelled the behaviour of Assuming this hypothesis turns out to be true—which
individual scout bees and have not yet explained what is consistent with Seeley’s empirical observations—we
it means for a consensus to emerge among them. From are also interested in explaining why this is the case. As

already indicated in the introduction, we suggest that two process under a range of empirically motivated, non-
characteristics of the bees’ decision process stand out. extremal assumptions about individual bees’ reliability
First, the bees are independent in that they individually (s) and their interdependence through signalling (l), in
inspect potential nest sites and dance to advertise them as each case assuming no mimicking (mZ0). As reported
a function of their individual quality assessments of these below, our simulations broadly confirm hypothesis 3.1.
sites; they do not blindly join a dance for a nest site Our second set of simulations was designed to test
without having inspected the site themselves (i.e. m is 0 or hypothesis 3.2, focusing on the mechanisms underlying
low). We can express this in the language of probability the reliability of the bees’ decision process. To isolate
theory by saying that, conditional on having identified a the effects of independent assessments of the various
particular site, each bee’s dance duration for that site is sites’ quality by the bees and communicative inter-
independent of other bees’ dance durations for it. dependence between them, we varied the parameters m
Second, the bees are interdependent in that they are and s such that one of these two characteristics
more likely to inspect nest-sites advertised by others (i.e. l was completely or partially absent from the bees’
is high). Expressed in probability-theoretic terms, the decision making.
identification of a particular site by one bee is correlated To model the full or partial absence of indepen-
with the identification of that site by others. We hypo- dence, we considered non-zero values of m, thereby
thesize that the reliability of the bees’ decision process allowing that a bee may join a dance for a site not on the
is driven by the interplay of these two characteristics: basis of her independent assessment of its quality but
merely as a result of mimicking other bees dancing for it
Hypothesis 3.2. The bees’ independence in assessing the
(we set the relevant constant K equal to the maximal
various sites’ quality and their interdependence
nest-site quality, but other values of K would yield
through communication are both necessary and
similar results). Recall that our original case mZ0
sufficient for the reliability of the bees’ decision process.
meant that bees always independently assess a site’s
Both hypotheses are deliberately stated informally quality before dancing for it. By contrast, the higher the
here, but their operationalization will become clearer in value of m, the less likely it is that a bee independently
the context of our computer simulations. assesses a site’s quality before dancing for it. In the
limiting case mZ1, bees join dances randomly, based
on only the probability distribution over sites, without
4. COMPUTER SIMULATIONS paying attention to any site’s quality (i.e. the dance
(a) Basic description duration is determined by the site-quality-independent
Implementing the model above as a MATHEMATICA random variable K exp(Ts), as defined above). Our
program, we ran a number of computer simulations of simulation results reported below are consistent with
the bees’ nest-site choice under various parameter hypothesis 3.2, showing that high values of m under-
conditions. To ensure comparability across simul- mine the reliability of the bees’ decision process, while
ations, we fixed the number of scout bees at nZ200 low values support it.
and the number of potential nest sites at kZ5. These To model the full or partial absence of interdepen-
assumptions are empirically motivated: there are dence between the bees, we varied the parameter l. As
usually several hundred scouts in a swarm, and there already noted, a value of zero implies that each bee’s
were typically five candidate nest sites in the experi- probabilities of finding the various nest sites remain the
ments conducted by Seeley and others on Appledore a priori probabilities, regardless of other bees’ dancing
Island, off the coast of Maine. activity; a value of one implies that each bee’s
We also fixed the objective quality levels q1, ., q5 of probabilities of finding those sites are perfectly
the five nest sites at 3, 5, 7, 9, 10, respectively, thus proportional to the numbers of bees dancing for
making it intuitively difficult for the bees to distinguish them; no other factor leads a bee to inspect any site.
the two or three best nest sites. (Even when individual Here, too, our simulation results are consistent with
reliability is high, e.g. sZ0.2, the intervals in which hypothesis 3.2; low values of l undermine the reliability
individual quality assessments of the two best sites are of the bees’ decision process while high values reinforce
likely to fall, namely [9 exp(K0.2), 9 exp(0.2)]Z[7.37, it. An exception arises for the limiting case lZ1, where
10.99] and [10 exp(K0.2), 10 exp(0.2)]Z[8.19, the bees’ probabilities of finding the different sites are
12.21], overlap significantly. When individual reliability given by the existing dance proportions for those sites.
is lower, e.g. sZ1, the overlap between these intervals, Here, there is not enough noise in the system for bees to
now [3.31, 24.46] and [3.68, 27.18], grows further.) discover any new sites not advertised by others. Small
We further assumed that when a bee flies out noisy deviations from perfect proportionality (i.e. l!1)
randomly without following any other bees’ advertise- are necessary to permit the discovery of new sites.
ment for a site, she has a 25 per cent probability of
finding some site, where the probability is equally (b) Results on hypothesis 3.1
distributed over the five sites (i.e. p1Z.Zp5Z5% and Our first simulations capture what may be described as
p0Z75%). In all simulations, we calculated the bees’ a best-case scenario: the reliability of bees in assessing
behaviour for 300 time periods, though a consensus, sites is good (sZ0.2), and their interdependence
under both criteria introduced above, often emerged in through communication is high (lZ0.8). Figure 1
less time. We verified that our findings are robust to shows an illustrative simulation for these parameter
changes in the choice of these fixed parameters. values. The figure shows the number of bees engaged in
Our first set of simulations was run to test hypothesis dance activity for each of the five nest sites at each of
3.1 by investigating the reliability of the bees’ decision the 300 time periods calculated. It is easy to see that

d for d for 5
ce
n
5 ce 4
da
4 an 3
e 3 ed 2
sit 1
2 i
s 1t
none none
175
150 175
ees
125 150
ees
100 125
no. of b
no. of b
75 100
50 75
25 50
25
0
100 0
tim 200 100
e (t tim
) 300 e (t 200
) 300
Figure 1. Illustrative simulations to test hypothesis 3.1. High
reliability, high interdependence. Figure 3. Illustrative simulations to test hypothesis 3.1. High
reliability, low interdependence.
for for
c ed 4 5 c ed 4 5
an 3 an 3
ed ed
sit 1 2 sit 1 2
none none
175
175 150
ees
150 125
ees
no. of b
125 100
75
no. of b
100
75 50
50 25
25
0
0 100
100 tim 200
e (t
tim 200 ) 300
e (t
) 300
Figure 4. Illustrative simulations to test hypothesis 3.1. Low
Figure 2. Illustrative simulations to test hypothesis 3.1. High reliability, high interdependence.
reliability, medium interdependence.
after some initial amount of dance activity for other periods. The consensus is less strong here. Again we
sites, the dance activity concentrates on the best site repeated the simulation for the same parameter values 250
(site 5). To confirm that this pattern is not accidental, times. Under the less demanding criterion of consensus,
we repeated the simulation with the same parameter the best nest-site emerged as the winner 226 times and
values 250 times. In each case, we determined the the second best 22 times, with no winner in the remaining
consensus winner using the two illustrative criteria two cases; under the more demanding criterion, the
introduced above, applied to the last time period of the best site won 104 times, with no winner in all other cases.
simulation. Recall that according to the first and A further reduction in the interdependence between
weaker criterion, a site wins if it receives more support bees (to lZ0.2) weakens the emergence of a consensus
than any other site; according to the second and even more significantly, as shown in figure 3. In 250
stronger criterion, a site wins if it receives more than repetitions, this effect is particularly evident when we
twice the amount of support received by the second employ the more demanding one of our two criteria for
most popular site and at least 20 per cent of the scout consensus. While under the weaker criterion the best
bees are engaged in dance activity. For the present site (site 5) still won 176 times (and the second best 63
parameter values (lZ0.8 and sZ0.2), the best nest-site times, the third best once and no winner 10 times),
emerged as the winner in nearly all cases, regardless of under the stronger criterion the best site won only 11
which criterion for a consensus was employed: Under times and the second best once, with no consensus in
the first criterion, the best site was chosen every time; the other 238 cases.
under the second, it was chosen 246 times, while no site Having focused so far on the case in which bees are
met the winner criterion in the remaining four cases. highly reliable in assessing nest sites, let us now introduce
In our next simulations, the interdependence more noise into the bees’ individual assessment of sites
between bees is reduced to a lower level (lZ0.5), (setting sZ1). Strikingly, if interdependence between
while the other parameter values remain as before. bees is high (i.e. lZ0.8), the overall pattern remains
Figure 2 shows a representative simulation. While the broadly the same as in the best-case scenario reported
best site (site 5) continues to receive the most support, earlier. Figure 4 shows a sample calculation. In 250
there is also considerable dance activity for other sites, repetitions of the simulation, the best site (site 5) emerged
particularly the second best (site 4), throughout all time as the consensus choice, even under the stronger

d for 5 d for 5
ce 4 ce 4
an 3 an 3
i t ed 2 i t ed 2
s 1 s 1
none none
175 175
150 150
ees
ees
125 125
no. of b
no. of b
100 100
75 75
50 50
25 25
0 0
100 100
tim 200 tim 200
e (t e (t
) 300 ) 300
Figure 5. Illustrative simulations to test hypothesis 3.1. Low Figure 6. Illustrative simulations to test hypothesis 3.1. Low
reliability, medium interdependence. reliability, low interdependence.
criterion, 199 times and the second best site five times from low (lZ0) to high (lZ1), keeping a high
(site 4), with no consensus in the remaining 46 cases. individual reliability and independence of bees
Under the weaker criterion, the best site won 237 times, (sZ0.2 and mZ0). Figure 9 shows the proportion of
the second best 12 times, with no winner once. wins for each of the five sites (in 250 simulations for
Figures 5 and 6 show similar calculations for each set of parameter values), for different levels of
medium and low levels of interdependence (i.e. interdependence, where the winner is calculated using
lZ0.5 and lZ0.2, respectively) and essentially the more demanding criterion. The results are
confirm the earlier results for higher individual qualitatively similar under the less demanding
reliability. In 250 repetitions of the simulations for a criterion. The best site (site 5) consistently emerges
medium level of interdependence, the best site emerged as the winner only when the level of interdependence is
as the winner under the strong criterion 94 times and above a certain threshold. (Recall our earlier remarks,
the second best once, with no consensus in all other in the basic description of the computer simulations,
cases. (Under the weaker criterion, the effect is less about the exceptional limiting case lZ1.)
pronounced: the best site won 220 times and the Similarly, we ran a large number of simulations with
second best 28 times, with no winner twice.) When the the level of independence ranging from low (mZ1) to
level of interdependence was low, the best site won only high (mZ0), keeping a high interdependence of bees
seven times under the strong criterion, with no and a high reliability in the event they do verify a site’s
consensus in the other 243 cases. (Under the weaker quality (lZ0.8 and sZ0.2). Figure 10 shows the
criterion, the best site won 190 times, the second proportion of wins for each site (again in 250
best 58 times and in the remaining two cases, no simulations for each set of parameter values), for
consensus emerged.) different levels of independence (displayed as 1Km).
Table 1 summarizes the frequencies of various Again, we use the more demanding criterion of
consensus choices in 250 repetitions of the simulations. consensus; the results are qualitatively similar under
the less demanding criterion. The best site (site 5)
consistently emerges as the winner only when the level
(c) Results on hypothesis 3.2 of independence is not too low.
As indicated, we isolated the effects of independence
and interdependence by running simulations for the
special cases in which one or the other of these two 5. DISCUSSION
characteristics was absent. Figure 7 shows the case in We have developed an agent-based model of nest-site
which bees independently assess sites and they are also choice among honeybees. The model not only explicitly
relatively reliable (sZ0.2), but there is no longer any represents the behaviour of each individual bee as a
interdependence between bees through communication simple stochastic process, but it also allows us to
(lZ0). No clear consensus winner emerges. Figure 8 simulate the bees’ decision-making behaviour under a
shows an illustrative case in which interdependence wide variety of empirically motivated as well as
between bees is high (lZ0.8), but they are no longer hypothetical assumptions. The model predicts that,
independent (mZ1); they all mimic the dances of other consistently with empirical observations by Seeley &
bees without independently verifying the sites’ quality. Buhrman (2001), the bees manage to reach a consensus
In this example, after relatively little initial dance on the best nest site for a large range of parameter
activity, a sudden cascade of support for site 2 emerges conditions, under both more and less demanding criteria
(the second worst site), which is reinforced by the bees’ of consensus. Moreover, the model shows that the
mimicking of others’ dances. Equally, a cascade of remarkable reliability of the bees’decision-making process
support for another site could have randomly emerged. stems from the particular interplay of independence and
More generally, we ran a large number of interdependence between them. The bees are indepen-
simulations with the level of interdependence ranging dent in assessing the quality of different nest sites on their

Table 1. Frequencies of various consensus choices in 250 replications of the simulations.
high individual reliability (sZ0.2) low individual reliability (sZ1)
strong consensus weak consensus strong consensus weak consensus

criterion criterion criterion criterion
high interdependence 1st best site 246 (98.4%) 250 (100%) 199 (79.6%) 237 (94.8%)
(lZ0.8) 2nd best site 0 (0%) 0 (0%) 5 (2%) 12 (4.8%)
3rd best site 0 (0%) 0 (0%) 0 (0%) 0 (0%)
none 4 (1.6%) 0 (0%) 46 (18.4%) 1 (0.4%)
medium interdependence 1st best site 104 (41.6%) 226 (90.4%) 94 (37.6%) 220 (88%)
(lZ0.5) 2nd best site 0 (0%) 22 (8.8%) 1 (0.4%) 28 (11.2%)
3rd best site 0 (0%) 0 (0%) 0 (0%) 0 (0%)
none 146 (58.4%) 2 (0.8%) 155 (62%) 2 (0.8%)
low interdependence 1st best site 11 (4.4%) 176 (70.4%) 7 (2.8%) 190 (76%)
(lZ0.2) 2nd best site 1 (0.4%) 63 (25.2%) 0 (0%) 58 (23.2%)
3rd best site 0 (0%) 1 (0.4%) 0 (0%) 0 (0%)
none 238 (95.2%) 10 (4%) 243 (97.2%) 2 (0.8%)
for 5 or
d e df 5
ce 4 c
an 3 an 4
i t ed 2 i t ed 2 3
s 1 s 1
none none
ins
175
on of w
150
ees
125
no. of b
100 proporti
75
50
25
0
0 inte 0.5
100 rde
tim pen 1.0
e (t 200 den
) ce (
300 )
Figure 7. Illustrative simulations to test hypothesis 3.2. Figure 9. Sequences of simulations to test hypothesis 3.2.
Independence without interdependence. Interdependence ranging from low to high.
for df
or
c ed 4 5 e 5
an 3 a 4 nc
ed ed 3
sit 1 2 sit 1 2
none none
wins
175
150
proportion of
ees
125
no. of b
100
75
50
25
0
0 ind
epe 0.5
100 nde
nce
tim
e (t 200 ( 1–µ 1.0
) 300 )
Figure 8. Illustrative simulations to test hypothesis 3.2. Figure 10. Sequences of simulations to test hypothesis 3.2.
Interdependence without independence. Independence ranging from low to high.
would still emerge, but it would no longer robustly be

own, but interdependent in giving more attention to nest on the best nest site; instead, many bees would end up
sites that are more strongly advertised by others. dancing for nest sites that accidentally receive some
Without interdependence, the rapid convergence of initial support through random fluctuations. It is only
the bees’ dances to a consensus would be undermined; when independence and interdependence are com-
there would not be a ‘snowballing’ of attention on the bined in the right way that the bees achieve their
best nest site. Without independence, a consensus remarkable collective reliability.

To illustrate the importance of both independence Dryzek, J. & List, C. 2003 Social choice theory and
and interdependence in collective decision making deliberative democracy: a reconciliation. Br. J. Pol. Sci.
more generally, consider the human example of 33, 1–28. (doi:10.1017/S0007123403000012)
restaurant choice. If we walk into a street with many Elster, J. 1986 The market and the forum. In Foundations of social
restaurants, knowing nothing about their quality, we choice theory (eds J. Elster &A. Hylland). pp. 103–132.
Cambridge, UK: Cambridge University Press.
may well pick whichever restaurant has the most diners,
Farrar, C., Fishkin, J. S., Green, D., List, C., Luskin, R. C. &
assuming that these people have chosen it for its Paluck, E. L. In press. Disaggregating deliberation’s
quality. But if their reasoning was the same, we may all effects: an experiment within a deliberative poll. Br. J.
end up in the worst restaurant; if we do not, this is Pol. Sci.
entirely accidental. This phenomenon is called an Grofman, B., Owen, G. & Feld, S. L. 1983 Thirteen
informational cascade (Bikhchandani et al. 1992). It is theorems in search of the truth. Theory Decision 15,
only if enough people choose a restaurant based on 261–278. (doi:10.1007/BF00125672)
independently gathered information that such cascades Knight, J. & Johnson, J. 1994 Aggregation and deliberation:
can be avoided. Interdependence without indepen- on the possibility of democratic legitimacy. Pol. Theory 22,
dence can lead everyone to a bad decision. (Of course, 277–296. (doi:10.1177/0090591794022002004)
in the restaurant case, additional complexities arise Lindauer, M. 1955 Schwarmbienen auf Wohnungssuche.
because differences in individual choices are due to not Zeitschrift für vergleichende Physiologie 37, 263–324.
(doi:10.1007/BF00303153)
only different information but also different prefer-
List, C. 2004 Democracy in animal groups: a political science
ences; in the case of the bees, we have made the perspective. Trends Ecol. Evol. 19, 168–169. (doi:10.1016/
simplifying assumption that differences in individual j.tree.2004.02.004)
choices are solely due to informational differences, List, C. & Goodin, R. E. 2001 Epistemic democracy:
presupposing that there is an objective quality standard generalizing the Condorcet jury theorem. J. Pol. Philos.
for nest sites.) 9, 277–306. (doi:10.1111/1467-9760.00128)
Likewise, independence without interdependence Luskin, R. C., Fishkin, J. S. & Jowell, R. 2002 Considered
can be suboptimal too. Imagine a world without any opinions: deliberative polling in Britain. Br. J. Pol. Sci. 32,
restaurant reviews or recommendations, where people 455–487. (doi:10.1017/S0007123402000194)
only ever gather information about restaurants by Miller, D. 1992 Deliberative democracy and social choice.
wandering around the streets alone and checking out Pol. Studies 40, 54–67. Special Issue.
Myerscough, M. R. 2003 Dancing for a decision: a matrix
the restaurants they see. It can easily happen, then, that
model for nest-site choices by honeybees. Proc. R. Soc. B
an outstanding but badly located restaurant may fail to
270, 577–582. (doi:10.1098/rspb.2002.2293)
get many customers because too few people stumble Passino, K. M. & Seeley, T. D. 2006 Modeling and analysis of
across it. nest-site selection by honey bee swarms. The speed and
The bees’ ingenuity lies in their avoidance of both of accuracy trade-off. Behav. Ecol. Sociobiol. 59, 427–442.
these problems. Nature has given them the right mix of (doi:10.1007/s00265-005-0067-y)
independence and interdependence. Perdriau, B. S. & Myerscough, M. R. 2007 Making good
choices with variable information: a stochastic model for
This paper is based on an unpublished manuscript of January nest-site selection by honeybees. Biol. Lett. 3, 140–143.
2005. We are grateful to Larissa Conradt, Kai Spiekermann
(doi:10.1098/rsbl.2006.0599)
and two anonymous referees for their helpful comments.
Seeley, T. D. & Buhrman, S. C. 2001 Nest-site selection in
honey bees: how well do swarms implement the ‘best-
REFERENCES of-N’ decision rule? Behav. Ecol. Sociobiol. 49, 416–427.
Austen-Smith, D. & Banks, J. S. 1996 Information aggrega- (doi:10.1007/s002650000299)
tion, rationality, and the Condorcet jury theorem. Am. Pol. Seeley, T. D. & Visscher, P. K. 2003 Choosing a home: how
Sci. Rev. 90, 34–45. (doi:10.2307/2082796) the scouts in a honey bee swarm perceive the completion
Austen-Smith, D. & Feddersen, T. J. 2009 Information of their group decision making. Behav. Ecol. Sociobiol. 54,
aggregation and communication in committees. Phil. 511–520. (doi:10.1007/s00265-003-0664-6)
Trans. R. Soc. B 364, 763–769. (doi:10.1098/rstb.2008. Seeley, T. D., Visscher, P. K. & Passino, K. M. 2006 Group
0256) decision making in honey bee swarms. Am. Sci. 94,
Bikhchandani, S., Hirshleifer, D. & Welch, I. 1992 A theory 220–229.
of fads, fashions, custom, and cultural change as Sunstein, C. R. 2002 The law of group polarization. J. Pol.
informational cascades. J. Pol. Econ. 100, 992–1026. Philos. 10, 175–195. (doi:10.1111/1467-9760.00148)
(doi:10.1086/261849) Sunstein, C. R. 2006 Deliberating groups versus prediction
Borland, P. J. 1989 Majority systems and the Condorcet jury markets (or Hayek’s challenge to Habermas). Episteme 3,
theorem. Statistician 38, 181–189. (doi:10.2307/2348873) 192–213.
Britton, N. F., Franks, N. R., Pratt, S. C. & Seeley, T. D. Zuber, J. A., Crott, H. W. & Werner, J. 1992 Choice shift and
2002 Deciding on a new home: how do honeybees agree? group polarization: an analysis of the status of arguments
Proc. R. Soc. Lond. B 269, 1383–1388. (doi:10.1098/rspb. and social decision schemes. J. Pers. Soc. Psychol. 62,
2002.2001) 50–61. (doi:10.1037/0022-3514.62.1.50)

Phil. Trans. R. Soc. B (2009) 364, 763–769

doi:10.1098/rstb.2008.0256
Review
Information aggregation and communication

in committees
David Austen-Smith* and Timothy J. Feddersen
MEDS Department, Kellogg School of Management, Northwestern University, Evanston, IL 60208, USA
In this paper, we attempt to explain the underlying strategic incentives confronting individuals when
they must make a collective decision over a set of alternatives and each has information that is
decision-relevant for others. A significant literature has emerged in formal political theory over the
past several years that focuses on such problems, paying particular attention, first, to the extent to
which voting can be expected to aggregate committee members’ information and, second, to the role
of communication among committee members prior to voting. Inter alia, this literature reveals a
surprisingly subtle interaction between the voting rules used to make decisions and the incentives for
committee members to share information prior to voting.
Keywords: information; communication; committees
1. INTRODUCTION of Don’s papers and see if any look promising. In that

Rather than providing a necessarily discursive sum- case, the dean might choose to hire Don if all of the
mary of the formal literature on information aggrega- papers he does read happen to be promising and not
tion and communication in committees, our goal in this otherwise. Similarly, because he cannot read all of
paper is to explain the underlying strategic incentives Don’s papers, he might choose not to hire at all in
confronting individuals when they must make a which case he runs the risk of rejecting Don when in
collective decision over a set of alternatives and each fact all of Don’s papers are promising. But in taking this
has information that is decision relevant for others. approach, the dean runs the risk of hiring Don when
A canonical example of this class of problem is that of a not all of his research is promising, illustrating the
jury charged to decide whether a defendant is guilty or problems facing a decision-maker with imperfect
innocent; all jurors prefer to convict the guilty and information about the relative merits of competing
acquit the innocent, but each individual is unsure of alternatives. Decision-makers (like deans) often try to
the defendant’s status in this respect. A significant minimize the likelihood of errors by appointing a
literature has emerged in formal political theory over faculty committee to look at job candidates and make
the past several years that focuses on such problems, recommendations. The advantage of a committee is
paying particular attention, first, to the extent to which that it is, in principle, better informed collectively than
voting can be expected to aggregate committee any individual on the committee (although each
members’ information and, second, to the role of individual may read no more papers than the dean,
communication among committee members prior to the committee as a whole may read all of the papers).
voting. Inter alia, this literature reveals a surprisingly And in 1785, de Condorcet (1994) demonstrated that
subtle interaction between the voting rules used to when committee members in such situations vote in a
make decisions and the incentives for committee way that reflects their information (i.e. they vote
members to share information prior to voting. informatively), the committee decision minimizes the
probability of mistakes (see also Ladha 1992).
However, what was not appreciated until recently is
2. STRATEGIC VOTING WITH PRIVATE
that voting creates incentives for committee members
INFORMATION
not to vote informatively even when all committee
To fix ideas, it is helpful to consider a simple example.
members share the same objectives. To see this,
The dean of a college has to decide whether to hire a
suppose the dean appoints a faculty committee of
job candidate (Don). The dean would like to hire Don
three people (Alice, Bob and Chris) who must choose
if and only if all of his papers are promising. Don has a
whether or not to hire Don. The dean instructs the
long vita, too long for the dean to read. Assuming he is
committee that the candidate will be hired if and only if
qualified to recognize promising research in the
a majority of the committee votes in favour. Each
candidate’s field, the dean might read a small subset
member of the faculty committee, like the dean,
favours hiring a candidate if and only if all of his
* Author for correspondence (dasm@kellogg.northwestern.edu). research papers are ‘promising’. For the moment, that
One contribution of 11 to a Theme Issue ‘Group decision making in is, the committee members’ preferences are presumed
humans and animals’. to satisfy common values with respect to the decision:

764 D. Austen-Smith & T. J. Feddersen Review. Information aggregation and communication
conditional on having the same information, all A large literature exists in formal theory exploring
individuals’ preferences over final outcomes are iden- the relationship between voting rules and information
tical. The committee decides to divide responsibility aggregation when, unlike in the Condorcet jury
for reading Don’s papers equally among themselves, theorem, voters are strategic (early contributions
so that each member reads a different set of papers. include Ordeshook & Palfrey 1988; Austen-Smith &
The dean then hires Don if a majority votes in favour. Banks 1996; Feddersen & Pesendorfer 1996, 1997).
How should a faculty member, say Alice, vote? As the preceding discussion suggests, such strategic
Alice knows that her individual vote does not always behaviour can be quite subtle and, perhaps, prima facie
determine the outcome; in particular, her vote is pivotal implausible as a description of how people actually
when the other faculty members are split with one vote. However, there now exists an experimental
voting in favour and the other voting against. Alice also literature that finds that individuals do indeed seem
knows that because each committee member reads a to behave in a manner consistent with the theory (see,
distinct subset of Don’s papers, all three of them have for example, Guarnaschelli et al. 2000; Battaglini et al.
private information that is potentially relevant for every in press a,b).
committee member’s decision. For example, if Alice
has read some papers that are not promising, then she
knows that no person on the committee would like Don 3. VOTING RULES AND TALK
to be hired, including those who have read only It can be shown that informative voting by all
promising papers. Similarly, Alice knows that even if committee members is often not an equilibrium. As
she has read only promising papers, Don should not be a result, in small committees, equilibrium behaviour
hired if either Bob or Chris has read a paper that is not produces outcomes that are sometimes unsatisfactory;
promising. As a result Alice’s problem of how to vote in that is, there exist situations in which the candidate is
the committee is qualitatively different from that of an not hired when all of his research is promising and
individual decision-maker choosing in isolation (e.g. others in which he is hired when not all of his
the dean above). research is promising. Say that an outcome from
Suppose that not all of Don’s papers read by Alice equilibrium voting satisfies full information equivalence
are promising. Clearly, since everyone favours hiring if the result of the voting is always the same as the
Don only if all his research is promising, Alice should result would be if all private information were shared.
vote against hiring Don. But what if all of the papers Feddersen & Pesendorfer (1996, 1997) demonstrated
she has read are promising? Alice’s decision depends on that Condorcet’s basic insight is robust to strategic
behaviour for sufficiently large committees or electo-
how she thinks Bob and Chris are voting. If Alice
rates under many (non-unanimous) voting rules, even
believes that Bob and Chris vote informatively to hire
without the assumption of common values (see also
Don if and only if all of the papers each individually
Wit 1988; McLennan 1998). But our interest in this
reads are promising, she must also believe that the only
essay is with small committees and, in this setting,
event in which her vote is pivotal is when exactly one of
Austen-Smith & Banks (1996) demonstrated that
the other two has read papers that are not promising
voting alone may fail to satisfy full information
while the other has read only promising papers. But in
equivalence if the voting rule is not aligned with
this case she would rather vote not to hire Don. That is,
committee members’ preferences and the information
in situations with common values and private infor- environment. One approach to the problem of
mation, informative voting behaviour makes the event aggregating information in small committees, there-
that one is pivotal informative about what others know. fore, is to choose the voting rule appropriately.
Since one’s vote only matters in the event it is pivotal, Consider our running example (with common
individuals have an incentive to condition their vote not values) and suppose that instead of using majority
only on their own private information but also on the rule in committee, the dean requires a unanimous vote
information that others must possess in that event. This in favour of Don for him to be hired. Under this rule, if
is the case even though Alice does not know how others Alice believes as before that both Bob and Chris are
have voted and may believe it to be very unlikely that voting informatively, then she must also believe that the
her vote is pivotal. only event in which her vote is pivotal is when both are
Returning to Alice’s problem of how to vote, it seems voting in favour of hiring Don. Therefore, conditional
from the above that if others are voting informatively, on being pivotal, Alice’s optimal response is to vote
then Alice should not and instead she ought always to informatively as well, that is Alice should vote to hire
vote against hiring Don. But then similar reasoning Don if and only if all the papers she has read are
applies equally to Bob and Chris, further complicating promising. Thus, informative voting by all committee
Alice’s decision. Formally, such situations are properly members constitutes equilibrium behaviour and
analysed as games rather than as individual decisions. assures full information equivalence: Don is hired
Game theory asks what strategic behaviours by if and only if all of the papers read by the committee
committee members constitute an equilibrium: a list of are promising.
decisions or strategies, one for each individual, that The relationship between voting rules and infor-
together constitute a mutually consistent set of best mation aggregation in committees is an important
responses. In other words, in equilibrium, each concern in the formal literature and one to which we
individual correctly anticipates the behaviour of others return shortly. Voting rules, however, are often fixed
and chooses optimally given their correct beliefs about prior to knowing the details of any particular decision
that behaviour. or information environment and cannot be tailored to

Review. Information aggregation and communication D. Austen-Smith & T. J. Feddersen 765
each and every separate collective choice (see Austen- committee. In effect, this simple communication
Smith & Banks 1996; Feddersen & Pesendorfer 1998). structure serves as a kind of straw poll. Each committee
An alternative and, at least for small committees with member reads the others’ reports and then submits
common values, obvious solution to the potential their vote to the dean. Suppose that all of the papers
failure to aggregate information under majority rule is Alice read are promising and she is considering what to
to have everyone on the committee share their private report to the others. If Alice believes Bob and Chris are
information before voting. Suppose, in our example, reporting truthfully and expect her to do so likewise,
that Alice, Bob and Chris recognize the possible Alice anticipates that Bob and Chris will vote to hire
difficulties with simply voting and decide to meet and Don if and only if all report that they have read only
talk prior to voting. Then they might share their private promising papers; thus Alice’s report is pivotal at the
information with each other and so condition their communication stage only if Bob and Chris report that
vote on all of the available information about Don. all of the papers they have read are promising. If Alice
But sharing information in conversation turns out to be falsely reports that Don’s papers were not all promis-
problematic, both substantively and technically. ing, the result is that Don is not hired (because Bob and
The canonical formal literature focuses on the Chris will vote against), so she has an incentive to
extreme case of cheap talk. An individual sends a report truthfully; and it is easy to confirm that, because
message (makes a speech) from an abstract set of they share common values, the same is true in the event
possible messages; a listener hears the message and that Alice has read some unpromising papers by Don.
draws an inference about the speaker’s private infor- It is important to observe here that the preceding
mation. The speaker anticipates the different inferences argument did not depend on whether the committee
listeners might make and chooses the message that, votes under majority rule or unanimity rule. Moreover,
given his or her preferences, elicits the best possible the event in which Alice’s report is pivotal under
behavioural response from the listener. If the speaker majority rule at the communication stage (that is, given
suffers no consequences for making any particular truthful information sharing, both Bob and Chris
speech beyond the behaviour that her speech elicits and report that they have read only promising papers), is
if the listener has no independent way to verify the distinct from the event in which Alice’s vote is pivotal
veracity of the content of the speech, we say talk is at the voting stage (that is, given informative voting,
cheap. The cheap talk model, therefore, allows for the exactly one of Bob and Chris is voting to hire Don).
possibility of deception and opacity. For example, Indeed, given common values and truthful communi-
suppose that Alice has not read any promising papers cation prior to the voting stage, no committee member
by Don. In the pre-vote meeting with Bob and Chris, is pivotal at the voting stage under majority rule since
she might tell them that in fact all of the papers she read all vote unanimously one way or the other, depending
are promising, even though that is not the case. on the realized distribution of reports. Under unani-
However, assuming Bob and Chris are rational mity rule, however, every committee member is pivotal
listeners, they take account of Alice’s incentives and, at the voting stage following full revelation during the
perhaps, discount her speech accordingly. For Alice to communication stage and, just as in the case of absent
be credible, it must be the case that whatever her communication, every member has an incentive to
private information, she prefers to tell the truth (see vote informatively.
Crawford & Siobel 1982; Austen-Smith 1992; Farell & With common values, talk can compensate for
Rabin 1996). inadequate voting rules. While the particular communi-
Of course, in the real world, talk may not (always) cation protocol used in the example above (reports etc.)
be cheap: Alice might feel bad about dissembling, or is highly stylized, the basic intuition is quite robust.
she might worry that others discover her misrepre- For example, the committee members could meet
sentation and impose some sort of sanction, or Alice and talk in sequence as they would in a real meeting.
might be able to reveal credibly that she has read a In all cases in which a member’s speech influences
promising paper by showing it to the others. Never- the voting outcome common values guarantee, the
theless, the assumption of cheap talk is a particularly member wants to speak truthfully. Perhaps unfortun-
useful baseline from which to evaluate the incentives ately, this insight does not extend beyond the common
for individuals to share information in various value setting. When the assumption of common
situations. Indeed, even when talk is cheap, if the values is relaxed, an interesting and subtle connec-
committee shares common values, then talking prior tion between voting rules and incentives to share
to voting creates an incentive for everyone to report information emerges.
their private information truthfully and, subsequently,
to vote unanimously to yield full information equi-
valence. This simple solution to difficulties with
information aggregation in committees is explored in 4. BIASES, BELIEFS AND HETEROGENEITY
Coughlan (2000) for a wide class of voting rules. And The assumption that committee members share
although quite intuitive, it proves convenient to walk common values is not general and, when it fails
through a simple model of communication that to hold, allowing people to talk prior to voting
supports the result. introduces a further layer of strategic complexity
To this end, suppose that, prior to voting, each beyond that of voting with private information. First,
committee member independently writes a report however, it is necessary to be a little more precise about
saying essentially that either all or not all of the papers how preferences and beliefs are normally modelled
read are promising, and sends it to everyone on the in this literature.

To discuss a world without common values, it will however, this second conclusion fails. If only one
be useful to introduce some additional terminology. negative vote is needed not to hire Don and, as before,
We say that committee members have beliefs and biases. Alice believes both Bob and Chris are voting informa-
In the formal theory literature, beliefs are modelled as tively, then Alice is pivotal if and only if both of the
conditional probabilities: conditional on some event, others are voting in favour of Don which implies that
an individual assigns a probability to each of a set of neither of them has read an unpromising paper. Hence
possible states of the world. In our example, there are Alice prefers to vote uninformatively in favour of hiring
four states of the world: zero, one, two or three Don however many unpromising papers she reads.
committee members have read exclusively promising Under common values, everyone has an incentive to
papers. An individual reads some subset of Don’s share their information truthfully at the talk stage. To
papers. If she has read an unpromising paper, then she see that this conclusion does not hold without common
believes that one state has surely not occurred, viz. that values, suppose each member is believed by the others
Don has not written only promising papers, but is to report their private information truthfully. Now
unsure as to what state has in fact occurred. By assume that at least one of Don’s papers that Alice
contrast, if all of the papers Alice read were promising, reads is not promising. When does it matter what Alice
then she assigns different probabilities to each of reports? If Alice’s report matters at this stage, then it
the states. must be because both Bob and Chris report that all the
The example further illustrates that beliefs depend papers that they have read are promising, otherwise
not only on individuals’ private information, but also both will vote against hiring Don and he will not be
on the behaviour of others. This is transparent if, say, hired under either majority or unanimity rule. (This
Alice knows that Bob and Chris vote to hire Don if and follows because both Bob and Chris have the most
only if they have read exclusively promising papers and demanding bias; if either reads a paper that is not
she observes them cast split votes. In such a case, even if promising then they would, by hypothesis, have
Alice has read only promising papers, her belief reported that information during the talk stage and
conditional on this event is that Don has, in fact, both vote against hiring.) Therefore, conditional upon
written an unpromising paper and so should not be Alice’s report being pivotal, she has an incentive to
hired. The surprising thing is that Alice need not report falsely that all the papers she has read are
observe the votes of others to draw this inference: promising. Unlike the common values setting, neither
because the outcome is determined by voting, necess- talk prior to voting nor changes in the voting rule itself
arily her vote only matters when the votes of the others are sufficient to guarantee full information equivalence.
are split and hence she may draw the inference that An important feature of the preceding scenario is
exactly one of Bob or Chris (it does not matter which) that Alice knows that Bob and Chris have biases
has read an unpromising paper. distinct from her own. Austen-Smith & Feddersen
Committee members also have biases, that is, for a (2006) showed that the combination of bias
given state of the world, an individual’s bias describes uncertainty and non-unanimous voting rules can
his or her preferences over whether or not to hire Don. sometimes recover full information equivalence (see
Under common values, all committee members share also Meirowitz 2006, 2007). Alice’s incentive not to
the same bias; in the example, to hire Don if and only if reveal her information in the preceding discussion
he has written only promising papers. arose because she knew that the two others had a
Coughlan’s (2000) result cited earlier shows that more stringent standard for hiring. When bias
talking can compensate for an inappropriate voting uncertainty is introduced, however, Alice must be
rule and induce full information equivalence. But this concerned that her colleagues may share her bias
result turns out to depend crucially on the presumption and, therefore, by misinforming them she may cause
of common values. Coughlan (2000) and Austen- them to vote against their interest and hers.
Smith & Feddersen (2005, 2006) showed that even a To illustrate the role of bias uncertainty, suppose
small degree of uncertainty about whether there are that Alice is unsure whether Bob and Chris share her
common values in the committee leads to two things. bias and prefer to hire Don if any of his papers are
First, there is no voting rule that can induce all promising or, as above, both prefer to hire Don only if
committee members to vote informatively and, second, all of his papers are promising. Under the assumption
whatever voting rule is chosen, there may be incentives that Bob and Chris truthfully reveal their private
for the members not to reveal their private information information and share Alice’s bias, it follows that the
prior to voting. It is easy to illustrate this observation only event in which Alice’s information influences
by perturbing our running example for the extreme Bob’s and Chris’s vote is when neither have read any
case, that is, where there is certainty that people promising papers. In that case, if Alice has not read any
have different biases. promising papers either but claims she did, the result is
Suppose everyone knows that Bob and Chris prefer that both Bob and Chris vote for Don while Alice votes
to hire Don if and only if he has written only promising against. Under majority rule, Don is hired even though
papers, but Alice wants to hire him if and only if he has everyone on the committee would prefer otherwise.
written at least one promising paper. Recall from the Thus Alice prefers to reveal her information truthfully.
argument above that, under common values, everyone On the other hand, when Bob and Chris have a more
voting informatively on the committee could not be an demanding bias than Alice, Alice’s information is only
equilibrium under majority rule, but such behaviour pivotal when both Bob and Chris have read exclusively
did constitute an equilibrium when a unanimous vote is promising papers. As before, Alice prefers to mislead
required to hire Don. Without common values, her colleagues in this event. Under majority rule,

Alice’s decision to reveal her information truthfully for a model of information aggregation with multiple
depends upon which event she thinks is more likely: the alternatives). Two intuitively appealing solutions to
event that Bob and Chris share her bias and have addressing such problems are, first, to tailor the voting
information similar to hers, or the event that they have rule to align incentives and induce individuals to vote
more stringent biases and have information different informatively and, second, to allow committee
from hers. If it is more likely that Alice’s information is members to communicate prior to voting.
similar to that of Bob and Chris then, by conditioning However, we have observed that even minimal
on being pivotal, Alice puts more weight on the event deviations away from common values make it imposs-
that they share her bias too and, therefore, prefers to ible to use voting rules alone to induce full information
tell the truth. revelation. When there is bias uncertainty among
Note that Alice’s incentive to reveal information committee members, so no member is sure of the
truthfully under majority rule does not extend to distribution of biases across the committee, it is
unanimity rule. Under unanimity rule, Alice can always sometimes possible to combine non-unanimous voting
ensure that Don is not hired by voting against him. It rules with pre-voting talk and recover full information
follows that she does not need to worry about the event equivalence. Under a more realistic, sequential, view of
in which she has the same preferences as Bob and information exchange as might occur in real committee
Chris, but misleads them into voting in favour of Don, deliberations, full information equivalence turns out to
for in this case she can veto Don being hired at the be difficult for any voting rule, at least when speaking
voting stage. Instead, the only event Alice must worry constitutes cheap talk.
about occurs when she prefers that Don be hired but The model of a committee using a simple voting rule
the other two do not, an event in which Alice has a strict preceded by cheap talk is useful for isolating the key
incentive not to report her information truthfully. This underlying strategic incentives that may frustrate
basic intuition is very general and leads to one of effective information aggregation. But there are many
the main results in Austen-Smith & Feddersen (2006): other ways in which committees might be organized,
full information revelation is not possible under communication could take place or a final decision
unanimity rule but may be achievable under other, could be reached. For example, a subset of committee
non-unanimous, voting rules (see also Doraszelski members might meet prior to any open discussion of
et al. 2003). the alternatives in committee, or there could be a
The Austen-Smith and Feddersen result exploits the disinterested mediator used to collect and disseminate
stylized communication protocol described earlier, in private information, and so on. The question such
which each committee member makes a report possibilities raise, therefore, concerns the best way to
simultaneously prior to voting. Suppose instead that organize committee communication to achieve collec-
committee members take turns speaking as they might tively desirable outcomes, in particular, full infor-
in a meeting. Later speakers can then condition their mation equivalence.
reports on what earlier contributors have revealed.
Although apparently more natural, this protocol has
the unfortunate by-product that later speakers can 6. GENERAL COMMUNICATION IN COMMITTEES
rule out some pivot events. In particular, if Alice is the It turns out that even without specifying details of
last speaker and has heard Bob and Chris both say exactly how people communicate prior to voting, it is
that they have read only promising papers, she knows possible to draw a critical distinction between unani-
that her speech only matters when the others have mity and non-unanimous voting rules. As before, by
more stringent biases than herself. Alice has an voting rule we mean that a given alternative is chosen if
incentive not to report truthfully but to induce Bob and only if it receives q or more votes. Under non-
and Chris to vote for Don by saying that she too has unanimous voting rules with at least three committee
read only promising papers, even when she has not. members, if everyone on the committee votes for the
Formalizing this logic more generally, Van Weelden same alternative then an individual’s vote cannot be
(2008) shows that Austen-Smith & Feddersen’s (2006) pivotal at the voting stage; given everyone is expected to
result on the impossibility of full information sharing vote for the same alternative under any non-unanimous
under the unanimity rule extends to all voting rules rule, changing any one person’s vote does not change
when communication is sequential. In other words, the outcome. Hence, any individual might as well vote
sequential public communication among committee with the consensus as vote against, in which case the
members, although more appealing descriptively, is in only strategic issue at stake is the impact of an
fact deleterious for information revelation. individual’s report on the consensus that eventually
forms. In an elegant contribution, Gerardi & Yariv
(2007) exploited this observation to show, first, that
5. THE STORY SO FAR any level of information aggregation achievable under
It is useful to sum up. In the simplest, non-trivial, any possible protocol for communication with some
setting in which three committee members with non-unanimous voting rule is equally achievable under
identical biases, or common values, make a decision any other such rule and, second, that the level of
between two given alternatives, if the members have information aggregation achievable under unanimous
private information about the relative worth of the voting rules is never more and can be strictly less than
alternatives under consideration, the voting rule may that with non-unanimous rules.
impede reaching full information equivalent to com- On the positive side, the Gerardi and Yariv result
mittee decisions or outcomes (see Austen-Smith (1990) suggests that the voting rule itself is essentially

irrelevant relative to the way in which communication Austen-Smith, D. & Feddersen, T. J. 2006 Deliberation,
takes place. This finding provides support for the preference uncertainty and voting rules. Am. Polit. Sci.
approach taken in the informal (i.e. not game theoretic) Rev. 100, 209–217. (doi:10.1017/S0003055406062113)
literature on deliberation that emphasizes the import- Austen-Smith, D. & Feddersen, T. J. 2007 A note on
ance of argument and reasons rather than the interplay preference uncertainty and communication in commit-
of voting rules and communication. Important contri- tees. Working Paper, Northwestern University.
Battaglini, M., Rebecca, M. & Palfrey, T. In press a. The
butions in the informal literature on communication
swing voter’s curse in the laboratory. Rev. Econ. Stud.
collective choice include the essays and references in Battaglini, M., Rebecca, M. & Palfrey, T. In press b.
the collections by Bohman & Rehg (1997) and Elster Information and pivotal voter models in large laboratory
(2000). Landa & Meirowitz (2007) attempted to elections. Am. Econ. Rev. Papers Proc.
connect the more formal approach reviewed in the Bohman, J. & Rehg, W. (eds) 1997 Deliberative democracy:
current essay to the concerns addressed in the informal essays on reason and politics. Cambridge, MA: MIT
literature. See also Hafer & Landa (2007) and Glazer & Press.
Rubinstein (2001, 2004), who develop alternative Coughlan, P. J. 2000 In defense of unanimous jury verdicts:
models of argument and communication to those mistrials, communication and strategic voting. Am. Polit.
deployed in the canonical game-theoretic approach. Sci. Rev. 94, 375–393. (doi:10.2307/2586018)
The Gerardi and Yariv result, however, does not imply Crawford, V. & Siobel, J. 1982 Strategic information
that details of the communication environment are transmission. Econometrica 50, 1431–1451. (doi:10.2307/
irrelevant. Indeed, there is a growing formal literature 1913390)
that explores such details. In particular, it is worth de Condorcet, M. 1994 [1785] Essai sur l’application de
l’analyse a la probabilite des decisions rendues a la pluralite des
noting recent efforts to incorporate incentives to
voix. Paris, France: L’Imprimerie Royale. [Trans. Iain
acquire information (see Persico 2004; Gerardi & McLean and Fiona Hewitt, Paris.]
Yariv 2008). Doraszelski, U., Gerardi, D. & Squintani, F. 2003
The literature on communication in committees Communication and voting with double-sided infor-
focuses on the relationship between communication mation. Contrib. Theor. Econ. 3, 6. (doi:10.2202/1534-
and voting. The central question in this literature 5971.1084)
concerns how committees might best share information Elster, J. (ed.) 2000 Deliberative democracy. Cambridge, UK:
given that they ultimately make collective choices by Cambridge University Press.
voting, a question that remains open. To provide an Farell, J. & Rabin, M. 1996 Cheap talk. J. Econ. Persp. 10,
answer, the issue of what the objectives of a hetero- 103–118.
geneous committee ought to be has to be resolved and Feddersen, T. J. & Pesendorfer, W. 1996 The swing voter’s
this is far from being a trivial problem. Indeed, to curse. Am. Econ. Rev. 86, 408–424.
evaluate the performance of any specific communication Feddersen, T. J. & Pesendorfer, W. 1997 Voting behavior and
protocol and voting rule relative to a given objective, it information aggregation in large elections with private
information. Econometrica 65, 1029–1058. (doi:10.2307/
is necessary to identify the best that is achievable
2171878)
relative to that objective under any conceivable protocol Feddersen, T. J. & Pesendorfer, W. 1998 Convicting the
or rule. In the simultaneous cheap talk setting, Austen- innoncent: the inferiority of unanimous jury verdicts. Am.
Smith & Feddersen (2005, 2007) found a variety Polit. Sci. Rev. 92, 23–35. (doi:10.2307/2585926)
of partially informative equilibria under both majority Gerardi, D. & Yariv, L. 2007 Deliberative committees.
and unanimity rule for three person committees. J. Econ. Theor. 134, 317–338. (doi:10.1016/j.jet.2006.05.
Comparing these equilibria suggests majority rule is 002)
superior to unanimity rule here, at least with respect to Gerardi, D. & Yariv, L. 2008 Information acquisition in
full information equivalence. But such results are very committees. Games Econ. Behav. 62, 436–459. (doi:10.
limited: they fail to establish any benchmark for what is 1016/j.geb.2007.06.007)
in fact more generally achievable even within the Gershkov, A. & Szentes, B. 2009 Optimal voting schemes
confines of this simple structure. To date, such results with costly information acquisition. J. Econ. Theor. 144,
have proved elusive (but see Meirowitz (2006) and 36–68. (doi:10.1016/j.jet.2008.02.004)
Gershkov & Szentes (2009) and for some steps in Glazer, J. & Rubinstein, A. 2001 Debates and decisions: on a
rationale of argumentation rules. Games Econ. Behav. 36,
this direction).
158–173. (doi:10.1006/game.2000.0824)
Glazer, J. & Rubinstein, A. 2004 On optimal rules of
persuasion. Econometrica 72, 1715–1736. (doi:10.1111/
REFERENCES j.1468-0262.2004.00551.x)
Austen-Smith, D. 1990 Information transmission in debate. Guarnaschelli, S., McKelvey, R. & Palfrey, T. 2000 An
Am. J. Polit. Sci. 34, 124–152. (doi:10.2307/2111513) experimental study of jury decision rules. Am. Polit. Sci.
Austen-Smith, D. 1992 Strategic models of talk in political Rev. 94, 407–423. (doi:10.2307/2586020)
decision making. Int. Polit. Sci. Rev. 13, 45–58. (doi:10. Hafer, C. & Landa, D. 2007 Deliberation as self-discovery
1177/019251219201300104) and the institutions for political speech. J. Theor. Polit. 19,
Austen-Smith, D. & Banks, J. S. 1996 Information aggrega- 329–360. (doi:10.1177/0951629807077573)
tion, rationality, and the Condorcet Jury Theorem. Am. Ladha, K. 1992 The Condorcet Jury Theorem, free speech
Polit. Sci. Rev. 90, 34–45. (doi:10.2307/2082796) and correlated votes. Am. J. Polit. Sci. 36, 617–634.
Austen-Smith, D. & Feddersen, T. J. 2005 Deliberation (doi:10.2307/2111584)
and voting rules. In Social choice and strategic decisions: Landa, D. & Meirowitz, A. 2007 Game theory, information
essays in honor of Jeffrey S. Banks (eds D. Austen-Smith & and deliberative democracy. Working Paper, New York
J. Duggan). Heidelberg, Germany: Springer. University.

McLennan, A. 1998 Consequences of the Condorcet Jury Ordeshook, P. & Palfrey, T. 1988 Agendas, strategic voting
Theorem for beneficial information aggregation by and signaling with incomplete information. Am. J. Polit.
rational players. Am. Polit. Sci. Rev. 92, 413–418. Sci. 32, 441–466. (doi:10.2307/2111131)
(doi:10.2307/2585673) Persico, N. 2004 Committee design with endogenous
Meirowitz, A. 2006 Designing institutions to aggregate information. Rev. Econ. Stud. 71, 165–194. (doi:10.
preferences and information. Q. J. Polit. Sci. 1, 373–392. 1111/0034-6527.00280)
(doi:10.1561/100.00005011) Van Weelden, R. 2008 Deliberation rules and voting.
Meirowitz, A. 2007 In defense of exclusionary deliberation: Q. J. Polit. Sci. 3, 83–88. (doi:10.1561/100.00007062)
communication and voting with private beliefs and values. Wit, J. 1998 Rational choice and the Condorcet Jury
J. Theor. Polit. 19, 329–360. (doi:10.1177/09516298070 Theorem. Games Econ. Behav. 22, 364–376. (doi:10.
77573) 1006/game.1997.0596)

Phil. Trans. R. Soc. B (2009) 364, 771–779

doi:10.1098/rstb.2008.0258
Evolution of signalling systems with multiple

senders and receivers
Brian Skyrms*
Department of Logic and Philosophy of Science, University of California Irvine,
3151 Social Science Plaza A, Irvine, CA 92717-5100, USA
Sender–receiver games are simple, tractable models of information transmission. They provide a
basic setting for the study the evolution of meaning. It is possible to investigate not only the
equilibrium structure of these games but also the dynamics of evolution and learning—with
sometimes surprising results. Generalizations of the usual binary game to interactions with multiple
senders, multiple receivers or both provide the elements of signalling networks. These can be seen as
the loci of information processing, group decisions, and teamwork.
Keywords: signalling; information; teamwork; population dynamics
1. INTRODUCTION They may sometimes be separable in human affairs,

To coordinate action, information must be transmitted, but elsewhere in nature it is more typical that they have
processed and used to make decisions. Transmission of coevolved. It is possible to construct simplified models
information requires the existence of a signalling system that capture essential aspects of these issues as
in which the signals that are exchanged are coordinated evolutionary games.
with the appropriate content. Signalling systems in These models may also be viewed as modules that,
nature range from quorum signalling in bacteria once evolved, may be put together to form more
(Schauder & Bassler 2001; Taga & Bassler 2003; Kaiser complex interactions. Evolutionary games may be
2004), through the dance of the bees (Dyer & Seeley studied from a both a static and dynamic point of
1991), birdcalls (Hailman et al. 1985; Gyger et al. view. Static analysis of equilibria reveals a lot about the
1987; Evans et al. 1994; Charrier & Sturdy 2005) and structure of the interaction, and it can be carried out at
alarm calls in many species (Cheney & Seyfarth 1990; a level of generality that does not commit one to a
Seyfarth & Cheney 1990; Green & Maegner 1998; particular dynamics. But dynamic analysis sometimes
Manser et al. 2002), to human language. reveals complexities that are not immediately apparent
Information processing includes filtering—that is from the study of equilibria. Dynamic analyses may be
discarding irrelevant information and passing along mathematically challenging. Computer simulations are
what is important—and integration of multiple pieces of always available as a tool, but in these simple game-
information. Filtering systems are ubiquitous. Quorum- theoretic models, analytic methods are also applicable.
sensing bacteria disregard low levels of signalling We start with dyadic sender–receiver games—one
molecules, and only respond to concentrations appro- sender and one receiver—and then generalize the
priate to action. The black-capped chickadee Poecile, model to multiple senders and receivers. It can be
(Poecile atricapilla) disregards calls that lack the syntactic shown that surprisingly sophisticated behaviour can
structure that identifies a chickadee origin. Every sensory emerge from the dynamics of evolution. A full analysis,
processing system of a multicelled organism decides what however, is non-trivial in even the simplest dyadic
information to discard and what to transmit. Integration signalling games, and much remains to be done.
includes computation, logical inference and voting.
Although we usually think of these operations in terms
of conscious human thought, they can also be performed 2. CLASSIC TWO-AGENT SENDER–RECEIVER
unconsciously by simple signalling networks. Finally, GAMES: EQUILIBRIUM CONSIDERATIONS
information must be used to make decisions. These In the basic model (Lewis 1969), there are two players:
decisions may have fitness consequences for the whole the sender and the receiver. Nature chooses a state with
group, down to the level of quorum sensing in bacteria some probability (each state having non-zero prob-
and up to alarm calls and signals indicating location and ability of being chosen) and the sender observes the
quality of food sources. state. The sender then sends a signal to the receiver,
From an evolutionary perspective, these three aspects who cannot observe the state directly but does observe
of coordination are best addressed simultaneously. the signal. The receiver then chooses an act, the
outcome of which affects them both, with the pay-off
*bskyrms@uci.edu depending on the state. We assume at the onset that the
Electronic supplementary material is available at http://dx.doi.org/10. numbers of states, signals and acts are equal. Where
1098/rstb.2008.0258 or via http://rstb.royalsocietypublishing.org. this number is N, we refer to this as an N!N!N game.
One contribution of 11 to a Theme Issue ‘Group decision making in There is pure common interest between sender and
humans and animals’. receiver—they get the same pay-off. There is exactly

772 B. Skyrms Signalling: multiple senders and receivers
one ‘correct’ act for each state. In the correct act–state Among these equilibria, the signalling systems yield
combination, they both get a pay-off of 1, otherwise optimal pay-off, but this is no guarantee that one will
pay-off is 0. We number the states and acts, so that in a arrive at them. They also, however, have the distinction
play of the game, hstate, signal, actiZhsi , mj, aki, the pay- of being strict; that is to say, any unilateral deviation
off is 1 if iZk, 0 otherwise. results in a strictly worse pay-off. This has the
A sender’s strategy consists of a function from states immediate consequence that in an evolutionary setting,
to signals and a receiver’s from signals to acts. Expected a signalling system is an evolutionarily stable state of the
pay-offs are determined by the probability with which population. This is true both in a two-population
nature chooses states and the population proportions evolutionary model, with a population of senders and
of sender’s and receiver’s strategies. For the purposes of receivers and in a one-population model in which an
evolution, individual senders and receivers are assumed individual is sometimes in a sender role and sometimes
to have deterministic strategies. in a position of being a receiver.
Signals are not endowed with any intrinsic meaning. It is also easy to see that signalling systems are the
If they are to acquire meaning, the players must only evolutionarily stable states ( Wärneryd 1993). In
somehow find their way to an equilibrium where the pooling example above, a mutant sender who
information is transmitted. When transmission is always sent signal 2 would do just as well as the native
perfect, so that the act always matches the state and population. Likewise, a mutant receiver whose strategy
the pay-off is optimal, Lewis calls the equilibrium a responded differently to the signal 3 (which is never
signalling system. For instance, in a 3!3!3 game, the sent) would not suffer for doing so. In the partial
following combination of strategies is a Lewis signalling pooling example, a mutant sender who sent signal 2 in
system equilibrium: states 2 and 3 would elicit the same receiver response,
and thus would have the same pay-off as the natives.
sender receiver In each of these cases, the mutants do not do better
state 10 signal 3 signal 30 act 1 than the natives. The pooling and partial pooling
; equilibria are equilibria. But the mutants do no worse,
state 20 signal 2 signal 20 act 2 so they are not driven out. That is to say, pooling and
state 30 signal 1 signal 10 act 3 partial pooling equilibria fail the test for evolutionary
stability (Maynard Smith & Price 1973; Maynard
as is any combination of strategies that can be gotten Smith 1982). Equilibrium analysis might then lead
from this one by permutation of signals. The ‘meaning’ one to suspect that evolutionary dynamics would
of the signals is thus purely conventional, depending on always (or almost always) take us to signalling
the equilibrium into which the agents have settled. systems. It is not so (Huttegger 2007a,b, forthcoming;
There are also other equilibria in the signalling Pawlowitsch 2008).
games. There are pooling equilibria, in which the sender
ignores the state and the receiver ignores the signal. For
example, suppose that state 3 is the most probable. 3. DYNAMICS
Then, the following is a pooling equilibrium: The simplest dynamic model of differential reproduc-
tion for a large population is the replicator dynamics
sender receiver (Taylor & Jonker 1978; Hofbauer & Sigmund 1998).
state 10 signal 1 signal 30 act 3 Replicator dynamics has an alternative interpretation
: as a model of cultural evolution by imitation of
state 20 signal 1 signal 20 act 3
successful strategies (Björnerstedt & Weibull 1995;
state 30 signal 1 signal 10 act 3 Schlag 1998). It has a third interpretation as a limiting
case of reinforcement learning (Beggs 2005; Hopkins &
Since the sender conveys no information, the Posch 2005).
receiver can do no better than choose the act that We can consider a one-population model where
pays off in the most probable state. Since the receiver strategies are conditional (if the sender does this, if the
ignores the signal, the sender can do no better by receiver does that), or a two-population model with one
changing his signalling strategy. population of senders and another population of
In N!N!N games with NO2, there are also partial receivers. Both have biological applications. A two-
pooling equilibria, for example, population model is clearly appropriate for inter-
sender receiver species signalling. In case of same species alarm calls,
individuals are sometimes in the role of sender and
state 10 signal 3 signal 30 act 1 sometimes that of receiver.
:
state 20 signal 1 signal 20 act 3 For a single population, let the strategies be {Si}, let
xi be the population proportion of those who use
state 30 signal 1 signal 10 act 3 strategy Si and let the fitness of strategy Si played
The sender’s strategy does not discriminate between against Sj be denoted W(SijSj). Then, assuming
states 2 and 3 and leaves signal 2 unused. Upon random matching, the average fitness of strategy Si is
X
receiving the ‘ambiguous’ signal, the receiver chooses W ðSi Þ Z x W ðSi jSj Þ;
j j
optimally, given the limited information that was
transmitted. For larger N, there are more kinds of and the average fitness of the population is
partial pooling equilibria, depending on which states X
are ‘pooled’. W ðSÞ Z i
W ðSi Þxi :

Signalling: multiple senders and receivers B. Skyrms 773
The replicator dynamics is the system of differential attraction. It has been shown analytically that it does
equations (Hofbauer & Huttegger 2008). Simulations show that
dxi the size of the basin of attraction need not be negligible.
Z xi ½W ðSi ÞKW ðSÞ: The size depends, as would be expected, on the
dt difference in the probabilities of the two states. If we
For the two-population case, let xi be the population were to depart from the assumption that the states have
proportion of those who use strategy Si in the equal pay-offs, it would also depend on the magnitudes
population of senders and yi be the population of of the pay-offs.
those who use strategy R i in the population of receivers. Even if we keep the states equiprobable and the
We again assume random matching of senders and magnitudes of the pay-offs equal, almost sure conver-
receivers, so that gence to a signalling system is lost as we move from 2!
X 2!2 to 3!3!3. In this game, total pooling equilibria
W ðSi Þ Z y W ðSi jRj Þ and
j j are dynamically unstable, but there are sets of neutrally
X stable partial pooling equilibria as the ones discussed in
W ðRj Þ Z i
xi W ðRj jSi Þ: the last section. It can be shown analytically that the set
of partial pooling equilibria has a positive basin of
The average fitnesses of the sender and receiver
attraction, and simulation shows that this basin is not
populations, respectively, are
negligible (Huttegger et al. in press).
X X Even with the strong common interest assumptions
W ðS Þ Z i
W ðSi Þxi and W ðRÞ Z j
W ðRj Þ yi :
built into Lewis’ signalling games, the emergence of
signalling is not quite the sure thing that it may initially
We consider the evolution of this two-population
have seemed on the basis of equilibrium consider-
system using bipartite replicator dynamics (Taylor &
ations. Perfect signalling systems can evolve, but it is
Jonker 1978; Hofbauer & Sigmund 1998)
not guaranteed that they will do so. Dynamic analysis
dxi has revealed unexpected subtleties.
Z xi ½W ðSi ÞKW ðSÞ; There are more subtleties to explore, because the
dt
sets of suboptimal equilibria are not structurally stable
dyj (Guckenheimer & Holmes 1983; Skyrms 1999) Small
Z yj ½W ðRj ÞKW ðRÞ: perturbations of the dynamics can make a big
dt
difference. The natural perturbation to pure differential
In both the one- and two-population models of reproduction that needs to be considered is the
Lewis’ signalling games, the strong common interest addition of a little mutation. We can move from the
between the sender and receiver assures global conver- replicator dynamics to the replicator–mutator
gence of the replicator dynamics; all trajectories must dynamics (Hadeler 1981; Hofbauer 1985). For a
lead to dynamic equilibria (Hofbauer & Sigmund two-population model with uniform mutation, this is
1998; Huttegger 2007a,b).
In the case of a 2!2!2 Lewis signalling game, with dxi
Z xi ½ð1KeÞW ðSi ÞKW ðSÞ C ðe=nÞW ðSÞ;
states equiprobable, the ‘hasty conclusion’ from dt
evolutionary stability equilibrium analysis is, in fact,
born out by the dynamics. Equilibria other than the dyj
signalling systems are all dynamically unstable. In both Z yj ½ð1KeÞW ðRj KW ðRÞ C ðe=nÞW ðRÞ;
dt
two- and one-population models, replicator dynamics
carries almost all possible population proportions to a where e is the mutation rate and n is the number of
signalling system (Huttegger 2007a,b,c; Hofbauer & strategies. We include all possible strategies. Evolution-
Huttegger 2008). ary dynamics is now governed by a sum of selection and
But if states are not equiprobable, this is no longer mutation pressures. Mutation pressure pushes towards
so. Suppose that state 2 is much more probable than all strategies being equiprobable, where mutation into a
state 1. Then, the receiver might just do the act that is strategy would equal mutation out. Mutation pressure
the best in state 2 and ignore the signal. And since the can be counterbalanced or overcome by selection
signal is being ignored, the sender might as well ignore pressure. But if selection pressure is weak or non-
the state. Consider a population in which receivers existent, mutation can cause dramatic changes in the
always do act 2, some senders always send signal 1 and equilibrium structure of the interaction.
some always send signal 2. Any such population is an We can illustrate by returning to the 2!2!2
equilibrium. We have described a set of polymorphic signalling game, two-population states with unequal
pooling equilibria. These equilibria are dynamically probability. Suppose state 2 is more probable than
stable, even though they are not evolutionarily stable in state 1. Then, as we have seen, there is a set of pooling
the sense of Maynard-Smith & Price (1973). They are equilibria for the replicator dynamics. In the receiver
not strongly stable attractors in the dynamics. Rather, population, the strategy of always doing act 2 (no
they are ‘neutrally stable’, in that points near them stay matter what the state is) goes to fixation. In the sender
near them under the action of the dynamics. But they population, there is a polymorphism between two types
do not attract all points near them. For instance, other of sender. One sends signal 1, no matter what the state
pooling equilibria near them are not moved at all by the is; the other sends signal 2, no matter what the state is.
dynamics. The question is whether this set of pooling Since there is no selection pressure between the
equilibrium, considered as a whole, has a basin of senders’ types, every such sender polymorphism is an

equilibrium. Addition of any amount of uniform Table 1. Pay-offs if sending signal is costly.
mutation leads the set of pooling equilibria to collapse
to a single point at which ‘Always send signal 1’ and receiver 1 receiver 2 receiver 3 receiver 4
‘Always send signal 2’ are represented with equal
probability (Hofbauer & Huttegger 2008). But all sender 1 2-c, 2 1-c, 1 1.5-c, 1.5 1.5-c, 1.5
other strategies are also present in small amounts at this sender 2 1-c, 1 2-c, 2 1.5-c, 1.5 1.5-c, 1.5
population state, due to the action of mutation. sender 3 1.5-2c, 1.5 1.5-2c, 1.5 1.5-2c, 1.5 1.5-2c, 1.5
sender 4 1.5, 1.5 1.5, 1.5 1.5, 1.5 1.5, 1.5
The big question concerns the stability properties of
this perturbed pooling equilibrium. Is it dynamically stable
or unstable? There is no unequivocal answer. It Table 2. Pay-offs if receiving signal is costly.
depends on the disparity in the probability between
the two states (Hofbauer & Huttegger 2008). A little receiver 1 receiver 2 receiver 3 receiver 4
mutation can help the evolution of signalling systems,
but does not always guarantee that they evolve. sender 1 2-0.1, 2-0.1 1-0.1, 1-0.1 1.33-0.1, 1.67-0.1,
1.33 1.67
sender 2 1-0.2, 1-0.1 2-0.2, 2-0.1 1.33-0.2, 1.67-0.2,
1.33 1.67
4. COSTS sender 3 1.5-0.3, 1.5-0.3, 1.33-0.3, 1.67-0.3,
Let us return to the case of 2!2!2, with states 1.5-0.1 1.5-0.1 1.33 1.67
equiprobable, but assume that one of the signals costs sender 4 1.5, 1.5-0.1 1.5, 1.5-0.1 1.33, 1.33 1.67, 1.67
something to send, while the other is cost free. (We could
interpret the cost-free signal as just keeping quiet.)
Now there are pooling equilibria in which the sender Table 3. Pay-offs if costs are state specific.
always sends the cost-free signal and there are various
proportions of receiver types. receiver 1 receiver 2 receiver 3 receiver 4
Denoting the sender’s strategies as
sender 1 2, 2-0.1 1, 1-0.1 1.33, 1.33 1.67, 1.67
sender 1 : state 10 signal 1; state 20 signal 2 sender 2 1-0.3, 1-0.1 2-0.3, 2-0.1 1.33-0.3, 1.67-0.3,
1.33 1.67
sender 2 : state 10 signal 2; state 20 signal 1 sender 3 1.5-0.2, 1.5-0.2, 1.33-0.2, 1.67-0.2,
sender 3 : state 10 signal 1; state 20 signal 1 1.5-0.1 1.5-0.1 1.33 1.67
sender 4 1.5-0.1, 1.5-0.1, 1.33-0.1, 1.67-0.1,
sender 4 : state 10 signal 2; state 20 signal 2 1.5-0.1 1.5-0.1 1.33 1.67
and the receiver’s strategies as
does strictly worse. Thus, in both one- and two-
receiver 1 : signal 10 act 1; signal 20 act 2
population evolutionary models, it is evolutionarily
receiver 2 : signal 10 act 2; signal 20 act 1 stable and a strong (attracting) equilibrium in the
: replicator dynamics.
receiver 3 : signal 10 act 1; signal 20 act 1
If costs are state specific, a rosier picture is possible
receiver 4 : signal 10 act 2; signal 20 act 2 (Zahavi 1975). We alter the previous example so that
If signal 1 is costly, costZ2c, states equiprobable signal 1 is free in state 1 but costs 0.3 in state 2
and a background fitness is 1, we have the pay-off and signal 2 is free in state 2 but costs 0.3 in state 1.
matrix (sender’s pay-off, receiver’s pay-off ), as shown Sender 1 now pays no penalty; sender 2 always pays
in table 1. 0.3; sender 3 pays 0.3 two-thirds of the time (Z0.2)
Sender’s strategies 1 and 2 pay the cost half the and sender 4 pays 0.3 one-third of the time (Z0.1).
time, strategy 3 all the time and strategy 4 never. Pure This is shown in table 3.
Nash equilibria of the game for small c are italic-faced. The pooling state, hsender 4, receiver 4i, is no longer
(If cO0.5, it is never worth the cost to send a signal, an equilibrium at all. Given that the receiver is ignoring
and the signalling system equilibria disappear.) There the message, the sender is better off switching to the
is also a large range of mixed strategies (corresponding costless strategy, sender 1. If so, the receiver is better off
to the receiver polymorphisms) that are equilibria. switching to receiver 1, yielding the optimal signalling
States when receiver types are approximately equally system hsender 1, receiver 1i. Optimality, however, may
represented and senders always send the costless signal not evolve. The suboptimal signalling system hsender 2,
are such pooling equilibria. receiver 2i, in which the sender uses the ‘wrong’ signals
It might also cost the receiver something to listen. Let us and always pays a signalling cost, is also a strict
combine this with a costly message and unequal state equilibrium. Both signalling systems are strong
probabilities. For example, let the probability of state 1 (attracting) equilibria in both one- and two-population
be 1/3, the cost of signal 1 0.3 and the cost of the replicator dynamic models.
receiver paying attention to the signals 0.1. The
background fitness is 1. Then, the foregoing pay-off
matrix changes to that displayed in table 2. 5. SIGNALLING NETWORKS
The pooling equilibrium, hsender 4, receiver 4i, where There is no reason to limit ourselves to signalling
the sender always sends signal 2 and the receiver between just two actors: one sender and one receiver.
always does act 2, is now a strict Nash equilibrium of In fact, most signalling systems in nature involve
the game. Either the sender or receiver who deviates multiple senders, multiple receivers or both. If a

Table 4. Pay-offs with two senders and one receiver. system can evolve. In each of these settings, a signalling
system is a strongly stable (attracting) equilibrium in
act 1 act 2 act 3 act 4 the replicator dynamics.
Each sender’s signal conveys perfect information
state 1 1,1,1 0,0,0 0,0,0 0,0,0 about her observation—about the partition of the states
state 2 0,0,0 1,1,1 0,0,0 0,0,0 of the world which she can see. The combination of
state 3 0,0,0 0,0,0 1,1,1 0,0,0 signals has perfect information about the states of the
state 4 0,0,0 0,0,0 0,0,0 1,1,1
world. Exactly one state corresponds to each com-
bination of signals. And the receiver puts the signals
receiver gets signals carrying different pieces of together. The receiver’s acts contain perfect infor-
information from different senders, the signalling mation about the state of the world. The signalling
system is called upon to solve some problem of system simultaneously solves problems of transmission and
information processing. Consider a toy model with integration of information.
two senders and one receiver The basic model admits of interesting variations. Of
course, there may be more senders. And depending on
†/ †) † the act set available to the receiver, he may draw the
appropriate logical ‘conclusion’ from the ‘premises’
Signalling complementary information. There are four supplied by the various senders (Skyrms 2000, 2004,
states of nature, each of which occurs with non-zero 2008). The senders’ partitions may not be fixed by
probability. Two individuals are situated so as to make nature, but may themselves evolve in the presence of
different incomplete observations of the state. The first information bottlenecks (Barrett 2006, 2007a,b).
sees whether it is in {S1, S2} or in {S3, S4} and the Error: There is another class of multiple sender
second sees whether it is in {S1, S3} or in {S2, S4}. models, where the question is not one of comp-
Together, they have enough information to pin down lementary information but one of error. In the previous
the state of nature, but separately they do not. Each example, senders observed different partitions but
sends one of two signals to a receiver who must choose there was no error in identifying the true element of
one of four acts. Let us say the first sender chooses ‘red’ the partition. Here, we suppose that the senders
or ‘green’ and the second chooses ‘blue’ or ‘yellow’. all observe the same states but with some error in
The pay-offs favour cooperation. Exactly one act is correctly identifying them. (An alternative, essentially
‘right’ for each of the states, in that each of the equivalent, interpretation of the model would locate
individuals is reinforced just in case the right act for the errors in the transmission of the signals.)
the state is chosen. For the simplest model, suppose that there are only
In this extended Lewis signalling game, the observa- two states and two acts. States are equiprobable. Three
tional situation of sender 1 is characterized by a senders observe the states with error probability of
partition of the states, O1Z{{S1, S2}, {S3, S4}}. Her 10 per cent, with the errors being independent
signalling strategy is a function from the elements of between senders and between trials. Each sender
this partition into her set of signals, {R, G}. Likewise sends a message to the receiver, who must then choose
sender 2 in observational situation O2Z{{S1, S3}, one of the two acts. As before, we assume that act 1
{S2, S4}} has a signalling strategy that maps the pays off 1 for everyone involved in state 1 and act 2 pays
elements of her partition into her signal set, {B, Y}. off 1 for everyone involved in state 2. Otherwise, no one
The receiver’s strategy maps pairs of signals {{R, B}, gets anything.
{R, Y}, {G, B}, {G, Y}} into her set of acts {A1, A2, Nature here first flips a coin to pick a state, and then
A3, A4}. picks apparent states to present to the three senders
All agents get pay-off 1 just in case the receiver according to the error probabilities. A sender’s strategy
correctly identifies the state and does the appropriate is a function from apparent state into the set of signals,
act. Pay-offs are shown in table 4. {S1, S2}. We have a choice about how to set up the
A signalling system equilibrium is a combination of receiver’s strategies. If we were to assume that the
sender and receiver strategies such that pay-off is equal receiver could distinguish between senders, we could
to 1 in each state. As before, a signalling system is a take the receiver’s strategy to be a function from
strict equilibrium of the game, and the signalling systems ordered triples of signals to acts. But here we assume
are the only strict equilibria. There are lots of pooling that the receiver cannot distinguish between hS1, S2,
and partial pooling equilibria. S1i, hS1, S1, S2i and hS1, S1, S2i. The receiver here has
In an evolutionary setting, this three-player game an observational partition and can only count signals.
gives rise to three-, two- and one-population models. In This might be thought of as discrete approximation to
a one-population model, an individual’s strategy would a situation where the receiver perceives an intensity
be of the form: if sender in observational situation O1 has arising from many chemical signals, or the sound
this sender’s strategy; if sender in observational situation O2 intensity arising from many calls. A receiver’s strategy is
has that sender’s strategy; and if receiver has this strategy. then a function from the frequencies of signal received
The most natural two-population model has a popu- to act.
lation of senders with different observational roles and Optimal signalling in this model consists in what we
a population of receivers. In all three evolutionary might call a Condorcet equilibrium (see List et al. 2009;
settings, signalling systems are the unique evolution- Sumpter & Pratt 2009). There is one signal that the
arily stable states. It is no longer certain that a signalling senders all use for apparent state 1 and another that
system must evolve, but it is certain that a signalling they all use for apparent state 2. The receiver goes with

a majority vote. For instance, if the senders all Table 5. Pay-offs in a simple teamwork situation.
send signal 2 in state 1, the receiver will do act 2 if
two or more senders send signal 2 and act 1 otherwise. hA1, A1i hA1, A2i hA2, A1i hA2, A2i
In our example, individuals at a Condorcet equilibrium
reduce their error rate from 10 per cent to under 3 per state 1 1,1,1 0,0,0 0,0,0 0,0,0
cent. This can be viewed as an example of information state 2 0,0,0 1,1,1 0,0,0 0,0,0
state 3 0,0,0 0,0,0 1,1,1 0,0,0
filtering, as explained in §1.
state 4 0,0,0 0,0,0 0,0,0 1,1,1
Rather than thinking of evolution taking place
solely in the context of this game, we might assume
that sender’s strategies already evolved in the context of always gets pay-off of 1. A signalling system equilibrium
single sender–receiver interactions. Then, receivers is again a strict equilibrium, and the unique strict
usually get one signal, or multiple agreeing signals equilibrium in the game. It is a strongly stable attractor
according to the evolved signalling system, but in the replicator dynamics.
occasionally get disagreeing signals. Slow adaptation The example can be varied in many ways, some
for mixed signals in such an environment is a simple more interesting than others. The two receivers can be
problem of optimization. thought of as playing a rather trivial two-person game,
Against these fixed sender strategies, receivers who but the game is different in every state of the world.
go with the majority of senders will have the greatest In a signalling system, the sender can be thought of
fitness. Then replicator dynamics will converge to the either as conveying information about the game or the
optimal receiver strategy (Hofbauer & Sigmund 1998). optimal act to be done. In these trivial games, these are
But suppose we forego this easy route and ask equivalent. The example could be varied by changing
whether Condorcet signalling equilibria can evolve in the four embedded two-person games and their effect
the context of the original four-person game. Both the on the pay-offs to the sender.
sender’s signals and the receiver’s voting rule must Chains: Information can flow further than that
coevolve. It is still possible for efficient signalling to shown in the models given so far. Signallers can form
evolve. The Condorcet equilibria are strict. Conse- chains, so that information is passed along until it
quently, they are stable attractors in the evolutionary reaches an endpoint at which it can be used. Consider a
versions of this game using replicator dynamics. In fact, little signalling chain
simulations show the Condorcet equilibria almost
†/ †/ †
always evolving in the foregoing model (see the
electronic supplementary material). There is a sender, an intermediary and a receiver.
Variations in the parameters of the model may well Nature chooses one of two states with equal probability.
lead to the evolution of voting rules different from The sender observes the state, chooses one of two
majority rule. This is an area open for exploration. signals and sends it to the intermediary, the inter-
Recent rational-choice literature on strategic mediary observes the sender’s signal, chooses one of
voting (Austen-Smith & Banks 1996; Feddersen & her own two signals and sends it to the receiver. The
Pesendorfer 1998) is a source of a rich set of models receiver observes the intermediary’s signal and chooses
that can be transposed to an evolutionary setting. one of two acts. If the act matches the state, sender,
Teamwork: It is sometimes the case that a well-placed intermediary and receiver all get a pay-off of 1,
sender knows what needs to be done, and can send otherwise a pay-off of 0.
messages to receivers who can act, but that no one Suppose that the set of potential signals available to
receiver can do everything that needs to be done. The the sender is {R, B}, and that available to the receiver is
sender may be the foreman, the commander or the {G, Y}. A sender’s strategy is a function from {S1, S2}
brain of an organism—the team leader. Success for all into {R, B}, an intermediary’s from {R, B} into {G, Y}
requires teamwork. and a receiver’s from {G, Y} into {A1, A2}. A signalling
There may be one sender and multiple receivers system here is a triple of strategies such that the
composition of sender’s, intermediary’s and receiver’s
†) †/ † strategies maps state 1 to act 1 and state 2 to act 2.
Signalling systems are the unique strict equilibria in this
For a simple teamwork problem, we suppose that game and the unique evolutionarily stable states in the
there are two receivers and one sender. The sender corresponding one-, two- and three-population signal-
observes one of four equiprobable states of the world ling games. They are attractors in the replicator
and sends one of two signals to each receiver. The dynamics. In principle, signalling chains can evolve
receivers must each choose between two acts, and out of nothing.
the acts must be coordinated in a way determined However, simulations show that in this case
by the state for all to get a pay-off. We take pay-offs to evolution is very slow when compared with the other
be as shown in table 5. signalling games discussed so far. This may simply be a
We assume that the sender can distinguish the consequence of the multiplicity of coordination pro-
members of the team; so the sender’s strategy maps blems that need to be solved simultaneously. The speed
states into ordered pairs of signals and a receiver’s with which the chain signalling system can evolve is
strategy maps her signal into her space of acts. Here, much improved if the sender and receiver have pre-
the problem to be solved is a combination of one of existing signalling systems. They could be the same
communication and one of coordination. It is solved in signalling system, which would be plausible if the
a signalling system equilibrium, in which everyone sender and receiver were the members of the same

population, but the signalling systems need not be the Table 6. Pay-offs in a dialogue situation.
same. The sender and receiver can have different
‘languages’, so that the intermediary has to act as a decision 1 decision 1 decision 2 decision 2
‘translator’ or signal transducer. Suppose that the act 1 act 2 act 3 act 4
sender sends red or blue and the ultimate receiver
reacts to green or yellow as follows: state 1 1 0 1 0
state 2 1 0 0 1
sender receiver state 3 0 1 1 0
state 10 R G0 act 2 : state 4 0 1 0 1
state 20 B Y0 act 1
a plasticity of signalling that is absent in fixed sender–
A successful translator must learn to receive one receiver games. Signalling systems are strict and
signal and send another, so that the chain leads to a evolutionarily stable as before.
successful outcome. Signalling systems can evolve in the dialogue
sender translator receiver interaction in isolation, but simulations show this
state 10 R see R0 send Y Y0 act 1 : process to be very slow. As in the case of chains,
evolution of a signalling system is much easier if we
state 20 B see B0 send G G0 act 2 assume that some of its parts have evolved in less
The translator’s learning problem is now really quite complicated interactions. Player 1 may already have
simple. The requisite strategy strictly dominates all signalling systems in place for the two different
alternatives. It pays off all the time, while the strategies observational partitions as a consequence of evolution
always send Y and always send G pay off half the time, in simple sender–receiver interactions. If so, the
and the remaining possibility always leads to failure. evolution of dialogue only requires that the second
The dominated strategies are eliminated (Hofbauer & player signals the problem and the first chooses what to
Sigmund 1998), and the correct strategy evolves. observe. This is no more difficult than the evolution of a
Dialogue: The chain model showed one way in which signalling system in the original Lewis signalling game.
simple interactions could be strung together to form
more complex signalling systems. Here is another.
Suppose that a sender’s observational partition is not 6. CONCLUSION
fixed. The sender can choose which observation to We have investigated the evolution of signalling in some
make. That is to say, she can choose which partition of modest extensions of Lewis signalling games with
states to observe. Suppose also that the receiver’s multiple senders and receivers. This discussion has
decision problem is not fixed. Nature chooses focused on one particular setting—a large (infinite)
a decision problem to present to the receiver. Different population or several large populations with random
sorts of information are relevant to different decision interactions between individuals. Different settings
problems. Knowing the actual element of partition would call for different relevant dynamics. A small
A (the element that contains the actual state) may population with random encounters calls for a
be relevant to decision problem 1, while knowing stochastic model of evolution, with either a growing
the actual element of partition B may be relevant population or one whose size is fixed at some carrying
to decision problem 2. This opens up the possibility capacity (Shreiber 2001; Benaim et al. 2004; Taylor
of signalling dialogue, where information flows in et al. 2004). Pawlowitsch (2007) has shown that in one
two directions kind of finite population model, efficient proto-
†4 † languages are the only strategies that are protected by
selection. Individuals might interact with neighbours in
In the simplest sort of example, nature flips a coin some spatial structure (Grim et al. 2002; Zollman
and presents player 2 with one or another decision 2005). Isolated individuals might invent signalling
problem. Player 2 sends one of two signals to player 1. systems by trial-and-error learning in repeated
Player 1 selects one of two partitions of the state of interactions (Skyrms 2004, 2008; Barrett 2006,
nature to observe. Nature flips a coin and presents 2007a,b), which might then spread by a process of
player 1 with the true state. Player 1 sends one of two cultural evolution (Komarova & Niyogi 2004). In fact,
signals to player 2. Player 2 chooses one of two acts. urn models of reinforcement learning are very close to
Suppose that there are four states, {S1, S2, S3, S4}, those in a small, growing population (Shreiber 2001;
with alternative partitions: P1Z{{S1, S2}, {S3, S4}}, Benaim et al. 2004). It has been recently proved that
P2Z{{S1, S3}, {S2, S4}}. The two decision problems reinforcement dynamics in the simplest Lewis signal-
require choices in different act sets: D1Z {A1, A2}, ling game—2!2!2 states equiprobable—converges
D2Z{A3, A4}. Pay-offs for the two decision problems with probability 1 to a signalling system (Argiento et al.
are shown in table 6. in press). Such an analytic treatment of reinforcement
Player 2 has a signal set {R, G} and player 1 has a learning does not yet exist for more complicated
signal set {B, Y}. A strategy for player 2 now consists of signalling interactions, but simulations tend to
three functions: a sender strategy from {P1, P2} into give results parallel to the evolutionary analysis
{R, G}; a receiver strategy form {B,Y} into {A1, A2}; given here. This is not entirely surprising, given
and a receiver strategy from {B, Y} into {A3, A4}. In a the close connections between reinforcement learning
signalling system equilibrium, each player gets always a and the replicator dynamics (Beggs 2005; Hopkins &
pay-off of 1. The possibility of dialogue introduces Posch 2005).

Simple models such as those discussed here can be Cheney, D. & Seyfarth, R. 1990 How monkeys see the world:
assembled into more complex and biologically inter- inside the mind of another species. Chicago, IL: University of
esting systems. The network topologies themselves may Chicago Press.
evolve (Bala & Goyal 2000; Skyrms & Pemantle 2000). Dyer, F. C. & Seeley, T. D. 1991 Dance dialects and foraging
There are all sorts of interesting variations. For range in three asian honey bee species. Behav. Ecol.
Sociobiol. 28, 227–233. (doi:10.1007/BF00175094)
instance, signalling networks may allow eavesdroppers,
Evans, C. S., Evans, C. L. & Marler, P. 1994 On the meaning
a case well studied in (McGregor 2005). But the main of alarm calls: functional reference in an avian vocal
business of signalling networks is to facilitate successful system. Anim. Behav. 73, 23 – 38.
collective action. The simple models studied here focus Feddersen, T. & Pesendorfer, W. 1998 Convicting the
on the crucial aspects of coordinated action. Infor- innocent: the inferiority of unanimous jury verdicts
mation is acquired by the units of the group. It is under strategic voting. Am. Pol. Sci. Rev. 92, 23 – 35.
transmitted to other units and processed in various (doi:10.2307/2585926)
ways. Extraneous information is discarded. Various Green, E. & Maegner, T. 1998 Red squirrels, Tamiasciurus
kinds of computation and inference are performed. hudsonicus, produce predator-class specific alarm calls.
The resulting information is used to guide group Anim. Behav. 55, 511–518. (doi:10.1006/anbe.1997.
decisions that lead to coordinated action. All this can 0620)
Grim, P., St. Denis, P. & Kokalis, T. 2002 Learning to
happen either with or without conscious thought.
communicate: the emergence of signaling in spatialized
These processes are instantiated in human organiz-
arrays of neural nets. Adap. Behav. 10, 45 –70. (doi:10.
ations, in the coordination of the organs and cells of a 1177/10597123020101003)
multicellular organism and even within the cells Gyger, M., Marler, P. & Pickert, R. 1987 Semantics of
themselves. Information flows through signalling net- an avian alarm call system: the male domestic fowl,
works at all levels of biological organization. Gallus domesticus. Behavior 102, 15 –20. (doi:10.1163/
156853986X00027)
I would like to thank Jeffrey Barrett, Simon Huttegger, Louis
Guckenheimer, J. & Holmes, P. 1983 Nonlinear oscillations,
Narens, Don Saari, Rory Smead, Elliott Wagner and Kevin
Zollman for many discussions. Rory Smead performed the dynamical systems, and bifurcations of vector fields.
‘Taking as Vote’ simulations reported in the electronic New York, NY: Springer.
supplementary material. I would also like to thank two Hadeler, K. P. 1981 Stable polymorphisms in a selection
anonymous referees who provided many helpful suggestions model with mutation. SIAM J. Appl. Math. 41, 1–7.
for the improvement of the manuscript. (doi:10.1137/0141001)
Hailman, J., Ficken, M. & Ficken, R. 1985 The ‘chick-a-dee’
calls of Parus atricapillus. Semiotica 56, 191–224.
Hofbauer, J. 1985 The selection–mutation equation. J. Math.
REFERENCES Biol. 23, 41–53.
Argiento, R., Pemantle, R., Skyrms, B. & Volkov, S. In press. Hofbauer, J. & Huttegger, S. M. 2008 Feasibility of
Learning to signal: analysis of a micro-level reinforcement communication in binary signaling games. J. Theor. Biol.
model. Stochastic Process. Appl. (doi:10.1016/j.spa.2008. 254, 843 –849. (doi:10.1016/j.jtbi.2008.07.010)
02.014) Hofbauer, J. & Sigmund, K. 1998 Evolutionary games and
Austen-Smith, D. & Banks, J. S. 1996 Information aggrega- population dynamics. Cambridge, MA: Cambridge
tion, rationality, and the condorcet jury theorem. Am. Pol. University Press.
Sci. Rev. 90, 34– 45. (doi:10.2307/2082796) Hopkins, E. & Posch, M. 2005 Attainability of boundary
Bala, V. & Goyal, S. 2000 A non-cooperative model of points under reinforcement learning. Games Econ. Behav.
network formation. Econometrica, 1181–1229. (doi:10. 53, 110–125. (doi:10.1016/j.geb.2004.08.002)
1111/1468-0262.00155) Huttegger, S. M. 2007a Evolution and the explanation of
Barrett, J. A. 2006 Numerical simulations of the Lewis meaning. Philos. Sci. 74, 1–27. (doi:10.1086/519477)
signaling game: learning strategies, pooling equilibria, and Huttegger, S. M. 2007b Evolutionary explanations of
the evolution of grammar. Working Paper MBS06-09. indicatives and imperatives. Erkenntnis 66, 409 – 436.
Irvine, UK: University of California. (doi:10.1007/s10670-006-9022-1)
Barrett, J. A. 2007a The evolution of coding in signaling Huttegger, S. M. 2007c Robustness in signaling games.
games. Theory Dec. (doi:10.1007/s11238-007-9064-0) Philos. Sci. 74, 839–847. (doi:10.1086/525625)
Barrett, J. A. 2007b Dynamic partitioning and the con- Huttegger, S. M., Skyrms, B., Smead, R. & Zollman, K.
ventionality of kinds. Philos. Sci. 74, 527–546. (doi:10. In press. Evolutionary dynamics of Lewis signaling
1086/524714) games: signaling systems vs. partial pooling. Working
Beggs, A. 2005 On the convergence of reinforcement Paper University of California, Irvine. Synthese.
learning. J. Econ. Theory 122, 1– 36. (doi:10.1016/j.jet. Kaiser, D. 2004 Signaling in myxobacteria. Annu. Rev.
2004.03.008) Microbiol. 58, 75 – 98. (doi:10.1146/annurev.micro.58.
Benaim, M., Shreiber, S. J. & Tarres, P. 2004 Generalized urn 030603.123620)
models of evolutionary processes. Ann. Appl. Probab. 14, Komarova, N. & Niyogi, P. 2004 Optimizing the mutual
1455 –1478. (doi:10.1214/105051604000000422) intelligibility of linguistic agents in a shared world. Artif.
Björnerstedt, J. & Weibull, J. 1995 Nash equilibrium and Intell. 154, 1– 42. (doi:10.1016/j.artint.2003.08.005)
evolution by imitation. In The rational foundations of Lewis, D. K. 1969 Convention. Cambridge, MA: Harvard
economic behavior (eds K. J. Arrow, E. Colombatto, University Press.
M. Perlman & C. Schmidt), pp. 155–171. New York, List, C., Elsholtz, C. & Seeley, T. D. 2009 Independence and
NY: MacMillan. interdependence in collective decision making: an agent-
Charrier, I. & Sturdy, C. B. 2005 Call-based species based model of nest-site choice by honeybee swarms. Phil.
recognition in the black-capped chickadees. Behav. Trans. R. Soc. B 364, 755–762. (doi:10.1098/rstb.2008.
Process. 70, 271–281. (doi:10.1016/j.beproc.2005.07.007) 0277)

Manser, M., Seyfarth, R. M. & Cheney, D. 2002 Suricate Skyrms, B. 2000 Evolution of inference. In Dynamics of human
alarm calls signal predator class and urgency. Trends Cogn. and primate societies (eds T. Kohler & G. Gumerman),
Sci. 6, 55 –57. (doi:10.1016/S1364-6613(00)01840-4) pp. 77–88. New York, NY: Oxford University Press.
Maynard Smith, J. 1982 Evolution and the theory of games. Skyrms, B. 2004 The stag hunt and the evolution of social
Cambridge, MA: Cambridge University Press. structure. Cambridge, MA: Cambridge University Press.
Maynard Smith, J. & Price, G. 1973 The logic of animal Skyrms, B. In press. Signals. Presidential Address of the
conflict. Nature 246, 15 –18. (doi:10.1038/246015a0) Philosophy of Science Association. Philos. Sci.
McGregor, P. 2005 Animal communication networks. Skyrms, B. & Pemantle, R. 2000 A dynamic model of social
Cambridge, MA: Cambridge University Press. network formation. Proc. Natl Acad. Sci. 97, 9340– 9346.
Pawlowitsch, C. 2007 Finite populations choose an optimal (doi:10.1073/pnas.97.16.9340)
language. J. Theor. Biol. 249, 606 –616. (doi:10.1016/j.jtbi. Sumpter, D. J. T. & Pratt, S. C. 2009 Quorum responses and
2007.08.009) consensus decision making. Phil. Trans. R. Soc. B 364,
Pawlowitsch, C. 2008 Why evolution does not always lead to 743–753. (doi:10.1098/rstb.2008.0204)
an optimal signaling system. Games Econ. Behav. 63, Taga, M. E. & Bassler, B. L. 2003 Chemical communication
among bacteria. Proc. Natl Acad. Sci. 100, 14 549 –14 554.
203 –226. (doi:10.1016/j.geb.2007.08.009)
(doi:10.1073/pnas.1934514100)
Schauder, S. & Bassler, B. 2001 The languages of
Taylor, C., Fudenberg, D., Sasaki, A. & Nowak, M. 2004
bacteria. Genes Dev. 15, 1468 –1480. (doi:10.1101/
Evolutionary game dynamics in finite populations.
gad.899601)
Bull. Math. Biol. 66, 1621–1644. (doi:10.1016/j.bulm.
Seyfarth, R. M. & Cheney, D. L. 1990 The assessment by 2004.03.004)
vervet monkeys of their own and other species’ alarm calls. Taylor, P. & Jonker, L. 1978 Evolutionarily stable strategies
Anim. Behav. 40, 754–764. (doi:10.1016/S0003-3472(05) and game dynamics. Math. Biosci. 40, 145 –156. (doi:10.
80704-3) 1016/0025-5564(78)90077-9)
Schlag, K. 1998 Why imitate and if so, how? A bounded Wärneryd, K. 1993 Cheap talk, coordination, and evolutionary
rational approach to many armed bandits. J. Econ. Theory stability. Games Econ. Behav. 5, 532–546. (doi:10.1006/
78, 130–156. (doi:10.1006/jeth.1997.2347) game.1993.1030)
Shreiber, S. 2001 Urn models, replicator processes and Zahavi, A. 1975 Mate selection—selection for a handicap.
random genetic drift. SIAM J. Appl. Math. 61, J. Theor. Biol. 53, 205 –214. (doi:10.1016/0022-5193
2148 –2167. (doi:10.1137/S0036139999352857) (75)90111-3)
Skyrms, B. 1999 Stability and explanatory significance of Zollman, K. 2005 Talking to neighbors: the evolution of
some simple evolutionary models. Philos. Sci. 67, 94–113. regional meaning. Philos. Sci. 72, 69 –85. (doi:10.1086/
(doi:10.1086/392763) 428390)

Phil. Trans. R. Soc. B (2009) 364, 781–789

doi:10.1098/rstb.2008.0233
Leadership, consensus decision making and

collective behaviour in humans
John R. G. Dyer1,*, Anders Johansson2, Dirk Helbing2, Iain D. Couzin3
and Jens Krause1
1
Institute of Integrative and Comparative Biology, University of Leeds, Leeds LS2 9JT, UK
2
Department of Humanities, Social and Political Sciences, ETH Zurich, UNO D 11,
Universitätstrasse 41, 8092 Zurich, Switzerland
3
Department of Ecology and Evolutionary Biology, Princeton University, 106a Guyot Hall,
Princeton University, Princeton, NJ 08544-1003, USA
This paper reviews the literature on leadership in vertebrate groups, including recent work on human
groups, before presenting the results of three new experiments looking at leadership and decision
making in small and large human groups. In experiment 1, we find that both group size and the
presence of uninformed individuals can affect the speed with which small human groups (eight
people) decide between two opposing directional preferences and the likelihood of the group
splitting. In experiment 2, we show that the spatial positioning of informed individuals within small
human groups (10 people) can affect the speed and accuracy of group motion. We find that having a
mixture of leaders positioned in the centre and on the edge of a group increases the speed and
accuracy with which the group reaches their target. In experiment 3, we use large human crowds
(100 and 200 people) to demonstrate that the trends observed from earlier work using small human
groups can be applied to larger crowds. We find that only a small minority of informed individuals is
needed to guide a large uninformed group. These studies build upon important theoretical and
empirical work on leadership and decision making in animal groups.
Keywords: leadership; consensus decision making; collective behaviour; human group
1. INTRODUCTION inclined to lead. In the words of William Shakespeare

This paper will begin by reviewing the literature on in Twelfth Night: ‘some men are born great, some
leadership in vertebrate groups, including recent men achieve greatness and some have greatness thrust
empirical work on human groups. We will then present upon them’.
the results of three new human crowd experiments that Evidence of leadership behaviour has been found in
build upon the work of Dyer et al. (2008). a number of vertebrate species across a range of taxa,
Consensus decisions are defined by Conradt & including ungulates (Leicester sheep, Ovis aries;
Roper (2005) as ‘when the members of a group choose Squires & Daws 1975; sable antelope, Hippotragus
between two or more mutually exclusive actions with niger: Stine et al. 1982; Charolais heifers, Bos taurus;
the aim of reaching a consensus’. They are very Dumont et al. 2005; zebras, Equus burchellii: Fischhoff
important for both animal and human groups as they et al. 2007), primates (gorillas, Gorilla gorilla beringei:
allow groups to remain together despite individual Fossey 1972; chimpanzees Pan troglodytes: Wilson
differences in preference and consequently help 1980), canids (bush dogs Speothos venaticus: Macdonald
prevent individuals from losing the benefits associated 1996; wolves Canis lupus: Peterson et al. 2002), birds
with being part of a large group (Conradt & Roper (bar headed geese Anser indicus: Lamprecht 1992; zebra
2009; Sumpter & Pratt 2009). Decision making almost finches Taeniopygia guttata: Beauchamp 2000; homing
always involves some form of leadership. Here, we pigeons Columba livia domestica: Biro et al. 2006) and
define leadership as ‘the initiation of new directions of fishes (roach Rutilus rutilus: Bumann et al. 1997; golden
locomotion by one or more individuals, which are then shiners Notemigonus crysoleucas: Reebs 2000, 2001). In
readily followed by other group members’ (Krause some species, it has been shown that groups tend to be led
et al. 2000). Leadership may either be designated or by dominant individuals (gorillas G. g. beringei: Fossey
emerge spontaneously due to individuals possessing 1972; Leicester sheep, Squires & Daws 1975; chimpan-
qualities or experience in certain situations, or because zees P. troglodytes: Wilson 1980; sable antelope H. niger:
they are of a personality type that is generally more Stine et al. 1982; bush dogs S. venaticus: Macdonald
1996; wolves C. lupus: Peterson et al. 2002). In other
* Author for correspondence (bgyjdd@leeds.ac.uk). species, it has been demonstrated that leadership can be
Electronic supplementary material is available at http://dx.doi.org/10. more variable and that there is no correlation with
1098/rstb.2008.0233 or via http://journals.royalsociety.org. dominance (bar-headed geese A. indicus: Lamprecht
One contribution of 11 to a Theme Issue ‘Group decision making in 1992; zebra finches T. guttata: Beauchamp 2000; white-
humans and animals’. faced capuchins Cebus capucinus: Leca et al. 2003).

782 J. R. G. Dyer et al. Leadership in humans
Research on fish shoals has provided insight into than 1208) then the group decides in favour of one set of
leadership in the absence of dominance hierarchies and informed individuals. This prediction has received
complex signalling between individuals. Reader et al. empirical support from work on pairs of homing pigeons,
(2003) demonstrated that four demonstrator guppies, C. l. domestica (Biro et al. 2006).
Poecilia reticulata, could guide four uninformed con- Dyer et al. (2008) tested some of the predictions of
specifics through a hole to escape an oncoming trawl the Couzin et al. (2005) model using small human
net. Similarly, Reebs (2000, 2001) found that a small groups. Similar to the model of Couzin et al.,
minority of informed fish (golden shiners N. crysoleucas) individuals had no information about which other
could guide uninformed conspecifics from a preferred individuals had information, and participants were not
area of a tank, to a less preferred brightly lit area where allowed to talk or gesture in order to minimize
food was expected. Reebs (2001) also found an influence information exchange through active signalling. All
of body size on leadership with large knowledgeable individuals were instructed that they must remain
fish being readily followed by small uninformed fish together as a group. The groups started in the centre of
whereas small knowledgeable fish were less readily a circular arena, and instead of a preferred direction of
followed by larger individuals. Krause et al. (1998) travel informed individuals were instructed to reach a
found that front positions tended to be occupied by target (a number between 1 and 16) on the edge of the
larger fish and food-deprived fish. Bumann et al. (1997) 10 m circle. This enabled Dyer et al. to measure not
showed that individuals in front positions in shoals of only the accuracy of group motion but also the time
roach R. rutilus led the group, finding that just a single taken to reach their intended target. Using mixed sex
individual in a front position could have a strong groups of eight people, they found that just one
influence on an entire shoal of 16 fish. informed individual (12.5% of the group informed)
Couzin et al. (2005) used a simple individual-based could guide the group with great accuracy. They found
model to look at the mechanisms of leadership and the effect on time to be less immediate with two
decision making in moving animal groups, in the informed individuals (25% of the group informed)
absence of complex signalling and in situations where being required to bring about a significant decrease in
it is not possible for individuals to establish which other the time taken to reach the circle periphery. Interest-
individuals, if any, have information. First, they ask ingly, Dyer et al. (2008) found no evidence of a trade-
how small numbers of individuals with information on off between the speed and accuracy of decision making
a migration route or the location of resources can as has been found from previous work on humans
influence the rest of the group; and second, how groups (Edwards 1965; Vitevitch 2002), ants (Franks et al.
can overcome conflicts in individual preferences in 2002, 2003, 2009), monkeys (Roitman & Shadlen
order to achieve consensus. These questions are 2002) and bees (Chittka et al. 2003). Dyer et al. (2008)
particularly relevant to an understanding of the also tested scenarios where a conflict (different
mechanisms of leadership and decision making in numbers of informed individuals were instructed to
large animal groups, such as insect swarms or fish reach targets 1808 apart) was introduced in the
schools, where individuals may not have the capacity preferences of informed individuals, finding that the
for individual recognition and where crowding may direction of group motion was almost always decided
limit the range over which individuals can see each by the majority. These results provide good initial
other. The model is a simple individual-based model in empirical support for the predictions of Couzin et al.
which individuals attempt to maintain personal space (2005) but leave a number of questions unanswered:
by avoiding other individuals within a certain region. first, it is not clear what the role is of uninformed
If no neighbours are detected within this region then individuals in reaching consensus decisions. Second,
the individual will become attracted towards and how the spatial positioning of informed individuals
aligned with neighbours within a local interaction affects the speed and accuracy with which they guide
range in order to maintain cohesion with neighbours. uninformed group members to a target. Third, whether
The model assumes that all individuals move at the the results found by Dyer et al. (2008) using small
same speed and are identical except that a certain groups are also applicable to large groups (‘crowds’).
proportion is given a directional preference (unit To answer these questions, we present the results of
vector) representing, for example, the direction to a three experiments using human groups that build upon
known resource, whereas all other individuals have no the work of Dyer et al. (2008). Whenever we refer to
preferred direction of travel. Couzin et al. (2005) groups in this paper we use the flexible definition of
predicted that a small proportion of informed individ- Forsyth (1999) who defines a group as ‘two or more
uals (approx. 5% of group members) can accurately interdependent individuals who influence each other
guide an uninformed group and that for any given through social interaction’. This definition does not
group size the accuracy of group motion increases as imply permanence, structure or psychological meaning
the proportion of informed individuals is increased. for members. When referring to crowds, we simply
Furthermore, they predict that where there are mean large groups of 100 individuals and over.
conflicts in the preferences of informed individuals
and the number of individuals with each preference is
unequal, the group will always go with the majority of 2. MATERIAL AND METHODS
informed individuals. When the number with each (a) Experiment 1: the importance of uninformed
preference is equal, the group averages over the individuals in reaching consensus
preferences if the differences are small (less than 1208). This experiment took place between February and March
However, when individual differences are large (more 2006 at the University of Leeds (England) and the University

Leadership in humans J. R. G. Dyer et al. 783
1 Each group was given the following standard set of

16 2 instructions: ‘when we tell you to begin you should start
15 3 walking at a normal speed and do not stop before being told to
do so. You can walk anywhere inside or outside the circle but
14 A 4 you have to stay within an arm’s length of another individual
H B and you should not talk or gesture to each other.’ Both walking
13 G C 2m 5 10 m speed and the distance they should keep from each other were
F D demonstrated to them before the experiment. These instruc-
E 6 tions attempted to make the participants as comparable as
12
possible to the agents in the Couzin et al. (2005) model in
7 which they move at the same standard speed, they can move
11
8 anywhere and are attracted to each other within a certain zone.
10 9 The instruction not to talk and gesture attempted to minimize
active information transfer between individuals. We found
25 (a) ( J. Krause & J. R. G. Dyer 2005, personal observation) that in
median time to perimeter (s)
real-life situations the conditions apply remarkably often

20 because strangers getting off planes, interacting in pedestrian
zones, entering or leaving buildings regularly do interact
15 * *** without talking to each other and without obvious gestures. In
fact, this seems to be the norm in many countries.
10 In addition to these standard instructions, participants
were each handed a slip of paper with an additional individual
5
behavioural rule to follow. They were instructed to read and
memorize the information, then hide the slip to ensure that
0
no other member of the group could see it. The slips of paper
20 (b) gave one of two different behavioural rules, one for
*** uninformed individuals and one for informed individuals.
no. of groups that split
Behavioural rule 1 gave instructions to simply ‘stay with the

15 ***
group’, resulting in uninformed individuals. Behavioural rule
2 gave instructions to ‘Go to number X, without leaving the
10 group’ creating informed individuals (X represents a randomly
chosen number on the outer circle between 1 and 16).
This rule creates a scenario that is similar to the model of
5
Couzin et al. (2005) in which although individuals have a
preferred direction, they are still attracted to other individuals
0 and so must balance social attraction against individual
2vs2 naive present 4vs4 directional preference.
treatment Each group was tested in three different treatments. In
each treatment, the informed individuals were given one
Figure 1. Experiment 1. (a) Median (G quartiles) time taken
of two separate targets, 1808 apart. In the first treatment
to reach the periphery of the circle by groups tested in the
(the ‘2 versus 2’ treatment), two individuals were each given a
three different experimental treatments. (b) Number of
groups that split up at least once during their trial in each target and no uninformed individuals were present (group
of the three different experimental treatments. Treatment sizeZ4). In the second treatment (the ‘uninformed present’
differences are indicated by FDR corrected pairwise treatment), two individuals were each given a target and four
comparisons. 0.001, 0.01 and 0.05. (Inset) Overhead uninformed individuals were also present (group sizeZ8).
view of the arena used in experiment 1. Letters represent In the third treatment (the ‘4 versus 4’ treatment), four
starting positions for participants and numbers were used to individuals were each given a target and no uninformed
orientate them and as targets. individuals were present (group sizeZ8). These three
different treatments allowed us to look at the effect of the
of Wales at Bangor. Participants were undergraduate presence and the absence of uninformed individuals both
students. In total, 22 mixed-sex groups of eight individuals when group size remains constant (by comparing the 4 versus
were used for testing. All experiments were carried out 4 with the uninformed present treatment) and when number
double-blind in that both the participants and the individuals of informed individuals remains constant (by comparing the
who measured the response variables were not aware of the 2 versus 2 with the uninformed present treatment).
purpose of the experiment. Four informed individuals were randomly assigned and
A circular arena with a 10 m diameter was marked on the were used as the informed individuals in both the 2 versus 2
floor and cards labelled 1–16 were spaced equally around its and the uninformed present treatments. Treatment order was
perimeter. A circle with a diameter of 2 m was marked out in the systematically rotated to minimize its effect on the results.
centre of the first circle with the letters A–H spaced equally This meant that on some occasions (e.g. if the 2 versus 2 or
around its perimeter (figure 1 (inset)). Individuals were asked the uninformed present treatments followed the 4 versus 4
to stand on a letter (A–H) on the inner circle to ensure that treatment) some individuals who were informed in a previous
all starting positions were equal and equidistant from the outer trial would then be uninformed in a subsequent trial.
periphery. To avoid any bias due to initial direction of Previously informed individuals have been shown by Dyer
locomotion, the initial orientation of each individual in a trial et al. (2008) not to affect the results of subsequent trials, but
was randomized by instructing them to face a number from the we also test for their effects here. During the 2 versus 2 trials,
outer circle chosen at random without replacement. the four individuals who were not assigned as informed

individuals and therefore not involved in the trial were taken 1

to the side of the arena. This may have given them a chance to 16 2
observe the trial, which could potentially have given them 15 3
clues as to the nature of the experiment. Consequently, we
tested for any potential effects of observing a previous trial by 14 A 4
comparing the performance of groups in each of the other H B
I
treatments separately, on occasions when they preceded and 13 G C 2m 5 10 m
followed the 2 versus 2 treatment. There was no significant J
F D
difference in time taken to reach the circle periphery E 6
12
when comparing groups within the uninformed present and
4 versus 4 treatments on occasions where these treatments 7
11
preceded and succeeded the 2 versus 2 treatment (uninormed
10 8
present treatment: Mann–Whitney U-test: zZK1.284, nZ11, 9
11, pZ0.210; 4 versus 4 treatment: Mann–Whitney U-test:
zZK1.320, nZ11, 11, pZ0.197). This suggests that there
median time to perimeter (s)

was no effect on subsequent trials of the four individuals 50 (a) *
watching the 2 versus 2 treatment from the side.
*
After we signalled the start of a trial, it lasted until any 40
member of the group came within 50 cm of the perimeter of
the circle. This was judged by two observers on either side 35
of the arena who were blind to the purpose of the experiment.
20
The observers recorded the time taken by the group to come
within 50 cm of the periphery and the target which they 10
finished closest to or the two targets if they finished in
between the two targets. 0
median deviation from target
(b) Experiment 2: spatial position of informed (b)

4
individuals
This experiment took place between January 2006 and March *
2007 at the University of Leeds (England) and the University 3 *
of Wales at Bangor. Participants were undergraduate
students. In total, 15 mixed-sex groups of ten individuals 2 *
were used for testing. All experiments were carried out
double-blind in that both the participants and the individuals 1
who measured the response variables were not aware of the
purpose of the experiment. 0
A circular arena was marked out in the same way as in mixed close far 2 core
experiment 1 except that this time the letters I and J were also treatment
placed in the centre of the inner circle of letters (figure 2
(inset)). Individuals were asked to stand on a letter (A–J ) on Figure 2. Experiment 2. (a) Median (G quartiles) time taken
the inner circle. The rest of the protocol was exactly the same to reach the periphery of the circle by groups tested in the four
as in experiment 1 except that this time each group of 10 was different experimental treatments. (b) Median (G quartiles)
tested in four different treatments that differed only in the deviation of groups from their target under the four different
experimental treatments (filled circles represent leaders’
starting positions of the two informed individuals with the
starting positions and empty circles represent uninformed
same target direction (see figure 2 and inset). In the first
individuals’ starting positions). Treatment differences are
treatment (‘mixed treatment’), one of the informed individ-
indicated by FDR corrected pairwise comparisons. 0.001,
uals started in a core position (position J) and one started on
0.01 and 0.05. (Inset) Overhead view of the arena used in
the periphery (position E). In the second treatment (‘close experiment 2. Letters represent starting positions for
treatment’), the two informed individuals started close participants and numbers were used to orientate them and
together both on the periphery (positions C and D). In the as targets.
third treatment (‘far treatment’), the two informed individ-
uals started far apart at opposite sides of the periphery of the in Cologne (Germany) and 5 May 2007 in Freiburg
group (positions B and F). In the final treatment (‘2 core (Germany). Participants were volunteers between the age of
treatment’), both leaders started in core positions within the 18 and 70 of both sexes who had answered TV or radio
group (positions I and J). Treatment orders were system- advertisements asking for participants for a swarm experi-
atically rotated to minimize any order effects. The two ment (no further information on the nature of the experiment
informed individuals were randomly assigned for the first was given until the experiment was finished).
treatment and these same individuals were the informed A circular arena with a 50 m diameter was marked on the
individuals in each of the other treatments. floor, and large mounted wooden boards raised approxi-
mately 2 m above the ground and printed with the numbers
(c) Experiment 3: leadership and decision making in 1–12 were spaced equally round its perimeter (as a clock
large human crowds face). Two more circles were marked out in the centre of the
This experiment will yield mainly anecdotal evidence as it is first circle with diameters of 12 and 32 m (figure 3). The
based on a small sample size consisting of, in part, a single smallest circle in the centre (12 m in diameter) represented
group of 200 people, and in other parts, an additional group the starting area for the group of participants with the middle
of 100 people. The experiments took place on 4 March 2007 circle (32 m in diameter) acting as a guide for us to observe

12 (d) Statistical analyses

11 All statistical analyses were carried out using R v. 2.5.1. All
1
data failed to meet the preconditions required for parametric
testing even after transformation. Therefore, generalized
10 2 linear mixed models (GLMMs) were used to analyse the
effects of the different treatments and the order the groups
200 experienced the treatments, and their potential two-way
32 m 9 12 m people 3 50 m interaction on the response variables of time to circle
periphery and the likelihood of a group splitting in
experiment 1, and time to circle periphery and deviation
8 4 from intended target in experiment 2. Group ID was entered
as the random factor in each model due to the repeated-
7 5 measures design of the study (each group was tested in each
6 treatment). In all cases, time and deviation data were
overdispersed and quasi-Poisson error distributions were
Figure 3. Overhead view of the arena used in experiment 3. found to be the best fit to the data. In the case of the splitting
The numbers were used to orientate individuals and data, a binomial error distribution was used. In no cases were
as targets. the response variables affected by treatment order and so this
variable was removed from the model and further analysis.
Where a significant effect of treatment was found from a
the distance moved by the group. Controls were run with no GLMM, pairwise comparisons were made between the
circles on the floor, which confirmed that the same collective treatments by correcting the alpha level using an FDR
behaviours were found in the presence or the absence of these correction (see Benjamini & Hochberg 1995). All p-values
floor markings. from pairwise tests throughout the paper were ranked and the
Each participant was given one of 10 different coloured lowest p-value was compared with an alpha level of 0.00278
caps (20 individuals with each colour). This was done in (0.05/no. of pairwise comparisons). The next lowest p-value
order to facilitate the organization of such large numbers of was then compared with 0.05/(no. of comparisonsK1) and so
people and also to later identify informed individuals in the on until the highest p value is compared with 0.05.
video recordings. Each individual was handed a slip of paper,
which they were instructed not to open before being told to do
so. On the left-hand side of each slip of paper was an 3. RESULTS
instruction that read ‘before we say ‘GO’ you should face (a) Experiment 1: the importance of uninformed
number X’ (with X representing one of the 12 numbers on the individuals in reaching consensus
outer circle). On the right-hand side was a behavioural rule (i) Time to periphery
for them to follow. Uninformed individuals were told to ‘Stay The time taken to reach the periphery differed signi-
with the group’. Informed individuals (leaders) were told to ficantly between treatments (GLMM PQL: F2,42Z
‘Go to 9 o’clock, but do not leave the group’. 12.417, p!0.0001). Groups with uninformed individ-
The participants were then read the same set of uals took significantly longer to reach the periphery than
standardized instructions as in experiments 1 and 2, and groups in the 2 versus 2 or 4 versus 4 treatments, but
normal walking speed and staying together as a group were there was no significant difference between groups in the
demonstrated to them. The participants of each hat colour 2 versus 2 and 4 versus 4 treatments (figure 1a).
group in turn were then asked to spread themselves out in the
smallest circle in the centre, starting with the hat colour that
(ii) Deviation from target
contained the informed individuals. This ensured that
Deviation was measured as how many targets away
whatever colour hats the informed individuals were wearing
from the closest intended target of the informed
they were spread out among the group.
individuals the group finished (e.g. if one set of
A single group of 200 people was tested in five different
informed individuals were given target 1 and the others
treatments. In the first treatment (20 leader treatment), 20
were given target 9 and the group finished at target 7,
individuals received the rule for informed individuals. In the
second treatment (five leader treatment), five individuals
then the deviation is 2). Most groups were highly
received the rule for informed individuals. In the third
accurate, finishing at their targets and therefore scoring
treatment (10 leader treatment), 10 individuals received the
no deviation (19/22 groups in the 2 versus 2 treatment,
rule for informed individuals. In the fourth treatment 20/22 groups in the uninformed treatment and 21/22
(control treatment), all individuals received the rule for unin- groups in the 4 versus 4 treatment). Consequently,
formed individuals. In the fifth treatment (20 versus 10 there was no significant difference between the
conflict treatment), we introduced conflict so that 20 treatments in deviation from target (GLMM PQL:
individuals were instructed to go to one target and 10 were F2,42Z0.0583, pZ0.944).
instructed to go to a target opposite (1808 away) from this.
After each treatment, all individuals wearing the colour of hat (iii) Group splits
worn by the informed individuals were removed from the The likelihood of a group splitting was significantly
group and replaced by 20 more participants wearing the same affected by treatment (GLMM PQL: F2,42Z15.604,
coloured hats. This was done in order that no individuals who p!0.0001). Groups were significantly more likely to
were previously informed would be uninformed individuals split in the 4 versus 4 treatment than in either the
in a subsequent trial. The uninformed treatment and the 10 uninformed or 2 versus 2 treatment, but there was no
leader treatment were repeated with a different group of difference between the uninformed and 2 versus 2
100 participants. treatments (figure 1b). Where group splits were

Table 1. Proportion of informed individuals needed to guide an uninformed group.
proportion of indi- proportion of uniformed individuals

viduals that are time until first subgroup/unsplit that reach target in first subgroup/
group size informed (%) group split? group reaches the target (s) unsplit group (%)
200 2.50 yes 222 5

200 5 yes 250 89
200 10 no 75 100
100 10 no 103 100
observed, we recorded the size of the separate groups. informed individuals. After approximately 110 s the 10
In the 2 versus 2 treatment, all five group splits (100%) informed individuals had also reached their target and
were into two separate groups of two individuals. In the managed to take at least 40 per cent of the group with
uninformed present treatment, five out of seven group them. A bridge of people remained between the two
splits (71%) were into one group of six and another targets with a constant oscillation of people between the
group of two individuals. In the 4 versus 4 treatment, two targets (see appendix 2 in the electronic supple-
9 out of 13 splits (69%) were into two separate groups of mentary material). The experiment was terminated
four individuals. after the oscillation had continued for a further 4 min.
(b) Experiment 2: spatial position of informed (ii) Controls: no informed individuals

individuals When there were no informed individuals, the group
(i) Time to periphery formed a torus with multiple lanes of people moving in
The time taken to reach the periphery differed opposite directions (see appendix 3 in the electronic
significantly between treatments (GLMM PQL: supplementary material). The torus formed after
F3,42Z3.712, pZ0.0186). Groups with one informed approximately 30 s and ranged between 14 and 17 m
individual starting in the core and one on the group in diameter (figure 4a). The torus was not stationary
periphery reached the perimeter in significantly less and moved position within the arena (figure 4b). The
time than groups with two core leaders and groups with same collective behaviour also occurred with a group
two leaders on opposite sides of the edge. There were size of 100 people without informed individuals.
no other significant differences between treatments in
time to periphery (figure 2a).
4. DISCUSSION
(ii) Deviation from target Experiment 1 demonstrates the importance of unin-
Deviation from target differed significantly between formed individuals in the process of reaching consensus
treatments (GLMM PQL: F3,42Z3.798, pZ0.0170). movement decisions. Interestingly, when we compare
Groups with informed individuals in core and periph- the 2 versus 2 treatment with the 4 versus 4 treatment
eral positions deviated from their targets signific- we find that increasing group size per se does not
antly less than groups in all other treatments. There increase the time taken to reach the periphery. Only the
were no other significant differences between the presence of uninformed individuals is associated with
treatments (figure 2b). an increase in this time. Our results for the number of
groups that split show that groups in the 4 versus 4
(c) Experiment 3: consensus decision making in treatment split significantly more frequently than those
large human crowds: proportion of leaders in the 2 versus 2 and the uninformed present
The results presented in this section are largely treatments. This suggests that by increasing group
anecdotal as they are based on a small sample size size we also increase the likelihood of group fragmenta-
(one group of 200 people and for some treatments a tion, but only if the additional group members are
further group of 100 people) due to the logistical informed individuals (significantly more groups split in
difficulties in testing such large groups of people. the 4 versus 4 than the 2 versus 2 treatment). When the
At least 5 per cent of group members had to be additional group members were uninformed, groups
informed in order to lead the group with reasonable were no more likely to split (no significant difference
effectiveness (90% of the group) to the target (table 1). between the 2 versus 2 and the uninformed present
If 10 per cent of the members were informed, the whole treatments). It is likely that the increased splitting in the
group reached the target without a split (see appendix 1 4 versus 4 treatment occurs due to a ‘strength in
in the electronic supplementary material). numbers’ effect, whereby the two sets of informed
individuals can split and still feel that they have stayed
(i) Conflict: 20 versus 10 informed with the group. This is backed up by the fact that 9 out
Within 60 s, approximately half of the group were of 13 (69%) of group splits in the 4 versus 4 treatment
together at the target of the 20 informed individuals, were into two separate groups of four individuals.
while the other half were still fairly close to the centre of To our knowledge, few empirical studies on
the arena joined to the other group by a bridge of vertebrate groups outside the social science literature
people being exchanged between the groups. The on human groups have looked at decision making in
group then proceeded to become increasingly stretched conflict situations. Dyer et al. (2008) have found
out across the arena between the targets of the different support for the model of Couzin et al. (2005) showing

18 howler monkeys, capuchins, elephants and swans,

(a) which suggest that conflicts about timings and move-
17 ment directions are resolved by majority decisions.
However, these are largely anecdotal.
group diameter (m)
16
In experiment 2, we find that the spatial starting
position of informed individuals affects both the speed
and the accuracy with which they can guide an
15
uninformed group to a target. We find that having
one informed individual starting in the centre and one
14 starting on the periphery of the group is the most
effective way of guiding the group quickly and
13 accurately to a target. Our results are in line with the
model of Aube & Shield (2004), which predicted that
12 having leaders positioned in a mix of places (centre,
0 50 100 150 200 250
peripheral and distant) meant that more people could
time (s)
be saved in a shorter time from a simulated evacuation
2 scenario. Although there are obviously clear differences
(b)
between our study and real evacuation situations where
0 there could be widespread panic and more erratic
behaviour, our results may offer some insight into
considerations on the best places to position officials/
distance, y (m)
–2 marshals in order to evacuate people most efficiently.

It is likely that the mixed treatment is most effective
–4 due to the benefits of having the two different types of
leader. The leader on the periphery is likely to be more
mobile and unconstrained and can move freely around
–6 the outside of the group and quickly find and align with
the target, while the other leader in the core position,
–8
although being initially more constrained and sur-
–4 –3 –2 –1 0 1 2 3 rounded by people, may be able to influence more
distance, x (m) uninformed individuals through his/her movements
towards the target. Beckman et al. (2006) found
Figure 4. Experiment 3: no informed individuals. (a) Time
evidence that informed scout bees guide largely
evolution of the central point (average position) of the group.
At time tZ0, the marker is black and the marker becomes uninformed swarms to a new nest site by flying through
increasingly light for later times. (b) Diameter of the group, as the swarm indicating the direction of travel. Leca et al.
a function of the time. Calculated as the mean distance to the (2003) found that white-faced capuchin monkeys
centre for all individuals, multiplied by p/2. starting from core positions were more likely to initiate
group movements than those on the edge of a group.
that where differences in preference are large and where Our results suggest possible navigational benefits to
there is an imbalance in the number of individuals with animal groups from informed individuals being spread
each directional preference, human groups tend to out through the group. For example, in migrating
choose the direction preferred by the majority. groups of birds, we may expect more accurate
However, in contrast to some recent studies (e.g. navigation of the route if the experienced older
Ward et al. 2008; Franks et al. 2009; Sumpter & Pratt individuals, who have already completed the migration
2009), Dyer (2008) did not find evidence that in past years, are spread out through the flock.
consensus decisions followed a quorum decision rule. Unfortunately, very little work has been carried out
Kerth et al. (2006) found that Bechstein’s bats, Myotis on the extent to which younger individuals use the
bechsteinii, can make group roost decisions that follow a experience of older individuals in bird flocks or on the
majority rule. They also found that the temporary relative positioning of adult and juvenile birds (Alerstam
splitting of groups could allow individuals to avoid 1990; Berthold 1993; Maransky & Bildstein 2001).
following majority decisions that did not favour them. One such study by Maransky & Bildstein (2001) found
Biro et al. (2006) provided further support for Couzin that in mixed-age flocks of broad-winged hawks, Buteo
et al. (2005); finding that when differences between the platypterus, adults were more likely (but not always)
directional preferences of two homing pigeons were the lead bird and were more likely to be (but were not
small, they would average over these preferences, but always) in the lead half of the flock.
when differences were large, one of the birds would In experiment 3, we found anecdotal evidence that
become leader. Here, we look more specifically at the the results of Dyer et al. (2008) on small human crowds
role of uninformed individuals and demonstrate that (eight individuals) can be scaled up to large human
both the presence of uninformed individuals and group crowds (100 or 200 individuals). First, we showed that
size can affect the speed with which a group decides a small informed minority (5%) could effectively guide
between two opposing directional preferences and the a large uninformed group to a target. This is in close
likelihood of group fission. There are also several agreement with theoretical results by Couzin et al.
further studies (briefly reviewed in Conradt & Roper (2005). Second, when there was a conflict in the
2003; table 1) on buffalo, red deer, gorillas, baboons, information given to different informed individuals, the

majority of the group initially went towards the target of Couzin, I. D., Krause, J., Franks, N. R. & Levin, S. A. 2005
the majority. However, the arena was not large enough Effective leadership and decision-making in animal groups
to decide whether the group would have reached the on the move. Nature 433, 513–516. (doi:10.1038/
majority preferred target cohesively, or split. Third, nature03236)
Dumont, B., Boissy, A., Achard, C., Sibbald, A. M. &
when no directional information was given to any
Erhard, H. W. 2005 Consistency of animal order in
members of the group, we observed the formation of a spontaneous group movements allows the measurement of
torus as seen with smaller groups (Dyer et al. 2008). leadership in a group of grazing heifers. Appl. Anim. Behav.
Our work indicates that this collective behaviour arises Sci. 95, 55–66. (doi:10.1016/j.applanim.2005.04.005)
when people are in continuous motion without any Dyer, J. R. G. 2008 Leadership, decision making and
strong directional cues. This potentially sheds interest- collective behaviour in animal groups. PhD thesis,
ing light on torus formation in animal groups where University of Leeds, UK.
the behaviour is frequently found in pelagic fish Dyer, J. R. G., Ioannou, C. C., Morrell, L. J., Croft, D. P.,
species such as barracuda, Sphyraena barracuda and Couzin, I. D., Waters, D. A. & Krause, J. 2008 Consensus
jack, Trachurus symmetricus, and has also been decision making in human crowds. Anim. Behav. 75,
46–470. (doi:10.1016/j.anbehav.2007.05.010)
described for wrinkle-lipped bats, Chaerephon plicata
Edwards, W. 1965 Optimal strategies for seeking infor-
(Siemers & Nill 2001).
mation: models for statistics, choice reaction times, and
Thanks are due to Darren Croft, Dean Waters, Hannah human information processing. J. Math. Psychol. 2,
Sargeant, Christos Ioannou, Jolyon Faria, Ben Chapman, 312–329. (doi:10.1016/0022-2496(65)90007-6)
Phil Thomas, John Ward, Dawn Thomas and Lesley Morrell Fischhoff, I. R., Sundaresan, S. R., Cordingley, J., Larkin,
who helped with carrying out the experimental work. We H. M., Sellier, M. & Rubenstein, D. I. 2007 Social
would also like to thank Ismeni Walter, Wolfgang Lemme, relationships and reproductive state influence leadership
Ranga Yogeshwar and the whole Quarks’ Team from the roles in movements of plains zebra, Equus burchellii. Anim.
WDR Fernsehen and SWR Fernsehen (who made the large Behav. 73, 825–831. (doi:10.1016/j.anbehav.2006.10.012)
crowd experiments possible) and all the volunteers who Forsyth, D. 1999 Group dynamics, 3rd edn. Belmont, CA:
participated in the experiments. Financial support was Wadsworth.
provided by the EPSRC and the NERC to J.K. Fossey, D. 1972 Vocalizations of the mountain gorilla (Gorilla
gorilla beringei ). Anim. Behav. 20, 36–53. (doi:10.1016/
S0003-3472(72)80171-4)
Franks, N. R., Pratt, S. C., Mallon, E. B., Britton, N. F. &
REFERENCES Sumpter, D. J. T. 2002 Information flow, opinion polling
Alerstam, T. 1990 Bird migration. Cambridge, UK: and collective intelligence in househunting social insects.
Cambridge University Press. Phil. Trans. R. Soc. B 357, 1567–1583. (doi:10.1098/rstb.
Aubé, F. & Shield, R. 2004 Modeling the effect of leadership 2002.1066)
on crowd flow dynamics. Lect. Notes Comput. Sci. 3305, Franks, N. R., Dornhaus, A., Fitzsimmons, J. P. & Stevens,
601–611. M. 2003 Speed versus accuracy in collective decision
Beauchamp, G. 2000 Individual differences in activity and making. Proc. R. Soc. B 270, 2457–2463. (doi:10.1098/
exploration influence leadership in pairs of foraging rspb.2003.2527)
zebra finches. Behaviour 137, 301–314. (doi:10.1163/156 Franks, N. R., Dechaume-Moncharmont, F.-X., Hanmore,
853900502097) E. & Reynolds, J. K. 2009 Speed versus accuracy in
Beckman, M., Fathke, R. L. & Seeley, T. D. 2006 How does decision-making ants: expediting politics and policy
an informed minority of scouts guide a honeybee swarm as implementation. Phil. Trans. R. Soc. B 364, 845–852.
it flies to its new home? Anim. Behav. 71, 161–171. (doi:10.1098/rstb.2008.0224)
(doi:10.1016/j.anbehav.2005.04.009) Kerth, G., Ebert, C. & Schmidtke, C. 2006 Group decision
Benjamini, Y. & Hochberg, Y. 1995 Controlling the false making in fission–fusion societies: evidence from two field
discovery rate—a practical and powerful approach to experiments in Bechstein’s bats. Proc. R. Soc. B 273,
multiple testing. J. R. Stat. Soc. Ser. B: Methodol. 57, 2785–2790. (doi:10.1098/rspb.2006.3647)
289–300. Krause, J., Reeves, P. & Hoare, D. 1998 Positioning
Berthold, P. 1993 Bird migration: a general survey. Oxford, behaviour in roach shoals: the role of body length and
UK: Oxford University Press. nutritional state. Behaviour 135, 1031–1039.
Biro, D., Sumpter, D., Meade, J. & Guilford, T. 2006 From Krause, J., Hoare, D., Krause, S., Hemelrijk, C. K. &
compromise to leadership in pigeon homing. Curr. Biol. Rubenstein, D. I. 2000 Leadership in fish shoals. Fish Fish.
16, 2123–2128. (doi:10.1016/j.cub.2006.08.087) 1, 82–89.
Bumann, D., Krause, J. & Rubenstein, D. 1997 Mortality risk Lamprecht, J. 1992 Variable leadership in bar-headed geese
of spatial positions in animal groups: the danger of being in (Anser indicus): an analysis of pair and family depar-
the front. Behaviour 134, 1063–1076. (doi:10.1163/ tures. Behaviour 122, 105–120. (doi:10.1163/15685399
156853997X00403) 2X00336)
Chittka, L., Dyer, A. G., Bock, F. & Dornhaus, A. 2003 Bees Leca, J. B., Gunst, N., Thierry, B. & Petit, O. 2003
trade off foraging speed for accuracy. Nature 424, 388. Distributed leadership in semifree-ranging white-faced
(doi:10.1038/424388a) capuchin monkeys. Anim. Behav. 66, 1045–1052. (doi:10.
Conradt, L. & Roper, T. J. 2003 Group decision-making in 1006/anbe.2003.2276)
animals. Nature 421, 155–158. (doi:10.1038/nature01294) Macdonald, D. W. 1996 Social behaviour of captive bush
Conradt, L. & Roper, T. J. 2005 Consensus decision making dogs, Speothos venaticus. J. Zool. 239, 525–543.
in animals. Trends Ecol. Evol. 20, 449–456. (doi:10.1016/ Maransky, B. P. & Bildstein, K. L. 2001 Follow your elders:
j.tree.2005.05.008) age-related differences in the migration behavior of broad-
Conradt, L. & Roper, T. J. 2009 Conflicts of interest and the winged hawks at Hawk Mountain Sanctuary, Pennsylvania.
evolution of decision sharing. Phil. Trans. R. Soc. B 364, Wilson Bull. 113, 350–353. (doi:10.1676/0043-5643(2001)
807–819. (doi:10.1098/rstb.2008.0257) 113[0350:FYEARD]2.0.CO;2)

Peterson, R. O., Jacobs, A. K., Drummer, T. D., Mech, L. D. Squires, V. R. & Daws, G. T. 1975 Leadership and
& Smith, D. W. 2002 Leadership behavior in relation to dominance relationships in merino and Border Leicester
dominance and reproductive status in gray wolves, Canis sheep. Appl. Anim. Ethol. 1, 263–274. (doi:10.1016/0304-
lupus. Can. J. Zool. 80, 1405–1412. (doi:10.1139/z02-021) 3762(75)90019-X)
Reader, S. M., Kendal, J. R. & Laland, K. N. 2003 Social Stine, W. W., Howell, L. L., Murdock, G. K., Newland,
learning of foraging sites and escape routes in wild M. C., Conradsen, L. & Maple, T. L. 1982 The control of
Trinidadian guppies. Anim. Behav. 66, 729–739. (doi:10. progression order in a captive herd of sable antelope
1006/anbe.2003.2252) (Hippotragus niger). Zoo Biol. 1, 89–110. (doi:10.1002/
Reebs, S. G. 2000 Can a minority of informed leaders zoo.1430010203)
determine the foraging movements of a fish shoal? Anim. Sumpter, D. J. T. & Pratt, S. C. 2009 Quorum responses and
Behav. 59, 403–409. (doi:10.1006/anbe.1999.1314) consensus decision making. Phil. Trans. R. Soc. B 364,
Reebs, S. G. 2001 Influence of body size on leadership 743–753. (doi:10.1098/rstb.2008.0204)
in shoals of golden shiners, Notemigonus crysoleucas. Vitevitch, M. S. 2002 Influence of onset density on spoken
Behaviour 138, 797–809. (doi:10.1163/156853901753 word recognition. J. Exp. Psychol. Hum. Percept. Perform.
172656) 28, 270–278. (doi:10.1037/0096-1523.28.2.270)
Roitman, J. D. & Shadlen, M. N. 2002 Response of neurons Ward, A. J. W., Sumpter, D. J. T., Couzin, I. D., Hart, P. J. B.
in the lateral intraparietal area during a combined visual & Krause, J. 2008 Quorum decision-making facilitates
discrimination reaction time task. J. Neurosci. 22, information transfer in fish shoals. Proc. Natl Acad. Sci.
9475–9489. USA 105, 6948–6953. (doi:10.1073/pnas.0710344105)
Siemers, B. & Nill, D. 2001 Fledermäuse: Ein Bildreise in die Wilson, E. O. 1980 Sociobiology. Cambridge, MA: Belknap
Nacht. München, Germany: BLV Verlagsgesellschaft mbH. Press.

Phil. Trans. R. Soc. B (2009) 364, 791–806

doi:10.1098/rstb.2008.0275
Review
Reciprocity, culture and human cooperation:

previous insights and a new
cross-cultural experiment
Simon Gächter1,2,3,* and Benedikt Herrmann1
1
Centre of Decision Research and Experimental Economics, School of Economics, University of Nottingham,
Sir Clive Granger Building, University Park, Nottingham NG7 2RD, UK
2
CESifo, Poschingerstrasse 5, 81679 Munich, Germany
3
IZA, Institute for the Study of Labor, Schaumburg-Lippe-Strasse 5-9, 53113 Bonn, Germany
Understanding the proximate and ultimate sources of human cooperation is a fundamental issue in
all behavioural sciences. In this paper, we review the experimental evidence on how people solve
cooperation problems. Existing studies show without doubt that direct and indirect reciprocity are
important determinants of successful cooperation. We also discuss the insights from a large literature
on the role of peer punishment in sustaining cooperation. The experiments demonstrate that many
people are ‘strong reciprocators’ who are willing to cooperate and punish others even if there are no
gains from future cooperation or any other reputational gains. We document this in new one-shot
experiments, which we conducted in four cities in Russia and Switzerland. Our cross-cultural
approach allows us furthermore to investigate how the cultural background influences strong
reciprocity. Our results show that culture has a strong influence on positive and in especially strong
negative reciprocity. In particular, we find large cross-cultural differences in ‘antisocial punishment’
of pro-social cooperators. Further cross-cultural research and experiments involving different socio-
demographic groups document that the antisocial punishment is much more widespread than
previously assumed. Understanding antisocial punishment is an important task for future research
because antisocial punishment is a strong inhibitor of cooperation.
Keywords: human cooperation; strong reciprocity; public goods experiments; culture;
antisocial punishment
1. INTRODUCTION strikes, embargoes and consumer boycotts; the volun-

Many important collective problems that human tary provision of public goods; donations to charities;
decision makers face are characterized by a conflict of teamwork; collusion between firms; and so on.
interest between individual and group benefit. The However, despite this bleak prediction, humans often
‘tragedy of the commons’ (Hardin 1968) is probably manage to avoid the tragedy of the commons and
the best known example. Each individual farmer has an achieve high levels of cooperation. This holds for
incentive to put as many cattle on the common meadow hunter-gatherer societies to complex modern nation
as possible. The tragic consequence may be over- states, which would not exist without large-scale
grazing from which all farmers suffer. Collectively, all cooperation. Thus, understanding cooperation is an
farmers would be better off if they were able to important challenge for all social sciences but also
constrain the number of cattle that grazes on the for evolutionary biology, because it needs to explain
commons. Yet, each individual farmer is better off by how natural and cultural evolution can lead to coope-
letting their cattle graze. Collective welfare is jeopar- ration (Hammerstein 2003; Gardner & West 2004;
dized by individual greed in such diverse areas as Henrich & Henrich 2007; West et al. 2007).
warfare; cooperative hunting and foraging; environ- This paper reviews existing evidence and presents
mental protection; tax compliance; voting; the partici- novel cross-cultural results from systematic experi-
pation in collective actions such as demonstrations, mental investigations on how people solve cooperation
problems. We believe that sound empirical knowledge
* Author and address for correspondence: Centre of Decision Research is an important input for the development of proximate
and Experimental Economics, School of Economics, University of
Nottingham, Sir Clive Granger Building, University Park, Nottingham, and ultimate theories of cooperation. Laboratory
NG7 2RD, UK (simon.gaechter@nottingham.ac.uk). experiments are probably the best tool for studying
Electronic supplementary material is available at http://dx.doi.org/10. cooperation empirically. The reason is that in the field
1098/rstb.2008.0275 or via http://journals.royalsociety.org. many factors are operative at the same time. The
One contribution of 11 to a Theme Issue ‘Group decision making in laboratory allows for a degree of control that is often
humans and animals’. not feasible in the field.1,2

792 S. Gächter & B. Herrmann Review. Reciprocity, culture and human cooperation
In particular, experiments are helpful for separating Most experiments on strong reciprocity were
out explanations why people cooperate. According to conducted with students as participants, which raises
some important proximate theories as developed in the the question of how general the observations on strong
social sciences, in particular economics, and ultimate reciprocity are across different socio-economic groups.
theories as developed in evolutionary biology, people Section 6 reviews some recent findings on this
cooperate only if it is in their (long-term) self-interest. question. Section 7 provides concluding remarks.
For instance, if the interaction is among genetic relatives
(‘kin selection’, Hamilton 1964) or if it is repeated and/
2. THE PUBLIC GOODS GAME
or if one’s reputation is at stake (‘direct reciprocity’
Many human cooperation problems—from hunter-
and ‘indirect reciprocity’, respectively), people might
gatherer societies to modern societies—often involve
have a selfish incentive to cooperate (Trivers 1971;
large numbers of individuals. The ‘public goods game’
Axelrod & Hamilton 1981; Fudenberg & Maskin 1986; is a suitable research tool for studying such n-person
Panchanathan & Boyd 2004; Nowak & Sigmund 2005; cooperation problems.4 In this game, each of n-group
Lehmann & Keller 2006; Nowak 2006). Experimental members receives an endowment of, say, 20 tokens.
approaches allow the researcher to control, by way of Participants have to decide how many tokens to keep
experimental design, the extent to which direct and for themselves and how many to contribute to a group
indirect reciprocity are possible. project, which collectively earns naO1 for each token
This paper presents attempts to study with the help invested. Each group member earns a tokens (where
of controlled laboratory experiments some important 0!a!1) for each token invested in the project,
factors that influence an important aspect of human regardless of whether he or she contributed any.
collective decision making: cooperation behaviour. Since the cost of contributing one token to the project
Our focus is on cooperation because this has been is exactly one token while the individual return on that
a particularly active research field in the behavioural token is only a!1 tokens, keeping all one’s own tokens
sciences recently (both theoretically and experimen- is always in any participant’s material self-interest—
tally), with a large potential for cross-disciplinary irrespective of how much the other group members
research (Hammerstein 2003; Hammerstein & Hagen contribute. Yet, if, for example, in a group of four
2005; Fehr & Camerer 2007; Sigmund 2007).3 and aZ0.5 each group member retains all of his or
Our paper is structured as follows. In §2, we her tokens then there are no earnings to be shared;
introduce our tool of investigation—the public goods on the other hand, each member would earn 0.5!
game. In §3, we review the evidence that shows that 80Z40 tokens if each of them invests their entire 20
both repeated interaction and possibilities for repu- token endowment.
tation formation are important determinants for The public goods game epitomizes the tension
people’s cooperation behaviour. However, there is between collective welfare and individual incentives in
also substantial cooperation in anonymous one-shot a simple and stark way because selfish rationality
games, where neither strategic reciprocity nor repu- implies full ‘free riding’ (i.e. zero contributions),
tation can matter. In experiments in which people whereas collective welfare is maximized if every player
have the possibility to punish their group members at makes maximal contributions. Owing to its simplicity
their own cost after having seen how much the other the public goods game has been used to answer
group members contributed, it turned out that the questions about how various institutional parameters,
punishment of freeloaders is an important factor to such as group size (n), the marginal gains from
explain cooperation in both one-shot and repeated cooperation (a), the strategic nature of interaction
interactions. Cooperation in one-shot games is an structures (one-shot versus repeated interaction) and
evidence for ‘strong reciprocity’ (Gintis 2000; Fehr possibilities for multilateral peer punishment influence
et al. 2002a; Fehr & Fischbacher 2003; Carpenter et al. cooperation. The public goods game is also a
in press). Strong reciprocators incur personal costs to prototypical game to study pro-social behaviour in a
punish and reward others even when this behaviour group context (Camerer & Fehr 2004). We discuss the
cannot be justified by kinship, reciprocal altruism or most important findings in §3.
reputational concerns. Thus, strong reciprocity
presents a particular challenge to evolutionary theories 3. FACTORS THAT INFLUENCE COOPERATION
of cooperation and has therefore been an important Under the assumption that agents are rational and want
topic for research in the last few years. We will provide to maximize their monetary pay-off theory predicts that
novel evidence for strong reciprocity in §4. people will not contribute to the public good. However,
Section 4 is the core of our paper because we show numerous experiments have falsified this prediction—
that people cooperate and punish in one-shot games there exists substantial cooperation in a variety of
without any repetition. The experiment we report in set-ups.5 Six sets of results are particularly noteworthy
this section also shows that there exists a substantial in the light of existing proximate and ultimate theories
cultural influence on strong reciprocity. Section 5 of cooperation.
follows up on the findings from §4 by briefly reviewing
a cross-cultural experiment conducted in 16 partici- (i) Contributions are higher, the higher the mar-
pant pools around the globe (Herrmann et al. 2008). ginal gains from contributing (i.e. a) are
This experiment demonstrates that cooperation and (Isaac & Walker 1988b; Brandts & Schram
punishment are substantially shaped by the cultural 2001; Goeree et al. 2002; Zelmer 2003;
background across a range of diverse societies. Carpenter 2007b). This is interesting because

Review. Reciprocity, culture and human cooperation S. Gächter & B. Herrmann 793
from the viewpoint of (selfishly) rational they have helped others in the past and
decision making the prediction of full free riding therefore have a favourable ‘image score’. The
in the public goods game described above experimental evidence is consistent with such a
does not depend on a, as long as a!1. mechanism (Engelmann & Fischbacher 2002;
However, this result shows that people appa- Milinski et al. 2002; Semmann et al. 2005;
rently find it easier to contribute to the public Seinen & Schram 2006).
good the higher the marginal gains from (v) Communication also greatly facilitates co-
cooperation are (Anderson et al. 1998). operation and helps in preventing its break-
(ii) Larger groups do not cooperate significantly less down (Dawes et al. 1977; Isaac & Walker
than smaller groups (Marwell & Ames 1979; 1988a; Ostrom et al. 1992; Sally 1995; Brosig
Isaac & Walker 1988b; Isaac et al. 1994; Zelmer et al. 2003; Bochet et al. 2006). Similarly,
2003; Carpenter 2007b; Cardenas & Jaramillo intergenerational advice, if common knowl-
2007). This finding goes against conventional edge, can also sustain high levels of cooperation
wisdom that maintaining cooperation should be (Chaudhuri et al. 2006). Communication is
easier in smaller groups (Olson 1965). One interesting because it is an important human
explanation might be that people are hetero- capacity that can often be fruitfully employed
geneous with respect to their willingness to in smaller groups. There are many behavioural
cooperate (more on this below). Some are ‘free reasons why communication is effective:
riders’ and others are ‘conditional cooperators’ communication might help the cooperators
who are willing to cooperate provided others to coordinate on high levels and it might
cooperate as well. Larger groups may have more involve social pressure and mutual promises
free riders than small groups, but they possibly which would induce feelings of guilt if broken
also have more cooperators. Group size per se is (Charness & Dufwenberg 2006).
therefore not decisive. (vi) There is even substantial cooperation in pure
(iii) Playing the public goods game repeatedly with one-shot public goods games without any
the same group members often leads to higher repetition (Marwell & Ames 1979; Gächter
contributions than playing it one shot and with et al. 2004; Walker & Halloran 2004; Dufwenberg
randomly changing group members (Croson et al. 2006; Gächter & Herrmann 2007;
1996; Sonnemans et al. 1999; Fehr & Gächter Cubitt et al. 2008). This evidence is consistent
2000; Keser & van Winden 2000).6 This with strong positive reciprocity. In §4, we will
finding (and related ones from indefinitely present an experimental design that sheds new
repeated prisoners’ dilemma games (Dal Bo light on strong positive reciprocity in the context
2005)) is consistent with ultimate and prox- of voluntary cooperation.
imate arguments that repeated interactions
offer strategic reasons to cooperate (Trivers An important observation in all repeatedly played
1971; Axelrod & Hamilton 1981; Kreps et al. games reported in (i)–(iv) is that people make high
1982; Fudenberg & Maskin 1986). The contributions initially but over time contributions
significance of the finding that cooperation is dwindle to low levels. The decay of cooperation
typically higher in repeated games than one- has been replicated numerous times and has also
shot games, and similar findings from related been observed across a variety of participant pools
cooperation experiments (e.g. Falk et al. 1999; (Herrmann et al. 2008). What explains this almost
Engelmann & Fischbacher 2002; Gächter & inevitable outcome? One possibility is learning the
Falk 2002; Cochard et al. 2004) is that people free-rider incentives. However, one problem with
are able to distinguish situations that require this explanation is that in experiments with a surprise
strategic cooperation from those that do not restart contributions start high again, which is
(Fehr & Fischbacher 2003). inconsistent with a pure learning hypothesis (Andreoni
(iv) Experiments under non-anonymity, where 1988; Croson 1996; Cookson 2000). People might
participants could identify the individual also have some willingness to cooperate due to feelings
behind a particular contribution, increased of ‘warm glow’ (which might explain restart effects)
contributions relative to an anonymity bench- but are otherwise confused decision makers who
mark (Gächter & Fehr 1999; Andreoni & need time to learn what is the optimal contribution
Petrie 2004; Rege & Telle 2004). People even for them. Palfrey & Prisbrey (1997) test this idea and
contribute more to public goods if they find some support for warm glow and reduced
are exposed to subconsciously activated cues confusion over time. A further explanation, long
of being observed (Bateson et al. 2006; argued by social psychologists (e.g. Kelley & Stahelski
Burnham & Hare 2007).7 This evidence is 1970), is that many people are conditional coopera-
consistent with ‘reputation effects’ noted in tors, who in principle are willing to cooperate if others
several decision tasks involving altruistic do so as well, but get frustrated if others do not pull
behaviour (Haley & Fessler 2005; Milinski & their weight. Therefore, the breakdown of cooperation
Rockenbach 2007). People might care for a is due to ‘frustrated attempts at kindness’ (Andreoni
favourable reputation because this is evolution- 1995; p. 900).
arily advantageous according to the models of There is now mounting evidence from psychological
indirect reciprocity ( Nowak & Sigmund 2005), and economic experiments for the importance of
where people are more likely to receive help if conditional cooperation both in the laboratory and

the field (Gächter 2007). In experiments that elicited group members have contributed to the public good.
participants’ beliefs about how much they think Each group member can then decide to punish each of
others will contribute, contributions are indeed the other group members. A punishment decision is
positively correlated with beliefs (Dufwenberg et al. implemented by assigning between 0 and 10 points to
2006; Croson 2007; Fischbacher & Gächter 2008; the punished member. Each point assigned reduces the
Neugebauer et al. in press). A correlation does of punished member’s income by kR1 tokens and
course not establish causation and it is perfectly costs the punishing member one token. Punishment
possible that a false consensus effect induces people decisions are also made simultaneously and people
to believe that others contribute the same as them (e.g. are not informed about who punished them. Note
Kelley & Stahelski 1970). To circumvent this problem, that a rational and money-maximizing individual
Fischbacher et al. (2001) developed an experimental will never punish (in a one-shot game) because pun-
design in which the contribution of others was fixed. In ishment is costly.
their design, people have to indicate how much they Numerous experiments have been conducted in this
contribute to the public good as a function of all framework. Some of the results that are particularly
possible average contribution levels of other group interesting from the viewpoint of proximate and
members. The results show that approximately 50 evolutionary theories of cooperation are as follows.
per cent are conditional cooperators, who increase
their contributions if others contribute more, whereas (i) Many people punish those who contribute less
approximately 25 per cent are free riders who never than them to the public good. In particular, the
contribute anything—irrespective of how much others more someone free rides, the more he or she
contribute. The rest show more complicated patterns.8 gets punished on average. This observation has
Fischbacher & Gächter (2008) use the same method as been made in all public goods experiments with
Fischbacher et al. (2001) and show that the interaction punishment we are aware of; there also seems
of differently motivated people explains the decay of to be little cross-cultural variation in the extent
cooperation. The significance of this finding is that to which people punish freeloaders (Herrmann
the decay of cooperation will occur not just because et al. 2008). Together with the cross-cultural
people eventually learn what is in their best interest but evidence from ultimatum games and third
because frustrated conditional cooperators reduce party punishment games conducted in
their contributions. Thus, after some time, all types complex large-scale and small-scale societies
behave as income-maximizing free riders, even though around the globe (Oosterbeek et al. 2004;
only the free rider types are motivated by income Henrich et al. 2005, 2006; Marlowe et al.
maximization alone. 2008), these observations suggest that punish-
The fact that many people are conditional coopera- ment of selfish behaviour is a ‘human uni-
tors but some are free riders has two important general versal’.
implications. First, the interaction structure matters (ii) The large majority of studies find that peer
(e.g. Gächter & Thöni 2005; Gunnthorsdottir et al. punishment increases and stabilizes co-
2007), i.e. there is an ‘ecology of collective action’ operation at higher levels than without punish-
(Ones & Putterman 2007). For instance, if cooperators ment. This is an important finding because the
know that they are among other ‘like-minded’ cooperation-enhancing effect of punishment is
cooperators, they are able to maintain very high levels predicted by both proximate and ultimate
of cooperation (Gächter & Thöni 2005). Second, theories of cooperation and punishment
because conditional cooperators will adjust their (Boyd et al. 2003; Fehr & Schmidt 2006;
cooperative behaviour to those observed around them Carpenter et al. in press). There are exceptions,
and to what they believe others will do, any factor that however. For instance, punishment does not
shifts people’s beliefs will shift their behaviour.9 work well if it is perceived as being unfair
Reciprocity is a likely source of conditional coope- (e.g. van Prooijen et al. 2008) or if the group
ration (Rabin 1993; Dufwenberg et al. 2006).10 The structure is asymmetric (Reuben & Riedl
reason is that cooperating is a nice act towards the other in press). There are also cross-cultural
group members and people may want to return the differences in the extent to which punishment
favour. By contrast, free riding is an unkind act which establishes cooperation (see Herrmann et al.
people may want to punish. However, in the public (2008) and §§4 and 5 of this paper).
goods experiments described above, the only way to (iii) The strategic nature of interaction (repeated
punish free riding is to withdraw cooperation, with the interaction versus one-shot interaction) mat-
consequence that other cooperators in the group get ters for cooperation but not much for punish-
punished as well. This raises two questions: will people ment (Fehr & Gächter 2000). Put differently,
be willing to punish if they could target a free rider while cooperation rates are significantly and
directly? Will the possibility to punish affect co- substantially higher in repeated interactions
operation? Numerous experiments since the seminal when compared with repeated one-shot
studies of Yamagishi (1986) and Ostrom et al. (1992) interactions, people punish free riding simi-
have given affirmative answers to both the questions. larly, irrespective of whether it occurs in a
A typical design of most recent studies is as follows repeated relationship or in random one-shot
(Fehr & Gächter 2000, 2002). After participants interactions. Moreover, as we will see in §4,
have made their contribution decisions, group people punish even in strict one-shot games
members are informed about how much the other with no repetition. Punishment is also often

harshest in the final period after people and shame that induce individuals to behave
had experienced as many as 50 rounds of pro-socially (Barr 2001; Fessler & Haley
cooperation and punishment (Gächter et al. 2003). Hopfensitz & Reuben (in press) provide
2008). Any learning about the selfish incen- direct evidence for the role of shame and guilt
tives of the game should have taken place by in response to being punished. However,
then. Thus, these observations suggest that the recent cross-cultural experiments suggest that
level of cooperation is influenced by strategic punishment might not trigger guilt and shame
considerations (free riding is less likely in in the same way everywhere, because in some
repeated interactions), whereas punishment is participant pools punishment does not induce
to a large part non-strategic. Punishment freeloaders to increase their contributions
seems to be an impulse triggered by negative (Gintis 2008; Herrmann et al. 2008).
emotions (Pillutla & Murnighan 1996; Bos- (vii) In most experiments in which punishment has
man & van Winden 2002; Fehr & Gächter material pay-off consequences, punishment
2002; Sanfey et al. 2003; de Quervain et al. turned out to be an inefficient tool to enforce
2004; Knoch et al. 2006; Ben-Shakhar et al. cooperation because resources are destroyed.
2007; Fehr & Camerer 2007; Seymour et al. Indeed, in most experiments—which typically
2007; Reuben & van Winden 2008) and not ran for 10 periods or less—net pay-offs in
much by forward-looking considerations. treatments with punishment were often lower
(iv) Although punishment is most likely to a large than in treatments without punishment (e.g.
extent non-strategic and not forward looking, it Fehr & Gächter 2000; Page et al. 2005; Bochet
follows economic rationality (cost–benefit et al. 2006; Botelho et al. 2007; Sefton et al.
considerations) in the sense that punishment is 2007; Dreber et al. 2008; Egas & Riedl 2008;
less likely used the more costly it is for the Herrmann et al. 2008; Masclet & Villeval 2008;
punishing individual (Anderson & Putterman Nikiforakis 2008). For instance, Herrmann
2006; Carpenter 2007a; Egas & Riedl 2008). et al. (2008) has reported public goods experi-
The monitoring frequency and the severity ments with and without punishment conducted
of punishment inflicted on the punished indi- in 16 comparable participant pools around the
vidual also matters for the effectiveness world. With the exception of three participant
of punishment to stabilize (or increase) pools, the average pay-off in the experiments
cooperation (Carpenter 2007b; Egas & Riedl with punishment opportunities was lower than
2008; Nikiforakis & Normann 2008). without punishment; and in those three partici-
(v) There exists an interaction effect between the pant pools with higher pay-offs, the increase
availability of punishment opportunities and was modest and amounted to 9.1, 2.8 and 0.5
direct reciprocity at the cooperation stage within per cent, respectively. Thus, 13 participant pools
stable groups. A repeated interaction and would have been better off not having had a
punishment are mutually reinforcing means to punishment opportunity. The detrimental con-
achieve high cooperation (e.g. Fehr & Gächter sequences of punishment are even more con-
2000; Masclet et al. 2003). If only direct spicuous if ‘counter-punishment’, i.e. multiple
reciprocity is possible, cooperation collapses, rounds of punishment, is possible (Denant-
albeit it is higher than in random interactions. If Boemont et al. 2007; Nikiforakis 2008).
only punishment is possible but groups are (viii) The observation that punishment leaves groups
formed randomly and hence direct reciprocity is worse off compared with experiments without
not feasible, cooperation is stabilized at inter- punishment raises several interesting questions.
mediate levels. One reason why this is so is that For instance, Dreber et al. (2008) replicated the
punishment gives selfish individuals an incen- finding of the inefficiency of punishment in
tive to cooperate and therefore also reinforces prisoner’s dilemma experiments, and argue with
the beliefs of conditional cooperators that others reference to evolutionary (group-selection)
will cooperate (Shinada & Yamagishi 2007). models of altruistic punishment (in particular
The experiment by Rockenbach & Milinski Boyd et al. 2003) that ‘[P]unishment therefore
(2006) suggests that indirect reciprocity and has no benefit for the group, which makes it hard
punishment mutually reinforce cooperation as to argue that punishment might have evolved by
well. The advantage of direct and indirect group selection’ (p. 349). However, the obser-
reciprocity is that both help keeping the vation that punishment is detrimental for group
absolute costs of punishment low because they pay-offs stems predominantly from experiments
provide additional reasons to cooperate, and that ran for 10 periods or less. Since punishment
therefore reduce the need to maintain co- is to a large extent emotional and not forward
operation by costly punishment.11 looking, and because punishment is particularly
(vi) Interestingly, punishment can also increase used when cooperation is low, which typically is
cooperation if it is purely symbolic and merely the case at the beginning of the experiment, the
expresses social disapproval, without any beneficial effects of punishment need more time
material consequences for the punished indi- to show up. Gächter et al. (2008) tested this
vidual (Masclet et al. 2003; Carpenter et al. possibility in experiments that ran for 50 periods
2004; Noussair & Tucker 2005). This suggests and they compared pay-offs with those in
that punishment also triggers feelings of guilt 10-period experiments. As in previous

experiments, in the 10-period experiments games, because punishment is costly and bears
punishment was detrimental in terms of pay- no future benefits. To test this prediction, Fehr &
offs when compared with 10-period experiments Gächter (2002) conducted six rounds of anon-
without punishment. In the 50-period experi- ymous public goods experiments with punish-
ments the opposite conclusion holds—co- ment under the perfect stranger matching
operation is high and punishment costs design. In contrast to predictions, they observed
negligible. Thus, if the time horizon is long substantial punishment of free riders in all
enough, punishment can be group beneficial, a rounds. Punishment under these circumstances
finding that supports models of group selection is therefore evidence for strong negative recipro-
(Sober & Wilson 1998; Henrich & Boyd 2001; city. Punishment is ‘altruistic’ because it is costly
Boyd et al. 2003; Bowles 2006; Bowles & Choi for the punisher, but due to the changed group
2007). A second interesting question is whether composition in each round a punisher has no
people would adopt a sanctioning institution if chance to benefit if the punished individual
they had a choice. Gürerk et al. (2006) answered subsequently increases his or her contribution;
this question affirmatively, but there is an only others benefit.12 People punish others even
interesting twist. At the beginning of the in strict one-shot games without any repetition
experiments, people predominantly chose the (Walker & Halloran 2004; Gächter & Herrmann
non-sanctioning institution. As usual, there was 2007; Cubitt et al. 2008). In §4, we will
substantial free riding, which tipped many provide further comprehensive evidence for
people over to the punishment institution. strong negative reciprocity as it occurs in strict
Punishment then became the predominant one-shot games.
choice for almost all people and very high levels
of cooperation were established. Again, humans In summary, there can be no doubt that direct and
can also often communicate and coordinate indirect reciprocity strongly shape human cooperation.
punishment (Boehm 1993; Wiessner 2005; However, there is also substantial cooperation when
Reuben & van Winden 2008), which can these channels are not available. We turn to this
minimize punishment costs. Finally, people observation in §4.
can also frequently choose with whom to
associate. Experiments show that both com-
munication (Bochet et al. 2006) and voluntary 4. STRONG RECIPROCITY AND CULTURAL
association (Page et al. 2005) are indeed effective BACKGROUND
means to avoid the detrimental effects of In this section, we present an experiment that sheds
punishment. new light on strong positive and negative reciprocity.
(ix) Given that people are willing to incur costs to This experiment also investigates how the cultural
punish others, would they also be willing to incur background influences patterns of both strong positive
costs to reward others and would rewards (which and negative reciprocity. The evidence on strong
are not efficiency reducing) steer people towards positive and negative reciprocity reviewed in §3 has
high contributions? Sefton et al. (2007) investi- contributed to the development of ultimate (e.g. Boyd
gated this question in a design in which people et al. 2003) and proximate theories of why people
could mutually reward each other such that a cooperate and punish (see Fehr & Schmidt (2006) for
reward was a mere transfer of money from the a survey). Among the most important proximate
rewarding subject to the rewarded subject. They psychological mechanisms are concerns for equity
compared this with punishment, i.e. a situation (Loewenstein et al. 1989; Dawes et al. 2007), and the
in which one punishment point assigned punishment of kind and unkind intentions (Falk et al.
reduced the punished participant’s income by 2005; Houser et al. 2008). These theories assume
one money unit whereas the punisher had to implicitly that motivations for strong reciprocity are
incur a cost of one. It turned out that people are similar across cultures (on average). Two reasons make
prepared to reward cooperators, but punish- it likely that the cultural environment exerts an
ment is more effective to increase contributions influence on strong reciprocity, however. First, people
than rewards (see also Sutter et al. (2008) who have an innate ability to learn from others (Boyd &
got a similar result in a related design). The Richerson 1985; Tomasello et al. 2005). Cultural
problem with rewards is that they need to be learning mechanisms will cause members of social
used when cooperation occurs, whereas punish- groups to adopt similar values and beliefs about how
ment can work as a mere threat and need not be others around them will reward and punish their
used much if people cooperate. behaviour (Sober & Wilson 1998; Henrich & Henrich
(x) Of particular relevance for evolutionary theories 2007). Second, both strong positive and negative
of cooperation are experiments where any future reciprocity might be shaped by local social norms
interaction with the same group members is about what constitutes the appropriate reaction to a
excluded by design (so-called ‘perfect stranger’ benefit or harm one has received from others
matching). The reason why this is interesting is (Gouldner 1960; Coleman 1990; Sober & Wilson
that the theories of direct and indirect reciprocity 1998; Henrich & Henrich 2007).
can explain why selfish people cooperate in We are not the first to study cultural influences on
repeated games with the same players but these strong reciprocity (seminal studies are by Henrich
theories predict little cooperation in one-shot et al. 2005, 2006).13 However, our methodology

differs in several important ways from previous (a) 8

approaches. First, we conducted public goods experi-
7
mean expected punishment

ments with and without punishment, whereas previous
studies mainly investigated bargaining games or 6
third-party punishment games. Our set of games also
5
allows us to study strong positive and negative
reciprocity within one framework. In the context of 4
our games, a strong reciprocator is predisposed to 3
punish the non-cooperators (strong negative recipro-
city) and to cooperate if others cooperate (strong 2
positive reciprocity). 1
Second, we conducted our experiments one shot,
0
anonymously and with people who did not know
each other (the average participant had known only 6 (b) 8
per cent of other participants), because we wanted 7
to measure strong reciprocity in a situation that was
mean actually received

not confounded with reputational or strategic consi- 6
derations coming from repeated play (Milinski et al.
punishment
5
2002; Fehr & Fischbacher 2003; Rockenbach &
Milinski 2006). 4
Third, we elicited beliefs about how much others 3
will contribute and how much they will punish. Owing
to the one-shot nature of our experiments, participants 2
deliberately could not base their expectation about how 1
others were likely to behave on any observation made in
0
the experiment. Participants had to form their expec- (–20, –2) (–2, 2) (2, 20)
tations based on their experiences in daily life outside
deviation from group average
the laboratory. When we elicited beliefs we also asked
participants how confident on a 10-point scale (1, very Figure 1. Cultural influences on strong negative reciprocity.
unconfident; 10, very confident) they were about their (a) Mean punishment expected and (b) mean punishment
estimate. This is a measure of how precise people think received from other group members for a given deviation of
their estimate is. own contribution from the group average. The error bars
indicate the bootstrapped 95 per cent confidence bounds for
Fourth, we conducted our experiments in two
country averages. Diamonds, Belgorod; squares, Yekaterin-
highly developed industrialized countries (Russia burg; circles, Russia; minus symbols, St Gallen; crosses,
and Switzerland). We are not interested in these Zurich; plus symbols, Switzerland.
countries per se, but they make interesting test cases
as the ‘cultural distance’ between these societies is sequence and strong negative reciprocity in the
almost the largest one compared with all developed P-experiment of the P–N sequence. The reason for
societies from which data are available.14 We ran the the two sequences is to see how participant pools
Russian experiments in Belgorod and Yekaterinburg react when punishment opportunities are added (in
and the Swiss experiments in St Gallen and Zurich.15 the N–P sequence), or removed (in the P–N
If the wider societal and cultural background influ- sequence). Moreover, we can compare cooperation
ences patterns of strong reciprocity then it should in the N-experiment of the N–P sequence with the
affect beliefs and behaviour similarly in the two P-experiment of the P–N sequence to see to what
participant pools within a society and differently extent people anticipate the presence of a punishment
between societies.16 option in their cooperation behaviour without any prior
The specifics of our design are as follows. Groups experience of the cooperativeness of others. A total of
of three participants played an anonymous one-shot 603 people (360 Russian and 243 Swiss students)
public goods game (with aZ0.5). We had two participated in either the N–P sequence (nZ336) or
treatment conditions, one with no punishment oppor- the P–N sequence (nZ267).
tunities (called the ‘N-experiment’, to measure strong Figure 1a shows that in a case where a group
positive reciprocity) and one with punishment oppor- member’s contribution was lower than the group
tunities (‘the P-experiment’, to measure strong average contribution, expected punishment was very
negative reciprocity). similar across participant pools (Kolmogorov–Smirnov
All participants took part in both a one-shot test (KS test), pZ0.821). However, very strong
N-experiment and a one-shot P-experiment. We had differences between participant pools emerge in a
two sequences: the N–P sequence, in which partici- case where a subject made similar contributions to
pants first played the N-experiment and then the those of his or her group members or contributed
P-experiment; in the P–N sequence this order was even more. In both cases we find that the Russian
reversed. In both sequences, participants were unaware participant pools expected much more severe punish-
about the second experiment until they had finished ment than their Swiss counterparts. While the Swiss
the first one. This ensures the one-shot nature of the participants expected to receive 1.5 punishment points
first experiments. We will therefore measure strong on average (with no significant differences (at aZ0.05)
positive reciprocity in the N-experiment of the N–P between the two Swiss participant pools), their

Russian counterparts expected to receive almost 4.5 (a) 100

punishment points (also with no significant differences 90
(at aZ0.05) between participant pools). This
cumulative frequency (%)

difference is highly significant (KS test, p!0.004). 80
Although participant pools held very different beliefs 70
about the punishment they expected from their group 60
members, people in all participant pools were similarly 50
confident about their estimate. The average subject
40
reports a confidence level of 6.03 and significantly
more participants have a confidence level in the upper 30
half than in the lower half of the scale (two-sided 20
binomial test, pZ0.005). 10
Actual punishment (figure 1b) also shows a striking
0
difference between the Swiss and the Russian partici- 0 2 4 6 8 10 12 14 16 18
pant pools. There are no significant differences (at beliefs about others’ average contribution
aZ0.05) between the participant pools within a
society. However, punishment is highly significant (b) 20
and substantially harsher in the Russian than in the 18
Swiss participant pools. This holds true for all
mean actual contributions

16
deviation intervals (KS test, p!0.005). The Russian 14
participant pools punished not only the low contribu-
12
tors more severely than the Swiss participant pools, but
also those who contributed at least as much as the 10
group average. In the Swiss participant pools, punish- 8
ment was almost exclusively directed at the low 6
contributors. Thus, the cultural differences in actual 4
punishment are not only in the severity with which
2
people punish low contributors, but also in the way
they punish high contributors. Such ‘antisocial punish- 0
(0, 6) (7, 13) (14, 20)
ment’ (Herrmann et al. 2008) is particularly puzzling,
beliefs about others’ average contributions
given that our one-shot design excludes retaliation
(Herrmann et al. 2008; Nikiforakis 2008) for punish- Figure 2. Cultural influences on strong positive reciprocity.
ment received in the past as an explanation. (a) Distribution of beliefs about the average contribution of
Are there also cultural influences on strong positive the other two group members, separately for each participant
reciprocity as measured in the N-experiments of the pool and pooled for the Russian and the Swiss participant
N–P sequence? pools, respectively. KS test indicates Kolmogorov–Smirnov
As figure 2a shows, beliefs about others’ contri- tests about the equality of distributions. (b) Mean actual
contribution of a given belief about others’ contribution.
butions are not significantly different either between
Error bars indicate bootstrapped 95 per cent confidence
societies or between participant pools within societies bounds of country averages. Diamonds, Belgorod (B);
(KS test, pO0.489). Strong positive reciprocity in squares, Yekaterinburg ( Y ); circles, Russia; minus symbols,
our one-shot game requires that people who believe St Gallen (S); crosses, Zurich (Z); plus symbols, Switzerland.
that others make a high (low) contribution will
reciprocate by contributing a high (low) amount as Russian than the Swiss participant pools (KS test,
well (Fischbacher et al. 2001; Fehr & Fischbacher p!0.006); there were no significant differences
2003; Dufwenberg et al. 2006; Croson 2007). Thus, within societies (KS test, pO0.143). Similarly, in the
beliefs about others’ contributions and own contri- P-experiment of the P–N sequence, contributions of
butions should be positively correlated. This is indeed the Swiss participant pools were significantly higher
the case in all participant pools (figure 2b). However, than the Russian participant pools (KS test, p!0.001).
despite the fact that beliefs are not significantly As in the N-experiment, there are virtually no
different between participant pools, we also find differences in the distribution of contributions within
cultural influences on strong positive reciprocity in both the Russian and the Swiss participant pools (KS
the sense that the relationship between contributions test, pO0.659).
and beliefs is steeper in both the Swiss pools than in As a consequence of different cooperation and
both the Russian pools. The main reason for this punishment patterns, earnings in the P-experiment are
difference is that contributions towards high beliefs highly significantly different between the Russian
about others (expected contributions in the interval and the Swiss participant pools, but not significantly
(14, 20)) are substantially lower in the Russian different within societies. Eighty per cent of the
participant pools than in the Swiss participant pools Russian participants earned less than 20 money units—
(KS test, pZ0.001); no significant differences can be the earnings predicted for selfishly rational players. In
detected in the other intervals (KS test; pO0.113). Switzerland, this was true for 33 per cent of participants.
The cultural differences in strong reciprocity also Our final steps are, first, to compare contributions to
had an impact on cooperation (figure 3). In the the N-experiments of the N–P sequence and the
N-experiment of the N–P sequence, the resulting P-experiments of the P–N sequence. This analysis
contributions levels were significantly lower in the informs us about the extent to which participants

(a) 20 Belgorod Yekaterinburg St Gallen Zurich

18
16
mean contribution
14
12
10
8
6
4
2
0
first second first second first second first second
experiment
(b) 20 lowest middle highest
18
16
mean contribution
14
12
10
8
6
4
2
0
N P N P N P
experiment (N–P sequence)
Figure 3. Cultural differences on the impact of strong reciprocity on cooperation. (a) Change in contributions when punishment
is added (in the N–P sequence) or subtracted (in the P–N sequence). The thickness of the connecting lines indicates the
significance level of the behavioural change according to Wilcoxon matched-pair tests (with group average contributions as
independent observations): , pO0.1; , p!0.05; , p!0.01. The error bars are the bootstrapped 95 per cent
confidence bounds of mean contributions (circles, N-experiment; triangles, P-experiment). (b) Change in contribution in the P-
experiment compared with the N-experiment in the N–P sequence by the groups’ minimum, middle and maximum contributors
of the N-experiment. We indicate the p-values of Kruskal–Wallis tests of equality of contributions in all four participant pools.
Diamonds, Belgorod; squares, Yekaterinburg; minus symbols, St Gallen; crosses, Zurich.
anticipate punishment in their contribution behaviour investigate how they change their contribution to the
without any prior experience of others’ behaviour. P-experiment. We classify each group member in the
Second, we look at the change in contributions to the N-experiment whether he or she is the lowest, middle
N–P sequence, where we introduce a punishment or highest contributor in his or her group (figure 3b).
opportunity after participants have had some experi- The lowest contributors in the N-experiment in the
ence with cooperation behaviour in the N-experiment. Swiss participant pools increased their contributions to
Zurich is the only participant pool where contri- the P-experiment substantially (by 6.83 tokens on
butions are significantly higher in the P-experiment average), whereas in Russia the lowest contributors
than in the N-experiment ( KS test, pZ0.006; raised their contribution to the P-experiment only
comparing the first experiments in a sequence). In modestly (by 1.60 tokens on average). Similarly, the
the other participant pools, contributions are only middle contributors raised their contributions to both
insignificantly higher ( Yekaterinburg and St Gallen; the Swiss participant pools, whereas in the Russian
KS test, pO0.215) or even slightly lower (Belgorod, participant pools contributions dropped. Surprisingly,
KS test, pZ0.996). the top contributors lowered their contributions in all
In the P–N sequence, contributions from all four four participant pools.
participant pools are highly significantly lower in the In summary, the experiment presented here
N-experiment than in the preceding P-experiment. By unambiguously shows two things: first, people on
contrast, in the N–P sequence in both the Swiss average are strong reciprocators who cooperate if they
participant pools, contributions to the P-experiment believe others cooperate and punish free riders.
are significantly higher than the N-experiment. The Second, strong reciprocity, especially strong negative
opposite is true in both the Russian participant pools.17 reciprocity, is subject to substantial cultural influences.
To shed light on the cultural differences in the A particularly noteworthy phenomenon is the anti-
dynamics of cooperation when a punishment option social punishment observed in the Russian participant
is added, we look at individual group members in pools—people punished not only the free riders but the
the N-experiments of the N–P sequence and cooperators too, and the latter even expected being

punished.18 In the remaining two sections, we present 6. SOCIO-DEMOGRAPHIC INFLUENCES ON

evidence on how general the findings are along two STRONG RECIPROCITY
important dimensions: different societies (§5) and In most experiments discussed above, researchers had
different socio-economic groups (§6). used participants who were similar in age, educational
and socio-economic background; in the cross-cultural
experiments, the rationale was to maximize compar-
ability across participant pools. However, there is
5. ANTISOCIAL PUNISHMENT ACROSS evidence that some socio-demographic characteristics
SOCIETIES (in particular, age) matter for social preferences
The results from §4 suggest that the cultural back- (e.g. Fehr et al. 2002b; Carpenter et al. 2005b, 2008;
ground matters for cooperation and punishment Holm & Nystedt 2005; Bellemare & Kröger 2007;
behaviour. Stimulated by this result, Herrmann et al. Sutter 2007; Sutter & Kocher 2007; Bellemare et al.
(2008) undertook a large-scale experiment across 16 2008; Dohmen et al. 2008; Egas & Riedl 2008). This
different participant pools in 15 different societies raises the question of whether the patterns of punish-
around the world. In their experiments, groups of four ment observed above also hold for a more representa-
played 10 periods of a public goods game without tive sample of the population, not just young people.
punishment followed by 10 periods without punish- To test for the generalizability of our findings, we ran
ment. The results showed striking similarities as well as experiments very similar to those reported in §4 with
differences in punishment behaviour. The striking 566 Russian urban and rural dwellers of all age cohorts
similarities occurred in the punishment of free-riding (Gächter & Herrmann 2007). We were also interested
behaviour: across all subject pools people punished in running the experiments in urban and rural areas,
freeloaders very similarly. Large differences arose in the because the gap between them is particularly pro-
punishment of cooperators (antisocial punishment). In nounced in Russia. Moreover, norm enforcement
some subject pools antisocial punishment was virtually may be easier in close-knit rural communities than
absent, whereas in others it was as prevalent as in anonymous urban areas (Bowles & Gintis 2002).
punishment of freeloaders. As a consequence, co- We ran our experiments in the urban area of Kursk,
operation levels were vastly different: some participant a city in the heartland of the former Soviet Union, and
pools invested almost all their endowment to the public in the rural areas surrounding Kursk. We had four
good, whereas in others people invested less than a participant pools: two mature pools (‘urban mature’
third. Punishment stabilized cooperation everywhere. and ‘rural mature’), i.e. people who on average were
In the experiment without punishment cooperation 44 years old and had spent most of their life in a big city
collapsed, as in almost all previous experiments. (a rural area); and two young participant pools with
What explains antisocial punishment? Tentative an average age of 21 years (‘urban young’ and ‘rural
answers can be given at two levels. At a macro level, young’). The design was the same as the one described
Herrmann et al. (2008) found that antisocial punish- above. The only exception was that for practical
ment occurred predominantly in societies with weak purposes the experiments were hand run and we
social norms of cooperation, weak rules of law and did not elicit beliefs.
weak democracies, according to measures developed by The results strongly resemble the ones reported
various social scientists using representative survey above. We found in all four participant pools not only
data. At the individual level antisocial punishment may high levels of punishment of people who contributed
be motivated by revenge (Denant-Boemont et al. 2007; less than the punishing subject but also substantial
Nikiforakis 2008), at least in some societies (Herrmann antisocial punishment of people who contributed the
et al. 2008; Mohan 2008). There might also be cultural same or even more. In no participant pool did
differences in the extent to which people are motivated punishment lead to an increase in cooperation. In
by relative pay-offs (Liebrand et al. 1986; Zizzo 2003; particular, contributions in all four pools dropped even
Fliessbach et al. 2007) and concerns for dominance in the N–P sequence, as in the experiments reported
(Clutton-Brock & Parker 1995). People might also above. None of the socio-demographic background
dislike ‘do-gooders’ ( Monin 2007), punish non- variables matters for punishment but some of them
conformists (Carpenter & Matthews 2005) and punish matter for cooperation behaviour. In particular, rural
displays of conspicuous generosity (Henrich et al. dwellers were more cooperative than their urban
2006). Some punishment might also be motivated by counterparts and the older people were the more they
selfish considerations to induce others to contribute contributed to the public good both in the N- and
even more (Eldakar et al. 2007). Finally, punishment the P-experiment.
might be linked to the perception of group boundaries: Our observation that age only matters for co-
some (traditional) societies are structured along strong operation behaviour but not for punishment stands in
private networks with a lot of cooperation within contrast to findings from public goods experiments
networks and little beyond. Because participants did with and without punishment conducted with more
not know each other (and were outside each others’ than 800 Dutch people from all age cohorts (average
networks), they might not have accepted punishment age 35 years; Egas & Riedl 2008). They found that age
from an outsider. Punishment might trigger anger, not was only (weakly) significantly (and not very robustly
guilt (Gintis 2008). Indeed, antisocial punishment against other specifications) correlated with contri-
occurred predominantly in more traditional, segmen- butions. However, unlike in our Russian experiments,
tary societies. Which of these explanations is important age was a significant predictor for punishment
is a task for future research. behaviour—the older people were the more they

punished others, ceteris paribus. Thus, the relevance of circumstances. See Friedman & Sunder (1994) for an introduction to
socio-demographic background variables may also be methods in experimental economics; Guala (2005) for a discussion of
the methodology of experimental economics; and Kagel & Roth (1995)
subject to cultural influences.
and Camerer (2003) for an overview of the important experimental
results across a variety of human decision making problems.
2
Conducting experiments in environments outside the university
7. CONCLUDING REMARKS laboratory (‘field experiments’) is a burgeoning area in experimental
From the experimental evidence we reviewed here, economics. See e.g. Carpenter et al. (2005a) for an overview and
there can be no doubt that direct reciprocity (aka some applications.
3
‘reciprocal altruism’) and indirect reciprocity (helping Other important areas of empirical research in collective decision
those in good standing) are very important making concern coordination problems and problems of collective
determinants of human cooperative behaviour. Yet, choice. For lack of space we do not discuss this research here. We refer
the reader to Camerer (2003) and Devetag & Ortmann (2007) for recent
there is substantial accumulated evidence that people
surveys on coordination games, and Palfrey (2008) on experiments in
also cooperate and punish in anonymous one-shot collective choice. For further aspects of human collective decision
games where future gains from cooperation, or making, see Austen-Smith & Feddersen (2009), Conradt & Roper
reputational benefits, are excluded by design. We (2009), Dyer et al. (2009), Hix (2009) and Skyrms (2009).
4
view the numerous observations of substantial co- The prisoner’s dilemma is another useful tool for studying cooperation.
operation and punishment in one-shot games as It was particularly popular in early experimental research on
supporting evidence for strong reciprocity. cooperation. See Rapoport & Chammah (1965) and Colman (1999)
for overviews, and Dreber et al. (2008) for a recent example. The
We believe that understanding strong reciprocity is
disadvantage of the prisoner’s dilemma is that it is restricted to bilateral
of importance for a variety of behavioural disciplines interactions, which have different theoretical properties from multi-
for which cooperation (and culture) are central issues lateral interactions, in particular in repeated interactions (e.g. Boyd &
(Ostrom 1998; Fehr & Fischbacher 2003; Hagen & Richerson 1988).
5
Hammerstein 2006; Sigmund 2007). The findings For overviews, see Dawes (1980), Ledyard (1995), Kollock (1998),
reviewed here, in particular those from the cross- Zelmer (2003), Gächter & Herrmann (2005) and Gächter (2007).
6
cultural experiments, support anthropological and There are some exceptions. See e.g. Andreoni (1988), Weimann
(1994) and Andreoni & Croson (2008) for an overview.
evolutionary theories of cooperation which predict 7
Cues of kinship also increase cooperation (Madsen et al. 2007; Krupp
that people’s social preferences are programmable et al. 2008).
and therefore culturally variable ( Henrich 2004; 8
Herrmann & Thöni (in press) and Kocher et al. (2008) replicated the
Henrich et al. 2005). Our results also demonstrate Fischbacher et al. (2001) study using the same parameters. They got
that to explain our patterns of strong reciprocity similar results. See Kurzban & Houser (2005), Bardsley & Moffatt
models of decision making in game theory, economics (2007) and Muller et al. (2008) for related studies that also report
and psychology need to develop models of social substantial individual differences in cooperative attitudes. See Doebeli
preferences (e.g. concerns for equity and the reward et al. (2004) for an evolutionary explanation of type heterogeneity.
9
See Gächter (2007) for several examples and a general discussion.
and punishment of kind and unkind intentions ( Falk 10
Conformity is another source of conditional cooperation—people just
et al. 2005)) that take cultural influences on those do what others do. Carpenter (2004) and Bardsley & Sausgruber (2005)
motivations into account. In particular, the role of provide evidence for the relevance of conformity in voluntary
cultural influences on strong negative reciprocity cooperation. See Gächter (2007) for an overview of studies on
deserves extensive scrutiny as here the cultural conditional cooperation and discussions of related issues.
11
differences appear to be largest (Herrmann et al. Another mechanism to keep the costs of altruistic punishment low is
2008). Previous explanations have focused predomi- when punishment leads to a reputational benefit for the punisher. See
Barclay (2006) for a study that suggests this possibility.
nantly on altruistic punishment of low contributors 12
Egas & Riedl (2008) replicated this result with a large number of
(Sigmund 2007). Our results show that there is also a Dutch residents across all age cohorts and various socio-demographic
need to understand why people punish those who backgrounds.
behave pro-socially and what the cultural determinants 13
See the supplementary materials for further references to cross-
of antisocial punishment are. cultural experiments.
14
We conducted our experiments in Yekaterinburg and Belgorod
We are grateful to the Universities of Belgorod, Yekaterin- (Russia) and St Gallen and Zurich (Switzerland). Both countries
burg, Zurich and St Gallen for their support in running the are highly industrialized, rely on large-scale division of labour and
experiments. We also gratefully acknowledge financial have extensive trade among genetically unrelated strangers.
support from the University of Nottingham, the Latsis Compared to the small-scale societies of previous studies (Henrich
Foundation (Geneva) and the EU-TMR Research Network et al. 2005, 2006), the main distinguishing features between Russia
ENDEAR (FMRX-CT98-0238). We received helpful com- and Switzerland are therefore not in the fundamentals of socio-
ments from the referees and various workshop audiences, in economic organization but in historical, religious, political and
particular the Arts and Humanities Research Council work- cultural values, which are hugely different between these societies
shops Culture and the Mind in Sheffield, and from Jo Morgan according to frequently used measures developed by various
and Daniel Scruton. S.G. also gratefully acknowledges the social scientists interested in quantifying cultural and societal
hospitality of the Center for Economic Studies in Munich and differences (Inglehart & Baker 2000; Hofstede 2001). The cultural
the Economics Department of the University of Sydney while distance between Switzerland and Russia (measured as the Euclidean
working on this paper. This paper is part of the MacArthur distance between country scores of the respective indicators) is
Foundation Network on Economic Environments and the almost the largest one compared with the 55 countries from which
Evolution of Individual Preferences and Social Norms. data are available. See the electronic supplementary material for
further details.
15
Belgorod is a medium-sized city (roughly 300 K inhabitants) in the
ENDNOTES southwest of Russia, near the border to Ukraine. Yekaterinburg is a
1
In all the laboratory experiments we discuss, participants, depending big city (more than 1000 K inhabitants) in the Ural region, 1000
on their decisions, earned considerable amounts of money. Thus, the miles east of Moscow. These cities are representative of Russia
laboratory allows observing real decision making under controlled outside Moscow. Zurich is located in the centre of Switzerland and

its urban area has roughly 1000 K inhabitants. St Gallen has Barclay, P. 2006 Reputational benefits for altruistic punish-
roughly 80 K inhabitants and is the major centre in the Northeast ment. Evol. Hum. Behav. 27, 325–344. (doi:10.1016/
region of Switzerland. Both cities are representative of German- j.evolhumbehav.2006.01.003)
speaking Switzerland. Bardsley, N. & Moffatt, P. G. 2007 The experimetrics of
16
To maximize comparability across participant pools, we public goods: inferring motivations from contributions.
implemented the following procedures: (i) We had all instructions Theory Decision 62, 161–193. (doi:10.1007/s11238-006-
translated into Russian, and back-translated, to control for language- 9013-3)
induced differences in meaning; (ii) All instructions were written in a
Bardsley, N. & Sausgruber, R. 2005 Conformity and
neutral language, to avoid evoking culture-specific meanings; (iii) We
reciprocity in public good provision. J. Econ. Psychol. 26,
followed exactly the same protocol in the manner in which we
664–681. (doi:10.1016/j.joep.2005.02.001)
conducted the experiments in all participant pools—in particular,
participants had to answer the same set of control questions that
Barr, A. 2001 Social dilemmas and shame-based sanctions:
tested their understanding of payoff calculations before the experi- experimental results from rural Zimbabwe. Center for the
ment could start; (iv) We conducted all experiments with people who Study of African Economies working paper. WPS2001.11.
did not know each other and (v) in computerized laboratories in Bateson, M., Nettle, D. & Roberts, G. 2006 Cues of being
which participants were visually separated from one another to ensure watched enhance cooperation in a real-world setting. Biol.
between-subject anonymity and to maximize subject– experimenter Lett. 2, 412–414. (doi:10.1098/rsbl.2006.0509)
anonymity; (vi) We used the same software ( Fischbacher 2007), i.e. Bellemare, C. & Kröger, S. 2007 On representative social
participants saw the same interface (except for different languages); capital. Eur. Econ. Rev. 51, 183–202. (doi:10.1016/
(vii) During the experiment, we calculated all incomes in ‘Guilders’, j.euroecorev.2006.03.006)
to avoid number and currency effects whose perception might differ Bellemare, C., Kröger, S. & Van Soest, A. 2008 Measuring
across cultures; (viii) We used the same stake size in relative monthly inequity aversion in a heterogeneous population using
income, i.e. we chose the exchange rate between ‘Guilders’ and the experimental decisions and subjective probabilities. Econ-
local currency such that real expected earnings were roughly the ometrica 76, 815–839. (doi:10.1111/j.1468-0262.2008.
same; and (viii) to minimize experimenter effects, the same 00860.x)
experimenter (B. Herrmann, who speaks German and Russian Ben-Shakhar, G., Bornstein, G., Hopfensitz, A. & van
fluently) organized and supervised all 25 sessions according to exactly Winden, F. 2007 Reciprocity and emotions in bargaining
the same script. See the supplementary information (available upon using physiological and self-report measures. J. Econ.
request) for further details.
17 Psychol. 28, 314–323. (doi:10.1016/j.joep.2007.02.005)
With regard to strong positive and negative reciprocity (and its
Bochet, O., Page, T. & Putterman, L. 2006 Communication
impact on cooperation and earnings), we get very similar results in the
and punishment in voluntary contribution experiments.
second experiments of our N–P and P–N sequence. Thus, our
findings are robust to order effects.
J. Econ. Behav. Organ. 60, 11–26.
18
Cinyabuguma et al. (2006) call the punishment of cooperators Boehm, C. 1993 Egalitarian behavior and reverse dominance
‘perverse punishment’. hierarchy. Curr. Anthropol. 34, 227–254. (doi:10.1086/
204166)
Bosman, R. & van Winden, F. 2002 Emotional hazard in a
power-to-take experiment. Econ. J. 112, 147–169. (doi:10.
1111/1468-0297.0j677)
REFERENCES Botelho, A., Harrison, G. W., Costa Pinto, L. M. & Rutström,
Anderson, C. M. & Putterman, L. 2006 Do non-strategic E. E. 2007 Social norms and social choice. Working paper
sanctions obey the law of demand? The demand for no. 05-23, Economics Department, University of Central
punishment in the voluntary contribution mechanism. Florida.
Games Econ. Behav. 54, 1–24. (doi:10.1016/j.geb.2004. Bowles, S. 2006 Group competition, reproductive leveling,
08.007) and the evolution of human altruism. Science 314,
Anderson, S. P., Goeree, J. K. & Holt, C. A. 1998 1569–1572. (doi:10.1126/science.1134829)
A theoretical analysis of altruism and decision error in Bowles, S. & Choi, J.-K. 2007 The coevolution of parochial
public goods games. J. Public Econ. 70, 297–323. (doi:10. altruism and war. Science 318, 636–640. (doi:10.1126/
1016/S0047-2727(98)00035-8) science.1149498)
Andreoni, J. 1988 Why free ride—strategies and learning in Bowles, S. & Gintis, H. 2002 Social capital and community
public-goods experiments. J. Public Econ. 37, 291–304. governance. Econ. J. 112, 419–436. (doi:10.1111/1468-
(doi:10.1016/0047-2727(88)90043-6) 0297.00077)
Andreoni, J. 1995 Cooperation in public-goods experi- Boyd, R. & Richerson, P. J. 1985 Culture and the evolutionary
ments—kindness or confusion? Am. Econ. Rev. 85, process. Chicago, IL: University of Chicago Press.
891–904. Boyd, R. & Richerson, P. J. 1988 The evolution of reciprocity
Andreoni, J. & Croson, R. 2008 Partners versus strangers: the in sizable groups. J. Theor. Biol. 132, 337–356. (doi:10.
effect of random rematching in public goods experiments. 1016/S0022-5193(88)80219-4)
In The handbook of experimental economic results (eds Boyd, R., Gintis, H., Bowles, S. & Richerson, P. J. 2003 The
C. Plott & V. Smith), pp. 776–783. Amsterdam, The evolution of altruistic punishment. Proc. Natl Acad. Sci.
Netherlands: Elsevier. USA 100, 3531–3535. (doi:10.1073/pnas.0630443100)
Andreoni, J. & Petrie, R. 2004 Public goods experiments Brandts, J. & Schram, A. 2001 Cooperation and noise in
without confidentiality: a glimpse into fund-raising. public goods experiments: applying the contribution
J. Public Econ. 88, 1605–1623. (doi:10.1016/S0047- function approach. J. Public Econ. 79, 399–427. (doi:10.
2727(03)00040-9) 1016/S0047-2727(99)00120-6)
Austen-Smith, D. & Feddersen, T. J. 2009 Information Brosig, J., Ockenfels, A. & Weimann, J. 2003 The effect of
aggregation and communication in committees. Phil. communication media on cooperation. Ger. Econ. Rev. 4,
Trans. R. Soc. B 364, 763–769. (doi:10.1098/rstb.2008. 217–241. (doi:10.1111/1468-0475.00080)
0256) Burnham, T. C. & Hare, B. 2007 Engineering human
Axelrod, R. & Hamilton, W. 1981 The evolution of cooperation—does involuntary neural activation increase
cooperation. Science 211, 1390–1396. (doi:10.1126/ public goods contributions? Hum. Nat. Interdiscip. Biosoc.
science.7466396) Perspect. 18, 88–108.

Camerer, C. F. 2003 Behavioral game theory. Princeton, NJ: Croson, R. 1996 Partners and strangers revisited. Econ. Lett.
Princeton University Press. 53, 25–32. (doi:10.1016/S0165-1765(97)82136-2)
Camerer, C. F. & Fehr, E. 2004 Measuring social norms Croson, R. 2007 Theories of commitment, altruism and
using experimental games: a guide for social scientists. In reciprocity: evidence from linear public goods games.
Foundations of human sociality: experimental and ethno- Econ. Inq. 45, 199–216.
graphic evidence from 15 small-scale societies (eds J. Henrich, Cubitt, R., Drouvelis, M. & Gächter, S. 2008 Framing and
R. Boyd, S. Bowles, C. F. Camerer, E. Fehr & H. Gintis), free riding: emotional responses and punishment in social
pp. 55–95. New York, NY: Oxford University Press. dilemma games. CeDEx discussion paper no. 2008-02,
Cardenas, J. C. & Jaramillo, C. R. 2007 Cooperation in large University of Nottingham.
networks. An experimental approach. Documento CEDE Dal Bo, P. 2005 Cooperation under the shadow of the future:
2007-06 (Edición Electrónica). Bogotá, Colombia: Uni- experimental evidence from infinitely repeated games.
versidad de los Andes. Am. Econ. Rev. 95, 1591–1604. (doi:10.1257/00028280
Carpenter, J. P. 2004 When in Rome: conformity and the 5775014434)
provision of public goods. J. Socio-Econ. 33, 395–408. Dawes, R. M. 1980 Social dilemmas. Annu. Rev. Psychol. 31,
(doi:10.1016/j.socec.2004.04.009) 169–193. (doi:10.1146/annurev.ps.31.020180.001125)
Carpenter, J. 2007a The demand for punishment. J. Econ. Dawes, R. M., McTavish, J. & Shaklee, H. 1977 Behavior,
Behav. Organ. 62, 522–542. (doi:10.1016/j.jebo.2005. communication, and assumptions about other peoples
05.004) behavior in a commons dilemma situation. J. Pers. Soc.
Carpenter, J. 2007b Punishing free-riders: how group size Psychol. 35, 1–11. (doi:10.1037/0022-3514.35.1.1)
affects mutual monitoring and the provision of public Dawes, C. T., Fowler, J. H., Johnson, T., McElreath, R. &
goods. Games Econ. Behav. 60, 31–51. (doi:10.1016/j.geb. Smirnov, O. 2007 Egalitarian motives in humans. Nature
2006.08.011) 446, 794–796. (doi:10.1038/nature05651)
Carpenter, J. & Matthews, P. 2005 Norm enforcement: de Quervain, D. J. F., Fischbacher, U., Treyer, V.,
anger, indignation or reciprocity? IZA discussion paper Schellhammer, M., Schnyder, U., Buck, A. & Fehr, E.
no. 1583. 2004 The neural basis of altruistic punishment. Science
Carpenter, J. P., Daniere, A. G. & Takahashi, L. M. 2004 305, 1254–1258. (doi:10.1126/science.1100735)
Cooperation, trust, and social capital in Southeast Asian Denant-Boemont, L., Masclet, D. & Noussair, C. N. 2007
urban slums. J. Econ. Behav. Organ. 55, 533–551. (doi:10. Punishment, counterpunishment and sanction enforce-
1016/j.jebo.2003.11.007) ment in a social dilemma experiment. Econ. Theor. 33,
Carpenter, J., Harrison, G. & List, J. (eds) 2005a Field
145–167. (doi:10.1007/s00199-007-0212-0)
experiments in economics. Greenwich, CT: JAI Press.
Devetag, G. & Ortmann, A. 2007 When and why? A critical
Carpenter, J. P., Burks, S. & Verhoogen, E. 2005b
survey on coordination failure in the laboratory. Exp. Econ.
Comparing students to workers: the effect of stakes, social
10, 331–344. (doi:10.1007/s10683-007-9178-9)
framing, and demographics on bargaining outcomes. In
Doebeli, M., Hauert, C. & Killingback, T. 2004 The
Field experiments in economics (eds J. P. Carpenter,
evolutionary origin of cooperators and defectors. Science
G. Harrison & J. List). Research in experimental
306, 859–862. (doi:10.1126/science.1101456)
economics. Amsterdam, The Netherlands: Elsevier JAI.
Dohmen, T., Falk, A., Huffman, D. & Sunde, U. 2008
Carpenter, J., Connolly, C. & Knowles Myers, C. 2008
Representative trust and reciprocity: prevalence and
Altruistic behavior in a representative dictator experiment.
determinants. Econ. Inq. 46, 84–90.
Exp. Econ. 11, 282–298. (doi:10.1007/s10683-007-9193-x)
Dreber, A., Rand, D. G., Fudenberg, D. & Nowak, M. A.
Carpenter, J., Bowles, S., Gintis, H. & Hwang, S.-H. In press.
Strong reciprocity and team production: theory and 2008 Winners don’t punish. Nature 452, 348–351.
evidence. J. Econ. Behav. Organ. (doi:10.1038/nature06723)
Charness, G. & Dufwenberg, M. 2006 Promises and Dufwenberg, M., Gächter, S. & Hennig-Schmidt, H. 2006
partnership. Econometrica 74, 1579–1601. (doi:10.1111/ The framing of games and the psychology of strategic
j.1468-0262.2006.00719.x) choice. CeDEx discussion paper 2006-20, University of
Chaudhuri, A., Graziano, S. & Maitra, P. 2006 Social Nottingham.
learning and norms in a public goods experiment with Dyer, J. R. G., Johansson, A., Helbing, D., Couzin, I. D. &
inter-generational advice. Rev. Econ. Stud. 73, 357–380. Krause, J. 2009 Leadership, consensus decision making
(doi:10.1111/j.1467-937X.2006.0379.x) and collective behaviour in humans. Phil. Trans. R. Soc. B
Cinyabuguma, M., Page, T. & Putterman, L. 2006 Can 364, 781–789. (doi:10.1098/rstb.2008.0233)
second-order punishment deter perverse punishment? Egas, M. & Riedl, A. 2008 The economics of altruistic
Exp. Econ. 9, 265–279. (doi:10.1007/s10683-006-9127-z) punishment and the maintenance of cooperation. Proc. R.
Clutton-Brock, T. H. & Parker, G. A. 1995 Punishment in Soc. B 275, 871–878. (doi:10.1098/rspb.2007.1558)
animal societies. Nature 373, 209–216. (doi:10.1038/ Eldakar, O. T., Farrell, D. L. & Wilson, D. S. 2007 Selfish
373209a0) punishment: altruism can be maintained by competition
Cochard, F., Van, P. N. & Willinger, M. 2004 Trusting among cheaters. J. Theor. Biol. 249, 198–205. (doi:10.
behavior in a repeated investment game. J. Econ. Behav. 1016/j.jtbi.2007.07.024)
Organ. 55, 31–44. (doi:10.1016/j.jebo.2003.07.004) Engelmann, D. & Fischbacher, U. 2002 Indirect reciprocity
Coleman, J. 1990 Foundations of social theory. Cambridge, and strategic reputation building in an experimental
MA: The Belknap Press of Harvard University Press. helping game. IEW Working paper no. 132, University
Colman, A. M. 1999 Game theory and its applications in the of Zurich.
social and biological sciences. London, UK; New York, NY: Falk, A., Gächter, S. & Kovacs, J. 1999 Intrinsic motivation
Routledge. and extrinsic incentives in a repeated game with incom-
Conradt, L. & Roper, T. J. 2009 Conflicts of interest and the plete contracts. J. Econ. Psychol. 20, 251–284. (doi:10.
evolution of decision sharing. Phil. Trans. R. Soc. B 364, 1016/S0167-4870(99)00009-4)
807–819. (doi:10.1098/rstb.2008.0257) Falk, A., Fehr, E. & Fischbacher, U. 2005 Driving forces
Cookson, R. 2000 Framing effects in public goods experi- behind informal sanctions. Econometrica 73, 2017–2030.
ments. Exp. Econ. 3, 55–79. (doi:10.1007/BF01669207) (doi:10.1111/j.1468-0262.2005.00644.x)

Fehr, E. & Camerer, C. 2007 Social neuroeconomics: the Gächter, S. & Thöni, C. 2005 Social learning and
neural circuitry of social preferences. Trends Cogn. Sci. 11, voluntary cooperation among like-minded people. J. Eur.
419–427. (doi:10.1016/j.tics.2007.09.002) Econ. Assoc. 3, 303–314. (doi:10.1162/15424760544
Fehr, E. & Fischbacher, U. 2003 The nature of human 73323)
altruism. Nature 425, 785–791. (doi:10.1038/nature Gächter, S., Herrmann, B. & Thöni, C. 2004 Trust,
02043) voluntary cooperation, and socio-economic background:
Fehr, E. & Gächter, S. 2000 Cooperation and punishment survey and experimental evidence. J. Econ. Behav. Organ.
in public goods experiments. Am. Econ. Rev. 90, 980–994. 55, 505–531. (doi:10.1016/j.jebo.2003.11.006)
Fehr, E. & Gächter, S. 2002 Altruistic punishment in Gächter, S., Renner, E. & Sefton, M. 2008 The long-run
humans. Nature 415, 137–140. (doi:10.1038/415137a) benefits of punishment. Science 322, 1510. (doi:10.1126/
Fehr, E. & Schmidt, K. M. 2006 The economics of fairness, science.1164744)
reciprocity and altruism—experimental evidence and new Gardner, A. & West, S. A. 2004 Cooperation and punish-
theories. In Handbook of the economics of giving, altruism ment, especially in humans. Am. Nat. 164, 753–764.
and reciprocity, vol. 1 (eds S.-C. Kolm & J. M. Ythier), (doi:10.1086/425623)
pp. 615–691. Amsterdam, The Netherlands: Elsevier B.V. Gintis, H. 2000 Strong reciprocity and human sociality.
Fehr, E., Fischbacher, U. & Gächter, S. 2002a Strong J. Theor. Biol. 206, 169–179. (doi:10.1006/jtbi.2000.
reciprocity, human cooperation, and the enforcement of 2111)
social norms. Hum. Nat. Interdiscip. Biosoc. Perspect. 13, Gintis, H. 2008 Punishment and cooperation. Science 319,
1–25. (doi:10.1007/S12110-002-1012-7) 1345–1346. (doi:10.1126/science.1155333)
Fehr, E., Fischbacher, U., von Rosenbladt, B., Schupp, J. & Goeree, J. K., Holt, C. A. & Laury, S. K. 2002 Private costs
Wagner, G. G. 2002b A nationwide laboratory. Examining and public benefits: unraveling the effects of altruism and
trust and trustworthiness by integrating behavioral noisy behavior. J. Public Econ. 83, 255–276. (doi:10.1016/
experiments into representative surveys. Schmoller’s Jahr- S0047-2727(00)00160-2)
buch 122, 519–542. Gouldner, A. W. 1960 The norm of reciprocity: a preliminary
Fessler, D. & Haley, K. J. 2003 The strategy of affect: statement. Am. Sociol. Rev. 25, 161–178. (doi:10.2307/
emotions in human cooperation. In Genetic and cultural 2092623)
evolution of cooperation (ed. P. Hammerstein), pp. 7–36. Guala, F. 2005 The methodology of experimental economics.
Cambridge, MA: The MIT Press. Cambridge, UK: Cambridge University Press.
Fischbacher, U. 2007 z-Tree: Zurich toolbox for readymade Gunnthorsdottir, A., Houser, D. & McCabe, K. 2007
economic experiments. Exp. Econ. 10, 171–178. (doi:10. Disposition, history and contributions in public goods
1007/s10683-006-9159-4) experiments. J. Behav. Organ. 62, 304–315. (doi:10.1016/
Fischbacher, U. & Gächter, S. 2008 Heterogeneous social j.jebo.2005.03.008)
preferences and the dynamics of free riding in public Gürerk, Ö., Irlenbusch, B. & Rockenbach, B. 2006 The
goods. CeDEx discussion paper 2008-07, University of competitive advantage of sanctioning institutions. Science
Nottingham. 312, 108–111. (doi:10.1126/science.1123633)
Fischbacher, U., Gächter, S. & Fehr, E. 2001 Are people Hagen, E. H. & Hammerstein, P. 2006 Game theory and
conditionally cooperative? Evidence from a public goods human evolution: a critique of some recent interpretations
experiment. Econ. Lett. 71, 397–404. (doi:10.1016/S0165- of experimental games. Theor. Popul. Biol. 69, 339–348.
1765(01)00394-9) (doi:10.1016/j.tpb.2005.09.005)
Fliessbach, K., Weber, B., Trautner, P., Dohmen, T., Sunde, Haley, K. J. & Fessler, D. M. T. 2005 Nobody’s watching?
U., Elger, C. E. & Falk, A. 2007 Social comparison affects Subtle cues affect generosity in an anonymous economic
reward-related brain activity in the human ventral game. Evol. Hum. Behav. 26, 245–256. (doi:10.1016/
striatum. Science 318, 1305–1308. (doi:10.1126/science. j.evolhumbehav.2005.01.002)
1145876) Hamilton, W. D. 1964 Genetical evolution of social behavior
Friedman, D. & Sunder, S. 1994 Experimental methods. I, II. J. Theor. Biol. 7, 1–52. (doi:10.1016/0022-5193
A primer for economists. Cambridge, UK: Cambridge (64)90038-4)
University Press. Hammerstein, P. (ed.) 2003 Genetic and cultural evolution of
Fudenberg, D. & Maskin, E. 1986 The folk theorem in cooperation. Cambridge, MA: The MIT Press.
repeated games with discounting or with incomplete Hammerstein, P. & Hagen, E. H. 2005 The second wave of
information. Econometrica 54, 533–556. (doi:10.2307/ evolutionary economics in biology. Trends Ecol. Evol. 20,
1911307) 604–609. (doi:10.1016/j.tree.2005.07.012)
Gächter, S. 2007 Conditional cooperation: behavioral Hardin, G. 1968 The tragedy of the commons. Science 162,
regularities from the lab and the field and their policy 1243–1248. (doi:10.1126/science.162.3859.1243)
implications. In Psychology and economics (eds B. S. Frey & Henrich, J. 2004 Cultural group selection, coevolutionary
A. Stutzer), pp. 19–50. Cambridge, MA: The MIT Press. processes and large-scale cooperation. J. Econ. Behav.
Gächter, S. & Falk, A. 2002 Reputation and reciprocity: Organ. 53, 3–35. (doi:10.1016/S0167-2681(03)00094-5)
consequences for the labour relation. Scand. J. Econ. 104, Henrich, J. & Boyd, R. 2001 Why people punish defectors.
1–26. (doi:10.1111/1467-9442.00269) Weak conformist transmission can stabilize costly enforce-
Gächter, S. & Fehr, E. 1999 Collective action as a social ment of norms in cooperative dilemmas. J. Theor. Biol.
exchange. J. Econ. Behav. Organ. 39, 341–369. (doi:10. 208, 79–89. (doi:10.1006/jtbi.2000.2202)
1016/S0167-2681(99)00045-1) Henrich, J. & Henrich, N. 2007 Why humans cooperate:
Gächter, S. & Herrmann, B. 2005 Human cooperation from a cultural and evolutionary explanation. Evolution and
an economic perspective. In Cooperation in primates and cognition series. Oxford, UK: Oxford University Press.
humans. Mechanisms and evolution (eds P. M. Kappeler & Henrich, J. et al. 2005 ‘Economic man’ in cross-cultural
C. P. van Schaik), pp. 267–289. Heidelberg, Germany: perspective: behavioral experiments in 15 small-scale
Springer. societies. Behav. Brain Sci. 28, 795–815. (doi:10.1017/
Gächter, S. & Herrmann, B. 2007 The limits of self- S0140525X05000142)
governance when cooperators get punished—experimental Henrich, J. et al. 2006 Costly punishment across human
evidence from urban and rural Russia. CeDEx discussion societies. Science 312, 1767–1770. (doi:10.1126/science.
paper no. 2007-11, University of Nottingham. 1127333)

Herrmann, B. & Thöni, C. In press. Measuring conditional Ledyard, J. O. 1995 Public goods: a survey of experimental
cooperation: a replication study in Russia. Exp. Econ. research. In The handbook of experimental economics (eds
(doi:10.1007/s10683-008-9197-1) A. E. Roth & J. H. Kagel), pp. 111–181. Princeton, NJ:
Herrmann, B., Thöni, C. & Gächter, S. 2008 Antisocial Princeton University Press.
punishment across societies. Science 319, 1362–1367. Lehmann, L. & Keller, L. 2006 The evolution of cooperation
(doi:10.1126/science.1153808) and altruism: a general framework and a classification
Hix, S., Noury, A. & Roland, G. 2009 Voting patterns and of models. J. Evol. Biol. 19, 1365–1376. (doi:10.1111/
alliance formation in the European Parliament. Phil. j.1420-9101.2006.01119.x)
Trans. R. Soc. B 364, 821–831. (doi:10.1098/rstb.2008. Liebrand, W. B. G., Jansen, R. W. T. L., Rijken, V. M. &
0263) Suhre, C. J. M. 1986 Might over morality: social values
Hofstede, G. 2001 Culture’s consequences: comparing values, and the perception of other players in experimental games.
behaviors, institutions, and organizations across nations. J. Exp. Soc. Psychol. 22, 203–215. (doi:10.1016/0022-
Thousand Oaks, CA: Sage. 1031(86)90024-7)
Holm, H. J. & Nystedt, P. 2005 Intra-generational trust—a Loewenstein, G., Thompson, L. & Bazerman, M. 1989
semi-experimental study of trust among different gener- Social utility and decision making in interpersonal
ations. J. Econ. Behav. Organ. 58, 403–419. (doi:10.1016/ contexts. J. Pers. Soc. Psychol. 57, 426–441. (doi:10.
j.jebo.2003.10.013) 1037/0022-3514.57.3.426)
Hopfensitz, A. & Reuben, E. In press. The importance of Madsen, E. A., Tunney, R. J., Fieldman, G., Plotkin, H. C.,
emotions for the effectiveness of social punishment. Dunbar, R. I. M., Richardson, J. M. & McFarland, D.
Econ. J. 2007 Kinship and altruism: a cross-cultural experimental
Houser, D., Xiao, E., McCabe, K. & Smith, V. 2008 When study. Br. J. Psychol. 98, 339–359. (doi:10.1348/
punishment fails: research on sanctions, intentions and 000712606X129213)
non-cooperation. Games Econ. Behav. 62, 509–532. Marlowe, F. W. et al. 2008 More ‘altruistic’ punishment in
(doi:10.1016/j.geb.2007.05.001) larger societies. Proc. R. Soc. B 275, 587–590. (doi:10.
Inglehart, R. & Baker, W. E. 2000 Modernization, cultural 1098/rspb.2007.1517)
change, and the persistence of traditional values. Am. Marwell, G. & Ames, R. 1979 Experiments on the provision
Sociol. Rev. 65, 19–51. (doi:10.2307/2657288) of public goods I: resources, interest, group size, and the
Isaac, R. M. & Walker, J. M. 1988a Communication and free- free-rider problem. Am. J. Sociol. 84, 1335–1360. (doi:10.
riding behavior—the voluntary contribution mechanism. 1086/226937)
Econ. Inq. 26, 585–608. Masclet, D. & Villeval, M. C. 2008 Punishment, inequality,
and welfare: a public good experiment. Soc. Choice Welfare
Isaac, R. M. & Walker, J. M. 1988b Group size effects in
31, 475–502. (doi:10.1007/s00355-007-0291-7)
public goods provision: the voluntary contributions
Masclet, D., Noussair, C., Tucker, S. & Villeval, M. C. 2003
mechanism. Q. J. Econ. 103, 179–199. (doi:10.2307/
Monetary and nonmonetary punishment in the voluntary
1882648)
contributions mechanism. Am. Econ. Rev. 93, 366–380.
Isaac, R. M., Walker, J. M. & Williams, A. W. 1994 Group-
Milinski, M. & Rockenbach, B. 2007 Spying on others
size and the voluntary provision of public goods—
evolves. Science 317, 464–465. (doi:10.1126/science.
experimental evidence utilizing large groups. J. Public
1143918)
Econ. 54, 1–36. (doi:10.1016/0047-2727(94)90068-X)
Milinski, M., Semmann, D. & Krambeck, H. J. 2002
Kagel, J. & Roth, A. E. 1995 The handbook of experimental
Reputation helps solve the ‘tragedy of the commons’.
economics. Princeton, NJ: Princeton University Press.
Nature 415, 424–426. (doi:10.1038/415424a)
Kelley, H. & Stahelski, A. 1970 Social interaction basis of Mohan, N. H. 2008 Vengeance. NBER working paper no.
cooperators’ and competitors’ beliefs about others. J. Pers. 14131.
Soc. Psychol. 16, 190–219. Monin, B. 2007 Holier than me? Threatening social
Keser, C. & van Winden, F. 2000 Conditional cooperation comparison in the moral domain. Int. Rev. Soc. Psychol.
and voluntary contributions to public goods. Scand. 20, 53–68.
J. Econ. 102, 23–39. (doi:10.1111/1467-9442.00182) Muller, L., Sefton, M., Steinberg, R. & Vesterlund, L. 2008
Knoch, D., Pascual-Leone, A., Meyer, K., Treyer, V. & Fehr, Strategic behavior and learning in repeated voluntary-
E. 2006 Diminishing reciprocal fairness by disrupting the contribution experiments. J. Econ. Behav. Organ. 67,
right prefrontal cortex. Science 314, 829–832. (doi:10. 782–793. (doi:10.1016/j.jebo.2007.09.001)
1126/science.1129156) Neugebauer, T., Perote, J., Schmidt, U. & Loos, M. In press.
Kocher, M. G., Cherry, T., Kroll, S., Netzer, R. J. & Sutter, Self-biased conditional cooperation: on the decline
M. 2008 Conditional cooperation on three continents. of cooperation in repeated public goods experiments.
Econ. Lett. 101, 175–178. (doi:10.1016/j.econlet.2008. J. Econ. Psychol. (doi:10.1016/j.joep.2008.04.005)
07.015) Nikiforakis, N. 2008 Punishment and counter-punishment in
Kollock, P. 1998 Social dilemmas: the anatomy of co- public good games: can we really govern ourselves?
operation. Annu. Rev. Sociol. 24, 183–214. (doi:10.1146/ J. Public Econ. 92, 91–112. (doi:10.1016/j.jpubeco.2007.
annurev.soc.24.1.183) 04.008)
Kreps, D., Milgrom, P., Roberts, J. & Wilson, R. 1982 Nikiforakis, N. & Normann, H. 2008 A comparative statics
Rational cooperation in the finitely repeated prisoners’ analysis of punishment in public goods experiments. Exp.
dilemma. J. Econ. Theor. 27, 245–252. (doi:10.1016/0022- Econ. 11, 358–369. (doi:10.1007/s10683-007-9171-3)
0531(82)90029-1) Noussair, C. & Tucker, S. 2005 Combining monetary and
Krupp, D. B., Debruine, L. M. & Barclay, P. 2008 A cue of social sanctions to promote cooperation. Econ. Inq. 43,
kinship promotes cooperation for the public good. Evol. 649–660. (doi:10.1093/ei/cbi045)
Hum. Behav. 29, 49–55. (doi:10.1016/j.evolhumbehav. Nowak, M. A. 2006 Five rules for the evolution of coope-
2007.08.002) ration. Science 314, 1560–1563. (doi:10.1126/science.
Kurzban, R. & Houser, D. 2005 Experiments investigating 1133755)
cooperative types in humans: a complement to evolution- Nowak, M. A. & Sigmund, K. 2005 Evolution of indirect
ary theory and simulations. Proc. Natl Acad. Sci. USA 102, reciprocity. Nature 437, 1291–1298. (doi:10.1038/nature
1803–1807. (doi:10.1073/pnas.0408759102) 04131)

Olson, M. 1965 The logic of collective action. Cambridge, MA: Semmann, D., Krambeck, H. J. & Milinski, M. 2005
Harvard University Press. Reputation is valuable within and outside one’s own social
Ones, U. & Putterman, L. 2007 The ecology of collective group. Behav. Ecol. Sociobiol. 57, 611–616. (doi:10.1007/
action: a public goods and sanctions experiment with s00265-004-0885-3)
controlled group formation. J. Econ. Behav. Organ. 62, Seymour, B., Singer, T. & Dolan, R. 2007 The neurobiology
495–521. (doi:10.1016/j.jebo.2005.04.018) of punishment. Nat. Rev. Neurosci. 8, 300–311. (doi:10.
Oosterbeek, H., Sloof, R. & van de Kuilen, G. 2004 Cultural 1038/nrn2119)
differences in ultimatum game experiments: evidence from Shinada, M. & Yamagishi, T. 2007 Punishing free riders:
a meta-analysis. Exp. Econ. 7, 171–188. (doi:10.1023/ direct and indirect promotion of cooperation. Evol. Hum.
B:EXEC.0000026978.14316.74) Behav. 28, 330–339. (doi:10.1016/j.evolhumbehav.2007.
Ostrom, E. 1998 A behavioral approach to the rational choice 04.001)
theory of collective action. Am. Polit. Sci. Rev. 92, 1–22. Sigmund, K. 2007 Punish or perish? Retaliation and
(doi:10.2307/2585925) collaboration among humans. Trends Ecol. Evol. 22,
Ostrom, E., Walker, J. M. & Gardner, R. 1992 Covenants 593–600. (doi:10.1016/j.tree.2007.06.012)
with and without a sword—self-governance is possible. Skyrms, B. 2009 Evolution of signalling systems with
Am. Polit. Sci. Rev. 86, 404–417. (doi:10.2307/1964229) multiple senders and receivers. Phil. Trans. R. Soc. B
Page, T., Putterman, L. & Unel, B. 2005 Voluntary 364, 771–779. (doi:10.1098/rstb.2008.0258)
association in public goods experiments: reciprocity, Sober, E. & Wilson, D. S. 1998 Unto others. The evolution and
mimicry, and efficiency. Econ. J. 115, 1032–1052. psychology of unselfish behavior. Cambridge, MA: Harvard
(doi:10.1111/j.1468-0297.2005.01031.x) University Press.
Palfrey, T. R. 2008 Laboratory experiments. In The Oxford Sonnemans, J., Schram, A. & Offerman, T. 1999 Strategic
handbook of political economy (eds B. Weingast & behavior in public good games: when partners drift apart.
D. Wittman), pp. 915–936. Oxford, UK: Oxford Econ. Lett. 62, 35–41. (doi:10.1016/S0165-1765(98)
University Press. 00203-1)
Palfrey, T. R. & Prisbrey, J. E. 1997 Anomalous behavior in Sutter, M. 2007 Outcomes versus intentions: on the nature
public goods experiments: how much and why? Am. Econ. of fair behavior and its development with age.
Rev. 87, 829–846. J. Econ. Psychol. 28, 69–78. (doi:10.1016/j.joep.2006.09.
Panchanathan, K. & Boyd, R. 2004 Indirect reciprocity 001)
can stabilize cooperation without the second-order free Sutter, M. & Kocher, M. G. 2007 Trust and trustworthiness
rider problem. Nature 432, 499–502. (doi:10.1038/ across different age groups. Games Econ. Behav. 59,
nature02978) 364–382. (doi:10.1016/j.geb.2006.07.006)
Pillutla, M. & Murnighan, K. J. 1996 Unfairness, anger, and Sutter, M., Haigner, S. & Kocher, M. G. 2008 Choosing the
spite: emotional rejections of ultimatum offers. Organ. carrot or the stick?—endogenous institutional choice in
Behav. Hum. Decis. Process. 68, 208–224. (doi:10.1006/ social dilemma situations. Working paper in economics
obhd.1996.0100) and statistics 2008-07, University of Innsbruck.
Rabin, M. 1993 Incorporating fairness into game-theory and Tomasello, M., Carpenter, M., Call, J., Behne, T. & Moll, H.
economics. Am. Econ. Rev. 83, 1281–1302. 2005 Understanding and sharing intentions: the origins of
Rapoport, A. & Chammah, A. M. 1965 Prisoners’ dilemma. cultural cognition. Behav. Brain Sci. 28, 675–691. (doi:10.
A study in conflict and cooperation. Ann Arbor, MI: The 1017/S0140525X05000129)
University of Michigan Press. Trivers, R. 1971 The evolution of reciprocal altruism. Q. Rev.
Rege, M. & Telle, K. 2004 The impact of social approval Biol. 46, 35–57. (doi:10.1086/406755)
and framing on cooperation in public good situations. van Prooijen, J. W., Gallucci, M. & Toeset, G. 2008
J. Public Econ. 88, 1625–1644. (doi:10.1016/S0047-2727 Procedural justice in punishment systems: inconsistent
(03)00021-5) punishment procedures have detrimental effects on
Reuben, E. & Riedl, A. In press. Public goods provision and cooperation. Br. J. Soc. Psychol. 47, 311–324. (doi:10.
sanctioning in privileged groups. J. Conflict Resolut. 1348/014466607X218212)
(doi:10.1177/0022002708322361) Walker, J. M. & Halloran, M. A. 2004 Rewards and sanctions
Reuben, E. & van Winden, F. 2008 Social ties and coordi- and the provision of public goods in one-shot settings. Exp.
nation on negative reciprocity: the role of affect. J. Public Econ. 7, 235–247. (doi:10.1023/B:EXEC.0000040559.
Econ. 92, 34–53. (doi:10.1016/j.jpubeco.2007.04.012) 08652.51)
Rockenbach, B. & Milinski, M. 2006 The efficient interaction Weimann, J. 1994 Individual behavior in a free riding
of indirect reciprocity and costly punishment. Nature 444, experiment. J. Public Econ. 54, 185–200. (doi:10.1016/
718–723. (doi:10.1038/nature05229) 0047-2727(94)90059-0)
Sally, D. 1995 Conversation and cooperation in social West, S. A., Griffin, A. S. & Gardner, A. 2007 Evolutionary
dilemmas: a meta-analysis of experiments from 1958 to explanations for cooperation. Curr. Biol. 17, R661–R672.
1992. Ration. Soc. 7, 58–92. (doi:10.1177/1043463195 (doi:10.1016/j.cub.2007.06.004)
007001004) Wiessner, P. 2005 Norm enforcement among the Ju/’hoansi
Sanfey, A. G., Rilling, J. K., Aronson, J. A., Nystrom, L. E. & Bushmen—a case of strong reciprocity? Hum. Nat.
Cohen, J. D. 2003 The neural basis of economic decision- Interdiscip. Biosoc. Perspect. 16, 115–145.
making in the ultimatum game. Science 300, 1755–1758. Yamagishi, T. 1986 The provision of a sanctioning system as a
(doi:10.1126/science.1082976) public good. J. Pers. Soc. Psychol. 51, 110–116. (doi:10.
Sefton, M., Shupp, R. & Walker, J. M. 2007 The effect of 1037/0022-3514.51.1.110)
rewards and sanctions in provision of public goods. Econ. Zelmer, J. 2003 Linear public goods experiments: a meta-
Inq. 45, 671–690. analysis. Exp. Econ. 6, 299–310. (doi:10.1023/A:1026
Seinen, I. & Schram, A. 2006 Social status and group 277420119)
norms: indirect reciprocity in a repeated helping experi- Zizzo, D. J. 2003 Money burning and rank egalitarianism
ment. Eur. Econ. Rev. 50, 581–602. (doi:10.1016/ with random dictators. Econ. Lett. 81, 263–266. (doi:10.
j.euroecorev.2004.10.005) 1016/S0165-1765(03)00190-3)

Phil. Trans. R. Soc. B (2009) 364, 807–819

doi:10.1098/rstb.2008.0257
Conflicts of interest and the evolution of

decision sharing
Larissa Conradt* and Timothy J. Roper
Department of Biology & Environmental Sciences, University of Sussex, JMS Building, Falmer,
Brighton BN1 9QR, UK
Social animals regularly face consensus decisions whereby they choose, collectively, between mutually
exclusive actions. Such decisions often involve conflicts of interest between group members with respect
to preferred action. Conflicts could, in principle, be resolved, either by sharing decisions between
members (‘shared decisions’) or by one ‘dominant’ member making decisions on behalf of the whole
group (‘unshared decisions’). Both, shared and unshared decisions, have been observed. However, it is
unclear as to what favours the evolution of either decision type. Here, after a brief literature review, we
present a novel method, involving a combination of self-organizing system and game theory modelling,
of investigating the evolution of shared and unshared decisions. We apply the method to decisions on
movement direction. We find that both, shared and unshared, decisions can evolve without individuals
having a global overview of the group’s behaviour or any knowledge about other members’ preferences
or intentions. Selection favours unshared over shared decisions when conflicts are high relative to
grouping benefits, and vice versa. These results differ from those of group decision models relating to
activity timings. We attribute this to fundamental differences between collective decisions about
modalities that are disjunct (here, space) or continuous (here, time) with respect to costs/benefits.
Keywords: collective decisions; conflict resolution; consensus decisions; cooperation;
information pooling; speed versus accuracy trade-off
1. GENERAL INTRODUCTION constraints and the resultant trade-offs (e.g. between

A major challenge to most social animals, including speed and accuracy of decision making) limit the
humans, is the regular need to make ‘consensus options for potential decision outcomes (Passino &
decisions’, i.e. to choose, collectively, between mutually Seeley 2006; Franks et al. 2009; Sumpter & Pratt
exclusive actions (Conradt 1998). Such decisions may 2009). Here, we concentrate on the second of these
involve choosing between different activities (Conradt three factors, namely, conflicts of interest between
1998; Ruckstuhl & Neuhaus 2000), movement desti- group members.
nations (Trillmich et al. 2004; Couzin et al. 2005), nest To reach a consensus, conflicts of interest with
sites (Seeley & Buhrman 1999), migration routes (Simons respect to decision outcomes have to be resolved
2004) or hunting strategies (Conradt & Roper 2005). (Conradt & Roper 2003, 2007; Rands et al. 2003;
In principle, three main factors influence consensus Couzin et al. 2005). In principle, this can be achieved in
decisions, namely, information, conflict and time two ways: decisions can either be ‘shared’ (some or all
constraints. Firstly, the relevant information held by group members contribute to the decision outcome)
individual group members, and how this information is or ‘unshared’ (the decision outcome is dictated by a
shared within the group, influences decision outcomes single individual; Conradt & Roper 2003; List 2004;
and accuracy (Seeley & Buhrman 1999; Conradt & Austen-Smith & Feddersen 2009). Shared and
Roper 2003; List 2004; Simons 2004; Couzin et al. unshared decisions influence the fitness of individual
2005; Ame et al. 2006; Biro et al. 2006; Passino & group members differently. Here, we are interested in
Seeley 2006; Codling et al. 2007; Lusseau 2007; Ward the conditions under which shared decision making
et al. 2008; Dyer et al. 2009; Franks et al. 2009; List might be expected to evolve when there are conflicts of
et al. 2009; Skyrms 2009; Sumpter & Pratt 2009). interest between group members. We provide a brief
Secondly, conflicts of interest between group members, overview of published work on conflicts of interest and
in the form of conflicting preferences for different the evolution of decision sharing in animals and then
decision outcomes, affect the fitness of individual group investigate the evolution of decision sharing using a
members and the evolution of the decision-making novel approach in which we integrate self-organizing
process (Conradt 1998; Ruckstuhl 1998, 1999; Con- systems simulations with game theory modelling.
radt & Roper 2000, 2003, 2005, 2007; Ruckstuhl &
Neuhaus 2000, 2002; Rands et al. 2003; Austen-Smith
& Feddersen 2009; Hix & Noury 2009). Thirdly, time 2. SHORT REVIEW: CONFLICTS OF INTEREST
AND THE EVOLUTION OF DECISION SHARING
* Author for correspondence (l.conradt@sussex.ac.uk). Consensus decisions in animals vary in the extent to
One contribution of 11 to a Theme Issue ‘Group decision making in which they involve conflicts of interest between group
humans and animals’. members (Conradt & Roper 2005). For example, nest

808 L. Conradt & T. J. Roper Conflicts of interest
site choice in eusocial insects (Seeley & Buhrman 1999; (primates: Stolba 1979; Stewart & Harcourt 1994;
Visscher & Seeley 2007; Franks et al. 2009), or Boinski & Campbell 1995; Byrne 2000; Milton 2000;
navigation in homing or migrating birds ( Wallraff Meunier et al. 2006; Sellers et al. 2007; ungulates:
1978; Simons 2004; Biro et al. 2006), involve relatively Conradt & Roper 2003; bats: Kerth et al. 2006; birds:
little conflict of interest because the optimal decision Black 1988). Conversely, there is no intrinsic reason
outcome is the same for all group members. By why self-organizing systems should only produce equal
contrast, synchronization of group activities and choice sharing of decisions (Conradt et al. in press). Indeed,
of short-term travel destinations in birds, fishes and several empirical observations suggest that in self-
mammals can involve significant conflict (Conradt organizing groups individuals that have greater ‘needs’
1998; Ruckstuhl & Neuhaus 2000), since individual might be able to gain a disproportionate weight in
group members often differ in physiological require- decision outcomes (fishes: Krause et al. 1992; Krause
ments (e.g. Clutton-Brock et al. 1982) and therefore in 1993; ungulates: Prins 1996; Fischhoff et al. 2007).
their optimal requirements with respect to activity One of the most important outstanding theoretical
timing and travel destinations (e.g. Gompper 1996; questions is to explain how shared decision making can
Prins 1996; Ruckstuhl 1998). evolve in decisions that involve conflict, without
Where there is conflict, reaching a consensus invoking ‘good-for-the-group’ arguments (Gächter &
decision will necessarily involve compromise for at Herrmann 2009). The main challenge in addressing
least some group members. Formally, this means that this question is the complexity arising from multi-
at least some members will pay a ‘consensus cost’, individual interactions in group decisions. Rands et al.
defined as the fitness cost to an individual of foregoing (2003) modelled a group of two in which each member
its own optimal action in order to comply with the tries to maximize its own chance of survival and both
consensus decision (Conradt & Roper 2003, 2005, gain from foraging together, but in which the optimal
2007). Consensus costs, and their distribution times for foraging can differ between individuals. In
between group members, are crucial for group stability such a situation, if one member starts to forage because
(Conradt 1998; Ruckstuhl & Neuhaus 2000). its reserves drop below a crucial threshold, the other
The consensus cost to individual group members member should also start to forage, because the
depends on their own ability to influence the advantages of foraging together outweigh the dis-
decision outcome versus that of other group members advantages of foraging at a less preferred time. The
(Conradt & Roper 2003). For example, if a dominant result is high activity synchrony. This constitutes, in
individual dictates the decision (unshared decision principle, a shared consensus decision, because neither
making), it will not incur a cost because it can choose of the members has intrinsically more weight in the
the decision outcome that is optimal for itself, whereas decision than the other, or differs in principle from the
subordinate members will incur costs according to how other at the start of the simulations. The decision
different their own optima are from that of the threshold in this shared decision is one since, if at least
dominant. By contrast, if all group members contribute one of the two starts to forage, this determines the
equally to the decision outcome (‘equally shared’ collective behaviour. However, because group members
decision making) the consensus costs will be relatively differ in reserves, and this difference is likely to be
lower for most members other than the dominant preserved for some time, it is likely that during several
(Conradt & Roper 2003). Decision-making processes consecutive occasions it will be the same individual that
that are intermediate between these two extremes starts to forage first. Thus, this individual will seem to
(i.e. ‘partially shared’ decisions in which a subset of be the apparent ‘leader’ of the activity. It is not clear
group members contributes to the decision outcome) what would happen if one were to extend Rands et al.’s
are also possible. model to a group of more than two.
Historically, scientists working on relatively small Dostalkova & Spinka (2007) also modelled activity
groups (i.e. groups in which global communication synchronization between two group members. Similar
between group members is possible; Conradt & Roper to Rands et al. (2003), they predict that individuals
2005), especially in mammals, have tended to assume maximize their benefits by switching activities simul-
that decision making is unshared, no doubt because taneously when the first prefers a change, provided that
such groups usually exhibit obvious dominance hier- there is no communication between members.
archies (e.g. Sapolsky 2005). By contrast, scientists However, if there is communication, and acting too
working on large groups (i.e. groups in which only local late is ‘cheaper’ (in terms of consensus cost) than acting
communication between neighbouring members is prematurely, the group should switch activity only
possible; Conradt & Roper 2005), especially those when both members prefer to do so (i.e. when a
using self-organizing system approaches, have tended decision threshold of two is reached). In order to
to assume equally shared decision making insofar as explain this difference, Dostalkova & Spinka (2007)
they apply similar behavioural parameters to all group raised the interesting notion of the inherent informa-
members (see Couzin & Krause 2003 for a review). tional asymmetry of time. Without communication,
The reason for this is no doubt the intrinsic difficulty of information about members’ preferred leaving times
imagining that any individual member of a large self- becomes available only once members attempt to leave,
organizing group could exert meaningful control over resulting in an informational bias towards earlier
the behaviour of the group as a whole. preferred leaving times. This can affect synchronization
However, emerging empirical evidence challenges decisions in an asymmetric manner. Again, it is not
the assumption that a dominant individual always clear what will happen if the model is extended to a
dictates what happens in small groups of animals group larger than two.

Conflicts of interest L. Conradt & T. J. Roper 809
A model by Conradt & Roper (2003), again focusing without reference to the global pattern (i.e. nobody has
on activity synchronization, suggests that the consensus a global ‘overview’ of the group; Camazine et al. 2003).
costs to a group as a whole are considerably lower for Typical examples for self-organizing groups are swarms
equally shared than for unshared decisions under most of honeybees or locusts, colonies of ants, large shoals of
conditions and for group sizes larger than two. fish, large flocks of birds and large herds of ungulates.
However, without invoking group selection arguments Here, we are investigating the evolution of shared and
this does not necessarily lead to the evolution of unshared decision making in such self-organizing
decision sharing. Indeed, Conradt & Roper (2007) groups during group travel.
showed that, in principle, unshared decisions, as well as Studying the evolution of shared decision making in
equally shared decisions, can be evolutionarily stable in self-organizing systems poses some logistic challenges.
groups of all sizes. The authors identified five factors Self-organizing animal systems are often studied with
that favour the evolution of shared versus unshared the help of computer simulations. However, these
decisions in the context of activity synchronization, simulations can require considerable computing power
namely, (i) large conflicts of interest between group (e.g. Couzin et al. 2005). Linking self-organizing
members (resulting in high potential consensus costs), systems modelling directly with evolutionary simula-
(ii) long-term differences between group members with tions, in order to answer evolutionary questions, would
respect to requirements (i.e. high heterogeneity within therefore require prohibitively large computing
the group), (iii) asymmetric consensus costs (i.e. one resources. We avoid this problem by adopting a novel
decision outcome is potentially much more costly than approach in which we combine two different modelling
another), (iv) marginal grouping benefits, and (v) a techniques, namely, self-organizing system modelling
group size that is close to, or above, optimal group size. and game theory modelling.
The predictions of all three of these models still First, we modify a self-organizing system model by
require empirical testing. Moreover, all three models Couzin et al. (2005) in order to enable individual group
deal with activity synchronization, leaving it unclear as members to differ in their local behaviours. We then use
to whether the same results apply to the evolution of self-organizing system simulations to determine global
consensus decisions involving, say, movement desti- group behaviour (e.g. group spatial position, group
nations. Decisions about the latter often differ in movement efficiency, group fragmentation risk)
principle from those about activity synchronization in depending on all members’ local behaviours (cf.
an important respect, as follows. While an average or Conradt et al. in press). Next, we translate this group-
median timing might often offer a sensible compromise level outcome into fitness consequences for individual
between different optimal timings (Conradt & Roper group members (Conradt & Roper 2003, 2007). In this
2003), such a procedure would more rarely be valid manner, we determine fitness consequences for each
when compromising between different spatial targets individual group member, depending both on its own
(Conradt et al. in press). For example, if a flock of geese local behaviour and on that of all other group members.
were trying to decide which of two lakes to land on, it The resulting pay-off matrices to individuals can then
would make no sense for it to land halfway between the be used in multiplayer game theory models to
two lakes where there was no water at all. Another determine evolutionarily stable strategies (‘ESSs’;
limitation of all three models is that they assume (at least Maynard-Smith 1989). Because multiplayer game
implicitly) that group members have a global overview of theory models are complex, we use computer simula-
the activities of other group members, and that this tions to determine the relevant ESSs. Here, we use this
global overview can influence their behaviour. This method for the first time to investigate the evolution of
assumption does not hold in many of the large groups decisions about movement destinations in a self-
that we observe in nature (Couzin & Krause 2003; organizing group containing three members.
Couzin et al. 2005; Sumpter 2006; Sumpter & Pratt
2009). In §3, we address the evolution of consensus (b) Methods
decisions about movement destinations. We avoid the (i) Conflicts of interest
assumption of global overview by considering a self- To create conflicts of interest between members, we
organizing system. assume that two of the three members (‘majority
members’) prefer to move to one spatial target, while
the third member (‘minority member’) prefers to move
3. EVOLUTION OF DECISION SHARING DURING to another spatial target. The spatial targets are
MOVEMENTS BY SELF-ORGANIZING GROUPS equidistant (500 spatial units) from the group’s starting
(a) Introduction point, but lie in opposite directions. ‘type 1’
Many animal groups, in particular many economically animals prefer to move to target 1 and ‘type 2’ animals
important groups, are thought to be ‘self-organizing’ to target 2. Thus, a group consists either of two type 1
(see Couzin & Krause 2003 and Sumpter 2006 for and one type 2 animal, or vice versa.
reviews), in the sense that (i) pattern at the global level
(e.g. group movement direction, group movement (ii) Self-organizing group movement simulation model
speed, group composition, group formation, etc.) The model is based on one by Couzin et al. (2005). The
emerges solely from numerous local interactions mathematical details are given in appendix Aa(i). All
among the individual group members (i.e. nobody is members move simultaneously in discrete time steps.
in overall ‘command’) and (ii) the local behavioural Each individual member moves at each time step
rules specifying the interactions among the group according to behavioural rules that depend only on
members are executed using only local information, local information (‘local movement rules’), as follows.

Table 1. Standardized relative pay-off matrixa to focal animal depending on its own and the other group member’s assertiveness,
and on majority/minority relationships within the group.
(a) focal player is minority type in group
one majority player plays
ulow umedium uhigh
other majority player plays
focal animal
plays ulow umedium uhigh umedium uhigh uhigh
ulow K0.6 K1 K1 K1 K1 K1
umedium K0.1 K0.7 K1 K1 K1 K1
uhigh 0 0 0.4$B20 KB30 K0.2$B30 KB30 KB30
(b) focal player is majority type in group
other majority player plays
ulow umedium uhigh
minority player plays
focal animal
plays ulow umedium uhigh ulow umedium uhigh ulow umedium uhigh
ulow K0.4 K0.9 K1 0 K0.4 K1 0 0 K0.4CB20 KB30

umedium K0.1 K0.4 K1 0 0 0.2$(B20 KB30 )K1 0 0 B20 KB30
uhigh 0 0 0.6$B20 KB30 0 0 B20 KB30 0 0 B20 KB30
a
Note that if all animals within a group preferred the same target, the group always remained cohesive and moved towards the common
preferred target; thus, for the evolution of behaviours, only those groups were relevant in which there was conflicts of interest between group
members, and only those combinations are shown in the table.
The highest priority for an individual at any time is to spatial range, the individual has, at least for that time
avoid collision by moving away from neighbours that step, lost contact with other group members and
come too close. If (and only if ) there is no immediate cannot move towards, or align with, neighbours.
danger of collision, an individual has the following two Thus, it simply moves towards its own preferred spatial
aims: (i) to keep cohesion with other group members target (see above). According to these local behavioural
and (ii) to move towards its own preferred spatial rules, at each time step, for each individual, the next
target, as follows. In order to maintain group cohesion, movement step direction was calculated (constrained
the individual is attracted to neighbours within a given by a maximum turning angle and including a random
spatial range, so as to move towards them and align movement error; see appendix Aa(ii) for details), and
travel direction with them (‘social attraction’). then all individuals simultaneously moved in their new
However, the resulting ‘social attraction direction’ direction with a constant speed.
might differ from the direction towards the individual’s
preferred spatial target. Therefore, the individual
(iii) Variation in individual behaviours
‘balances’ attraction to its preferred spatial target
The degree of assertiveness ui determines an individual’s
against social attraction to neighbours with its own
readiness to compromise (see above). Therefore, it is the
individual factor ui (‘degree of assertiveness’). In other
words, it decides to move in the following direction: behavioural parameter that is most likely to influence
the sharing/non-sharing of consensus decisions and the
movement direction Z social attraction direction
fitness of individual group members (Conradt et al.
C ui $preferred target direction: ð3:1Þ in press). Hence, we varied ui between individual
It follows that, if ui is large (and the individual is highly members, such that they could exhibit low (ulowZ
assertive), the individual decides to move at each time 0.05), medium (umediumZ0.5) or high (uhighZ5.0;
step predominantly in the direction of its preferred see Couzin et al. 2005 for a biologically meaningful
target; if ui is small (and the individual is of low range of values for ui) assertiveness. We consider all
assertiveness), it decides to move predominantly possible combinations u1, u2, u3 (with u1, u2, u32{ulow,
towards, and align with, neighbours; and if ui is umedium, uhigh}) for three members (i.e. 18 biologically
intermediate (medium assertiveness), the individual different sets of behaviours; see also table 1). For each
compromises between its own target preference and set of behaviours, we ran 500 simulations since, in pilot
moving towards, and aligning with, neighbours. Should runs, results had stabilized to the required accuracy well
the individual have no neighbours within the given below this number of replicates.

(iv) Results of self-organizing system simulations benefits to individuals in each combination of

At the end of every simulation (after 5000 time steps), behaviours in terms of grouping benefits relative to
we determined for each group member (i) with how potential consensus costs. That is, B30 and B20 were the
many other group members it was still in a group (see grouping benefits of remaining in a group with three
Couzin et al. 2005 for criteria of group cohesion) and and two members, respectively, relative to the potential
(ii) its end position. If the individual had moved more consensus costs (i.e. the difference in benefits between
than 500 steps towards target 1 or 2, respectively, it was moving to preferred versus non-preferred spatial
considered to head towards target 1 or 2, respectively. target). Based on these two parameters, we could
Otherwise, it was considered to move too inefficiently calculate the standardized relative pay-off matrix for
(i.e. less than 10% movement efficiency) to be heading majority and minority group members depending on
towards either target. their own behaviour and that of other group members
(table 1). This pay-off matrix is the basis for the game
(v) Translating the results into individual fitness and theory model that follows below.
pay-off matrices
We determined for each group member at the end of
each simulation (i) the remaining number of members (vi) Game theory model
in their group and (ii) which target they were moving to We make the conservative assumption that individuals
(see above). This information can be translated into in self-organizing groups do not have information about
individual benefits, as follows. Firstly, social animals whether they are majority or minority types (Couzin
usually benefit by being in a group, and these ‘grouping et al. 2005). Thus, individuals behave independently of
benefits’ depend on group size (Krause & Ruxton whether they are in a majority or minority. A focal
2002). We can, thus, determine the grouping benefits individual can play the pure strategies ulow, umedium or
to an individual at the end of a simulation, depending uhigh. It can also play any combination of these pure
on how many other group members it was still together strategies. Assume that rlow, rmedium and rhigh are the
with (solitary individuals gain no grouping benefits). probabilities that the focal plays ulow, umedium or uhigh,
Secondly, individuals are likely to gain from moving respectively. All possible complex behaviour strategies
towards their preferred spatial target and from moving for a focal can, thus, be described as {rlow, rmedium, rhigh}
towards the target preferred by other group members, (with 0%rlow%1; 0%rmedium%1Krlow; and rhighZ
and more so from the former than the latter (e.g. 1KrlowKrmedium). Individuals drawn randomly from
Conradt 1998; Ruckstuhl & Neuhaus 2000; Conradt & the population can also play the pure strategies ulow,
Roper 2003, 2007). The difference in fitness gains umedium or uhigh, or complex combinations. Assume that
between moving towards a non-preferred target versus plow, pmedium, phigh and qlow, qmedium, qhigh are the
a preferred target constitutes a ‘consensus cost’ respective probabilities with which individuals that are
(Conradt & Roper 2003, 2005). Therefore, from the drawn randomly from the population, and that have the
final position of individuals, we could determine what same ( pxx) or opposite (qxx) target preference as the
moving-to-preferred/non-preferred-target benefits an focal, play ulow, umedium or uhigh, respectively. Thus, all
individual had gained, and whether an individual possible complex behaviour strategies in the population
had incurred a consensus cost, or not. In this manner, can be described as {plow, pmedium, phigh, qlow, qmedium,
we could, for each combination of behaviours, determine qhigh} (with 0%plow%1; 0%pmedium%1Kplow; and phigh
the net gains (grouping benefitsCspatial target benefits) Z1KplowKpmedium; 0%qlow%1; 0%qmedium%1Kqlow;
for each group member at the end of each simulation. We and qhighZ1KqlowKqmedium). Using our derived pay-off
averaged those gains over the 500 replicated simulations matrix (table 1), and assigning animals drawn from the
for each combination of behaviours. population into groups randomly with respect to
For the purpose of the game theory model, only preferred target, we can now calculate the expected
differences in relative benefits for different behavioural relative net gains to a focal individual playing any focal
strategies are relevant (rather than absolute benefits; behaviour strategy {rlow, rmedium, rhigh} in a population
Maynard-Smith 1989). Therefore, in order to keep with any behaviour strategy {plow, pmedium, phigh, qlow,
parameters to a minimum, we expressed the relative qmedium, qhigh}, as follows:
Relative gainsðfrlow ; rmedium ; rhigh g; fplow ; pmedium ; phigh ; qlow ; qmedium ; qhigh gÞ
0 2 31
plow hK0:4qlow K0:9qmedium K qhigh i
B 6 7C
Z rlow B 2 2 6
@ð1KtÞ ½K1 C 0:4qlow C 2tð1KtÞ4 Cpmedium hK0:4qmedium K qhigh i
7C
5A
0 0
Cphigh qhigh hK0:4 C B2 KB3 i
0 2 2 1
ð1KtÞ ½K1 C 0:9qlow C 0:6qlow qmedium
B 2 3C
B plow hK0:1qlow K0:4qmedium K qhigh i C
B C
Crmedium B 6 7C
B C2tð1KtÞ6 Cpmedium qhigh h0:2 B20 KB30 K1i 7 C
@ 4 5A

Cphigh qhigh B20 KB30
!
ð1KtÞ2 qhigh ð2K qhigh Þ C 0:2q2medium KB30 C 0:8 qlow qhigh B20
Crhigh ; ð3:2Þ
C2tð1KtÞqhigh ½B20 KB30 K0:4plow B20

where t is the proportion of individuals in the (i) Complete segregation of target types
population, which have the same target preference If the benefits of being in a group of three are lower
as the focal. We assumed that these relative gains than the potential consensus costs (i.e. if B30 !1;
translate into fitness consequences to individual players figure 1, black areas) for both target types, the only
(Maynard-Smith 1989) and, thus, that plow, pmedium, ESS in the system is: ‘both target types always play
phigh and qlow, qmedium, qhigh evolve accordingly (see also uhigh’ (table 2). That is, none of the group members
appendix Ab). We have further assumed that the should compromise and incur consensus costs in order
proportion of different types of members (t) does not to maintain group cohesion (and gain grouping benefits
evolve. The rationale for this is as follows. Group that do not outweigh the consensus costs). This ESS
members usually differ with respect to target preference leads to complete segregation of different target types
because they are of different sex, age, size or into separate groups. In the rest of the results section,
physiological state (Conradt & Roper 2007), so t is we only consider what happens if grouping benefits
commonly determined by factors (e.g. reproduction) relative to potential consensus costs are high enough
that are independent of group decision making. not to automatically lead to complete segregation (i.e. if
However, this is not invariably so (see Rands et al. it is B30 O1 for at least one target type).
2003) and, by covering a wide parameter range for t
(0.1–0.9), we will be able to assess the potential effect t (ii) Partial segregation of target types
could have had if it had been allowed to co-evolve. If group size three is above optimal group size (i.e. if
B30 ! B20 ), and grouping benefits of being in a group of
three are less than five times as high as potential
(vii) Evolutionarily stable strategies consensus costs for at least one of the two target types
We assume that relative gains (see equation (3.2)) (i.e. if B30 !5), three possible situations can arise, as
translate into fitness consequences to individuals, and follows (see table 2 for criteria for each case). Firstly, no
that the system evolves until it either reaches an ESS exists and both types oscillate between playing
evolutionarily stable strategy (Maynard-Smith 1989) uhigh and umedium (figure 1, horizontally striped areas).
or starts to oscillate. In order to find the ESSs in the Secondly, one ESS exists: ‘more common type/type
system, we used a combination of analytical and with lower grouping benefits relative to potential
simulation methods, as follows. Assume target type 1 consensus costs always plays uhigh and other type
is the focal target type, and target type 2 the other target always plays umedium’ (figure 1b–e, light grey areas).
type. First, we analytically determined stability con- Thirdly, two ESSs exist: (i) ‘type 1 always plays uhigh
ditions for pure strategies (i.e. plowZ1, pmediumZ1 or and type 2 always umedium’ and (ii) ‘type 2 always plays
phighZ1; and qlowZ1, qmediumZ1 or qhighZ1; see uhigh and type 1 always umedium’ (figure 1a, light grey
appendix Ab(i)). The biological effect of the emerging area). The biological effect is the same in all three
pure ESSs (e.g. whether certain individuals or the situations, namely members will not always reach
majority led the group, and whether the group consensus and at least some group fragmentation will
fragmented or not) can be derived from the self- occur in mixed-type groups. This will lead to at least
organizing system simulations, given the pure strategies partial segregation of the two target types into separate
of both types (table 2). To examine the system for more groups.
complex ESSs, we used simulations. Starting from
different ‘starting population strategies’, we simulated (iii) Equally shared consensus decisions
the evolution of population strategy, based on the If, and only if, grouping benefits of being in a group of
relative gains given in equation (3.2) (assuming that three are more than five times as high as potential
these gains translate into fitness benefits; Maynard- consensus costs for both target types (i.e. if B30 O5), the
Smith 1989), until there was either no further change following strategy was always an ESS: ‘both target types
in population strategy (i.e. we had reached an ESS), or always play umedium’. The biological effect of this ESS is
the system began to oscillate. The details for these equally shared consensus decision making (see table 2;
simulations are given in appendix Ab(ii). By closely cf. Couzin et al. 2005). Note, however, that only
covering the whole parameter range for potential disjunct parameter values for u could be tested here,
starting population strategies, we made sure we and that ‘both target types always play umedium’ might
detected all potential ESSs in the system. cease to be an ESS in a continuous parameter space
for u. It would then be likely to be replaced by
another ‘equal-sharing’ ESS with a slightly higher
(c) Results common u value and slightly different stability criteria,
All of the ESSs that emerged were ‘pure’ ESSs, i.e. they while qualitative results would be similar (Conradt &
consisted of a single pure strategy rather than an Roper 2007).
equilibrium mix of pure strategies (figures 1 and 2, If the minimum condition B30 O5 was met for both
table 2). There were six possible ESSs (table 2). Their target types, and group size three was above optimal
biological consequence for group decisions ranged group size (i.e. if B30 ! B20 ), both target types always
from (i) complete segregation of different target types, play umedium was the only ESS in the system, and equal
via (ii) partial segregation, and (iii) equally shared sharing of consensus decisions always evolved (figure 1,
consensus decision making, to (iv) consensus decision spotted area). If group size three was not above optimal
making by individuals of only one of the two target group size (i.e. if B30 O B20 ; figure 1, white areas), two
types. We examine the ESSs, assembled by their further ESSs could exist: ‘type 1 always plays ulow, and
biological consequences, in detail, in the following. type 2 always uhigh’ (which would mean that type 2

individuals lead all decisions; tables 1 and 2) and (ii)
type 2 leads mixed groups towards target 2;
type 1 leads mixed groups towards target 1;

type 2 leads; mixed groups move cohesively
type 1 leads; mixed groups move cohesively

some group fragmentation, thus, partial
some group fragmentation, thus, partial

‘type 1 always plays uhigh, and type 2 always ulow’
equally shared decision; majority leads

(which would mean that type 1 individuals lead all
complete segregation of the two types

decisions). To which ESS the system evolves in such a
case depends on the starting point of the system, the
proportion t of type 1 in the population and on the
grouping benefits relative to potential consensus costs
(i.e. B30 and B20 ). If grouping benefits are large relative
to consensus costs (i.e. B30 and B20 are large), equal
sharing of decision making is more likely to evolve. If
cohesive group
biological effect
grouping benefits are small relative to consensus costs
segregation
segregation
to target 2
to target 1
(i.e. B30 and B20 are small), leading by one type is more
likely to evolve (see also next section).
(iv) One target type leads consensus decisions

We consider only situations without complete or partial
B30 ðtype 2ÞO ðð2KtC 1:6ð1KtÞB20 ðtype 2ÞÞ=ð1:6K0:6tÞÞ segregation (see above and table 2 for criteria). If group
size three is not above optimal group size (i.e. if
B30 ðtype 2ÞO 1C ð1:2ð1KtÞ=ð1C ð1KtÞÞÞ$B20 ðtype 2Þ
B30 O B20 ), the following two strategies are potential

B30 ðtype 1ÞO ðð1C t C 1:6tB20 ðtype 1ÞÞ=ð1C 0:6tÞÞ
ESSs: (i) ‘type 1 always plays uhigh and type 2 always

ulow’ and (ii) ‘type 2 always plays uhigh and type 1
always ulow’. Either of these ESSs results in decisions
B30 ðtype 1ÞO 1C ð1:2t=ð1C tÞÞ$B20 ðtype 1Þ
ESS, if B30 (type 1)O5 and B30 (type2)O5
ESS, if B30 (type1)!1 and B30 (type 2)!1
being made by individuals of only one target type. In

particular, the first ESS means that target type 1
ESS, if B20 (type 1)!B30 (type 1) and
ESS, if B20 (type 2)!B30 (type 2) and
individuals lead all consensus decisions (table 2;

ESS, if B20 (type 1)OB30 (type 1);
ESS, if B20 (type 2)OB30 (type 2);
figure 1, dark grey and vertically striped areas), the

second ESS that target type 2 individuals lead all
consensus decisions. Since individuals of the leading
target type always play uhigh, they do not compromise
and B30 (type 2)!5
and B30 (type 1)!5
with other group members but behave domineeringly.

Decisions are no longer equally shared, and the decision
stability criteria
is not necessarily made by members of the target type

never stable
never stable
never stable
that are in the majority within the group. If only one

member in a group is of the ‘leading’ target type, the
decision will, strictly speaking, even be ‘unshared’.
To which ESS the system evolves, depends on the
starting point of the system, the proportion t of type 1
in the population and on the grouping benefits relative
Table 2. Evolutionary stability criteria for pure strategies of either target type.
to potential consensus costs (i.e. B30 and B20 ). If the ratio

play always umedium (qmediumZ1)
of grouping benefits to potential consensus costs is

similar for individuals of both types (i.e. if B30 (type 1)Z
strategy of type 2 individuals
play always uhigh (qhighZ1)
B30 (type 2) and if B20 (type 1)ZB20 (type 2)), and if type
play always ulow (qlowZ1)
1 and 2 individuals are equally frequent in the

population (i.e. if tZ0.5), the ESSs ‘type 1 individuals
lead all consensus decisions’ and ‘type 2 individuals
lead all consensus decisions’ are equally likely to evolve
(figure 1a). Otherwise, if grouping benefits are small
relative to potential consensus costs (i.e. B30 is small;
figure 1, dark grey area; see table 2 for threshold), only
that ESS can evolve in which those individuals lead
all decisions, which are of the more common type
(figure 1b) and/or of the type with lower grouping
benefits relative to potential consensus costs
play always umedium ( pmediumZ1)
(figure 1c–e). As grouping benefits relative to potential

consensus costs increase (i.e. as B30 increases), the
strategy of type 1 individuals
play always uhigh ( phighZ1)

play always ulow ( plowZ1)
system can evolve also to ‘leading by the other type’

(figure 1, vertically striped areas). If grouping benefits
of being in a group of three exceed five times the
potential consensus costs (i.e. if B30 O5) for both target
types, a third ESS exists that leads to equal sharing of
consensus decisions (see section above). If relative
grouping benefits increase further (i.e. B30 increases
further, and in particular if B20 increases), the system is
more and more likely to evolve to equal sharing of

(a) 25 (b)
no ESS; 1 ESS: equal no ESS; 1 ESS: equal
oscillations sharing of decisions oscillations sharing of decisions
and partial and partial
20 equal sharing equal sharing
segregation segregation
1 ESS: complete segregation

is more frequent is more frequent

1 ESS:
more common type
15 3 possible ESSs: 3 possible ESSs:
B2′ (type 1)
decides; but: partial

2 ESSs: (i) equal sharing of (i) equal sharing of decisions
segregation
(i) type 1 decides decisions (ii) more common type decides
(ii) type 2 decides (ii) type 1 decides (iii) less common type decides
10 (iii) type 2 decides
but: partial segregation
‘one type decides’ ‘one type decides’
is more frequent is more frequent
5
2 possible ESSs: 2 possible ESSs:
(i) type 1 decides (i) more common type decides
(ii) type 2 decides (ii) less common type decides
0
1 ESS:
more common type decides
(c) 25 (d)
no ESS; 1 ESS: 1 ESS: no ESS; 1 ESS:
oscillations and type 2 equal oscillations and 1 ESS: equal
partial segregation decides; sharing of partial segregation type 2 sharing of
20
but: partial decisions decides; decisions

segregation but: partial

segregation
15 3 possible 3 possible
B2′ (type 1)
ESSs: ESSs:
(i) equal (i) equal
sharing of sharing of
10 decisions
decisions
(ii) type 1 (ii) type 1
2 possible ESSs: decides 2 possible ESSs: decides
5 (i) type 1 decides (iii) type 2 (i) type 1 decides (iii) type 2
(ii) type 2 decides decides (ii) type 2 decides decides
1 ESS: type 2 1 ESS: type 2
decides decides
0 5 10 15 20 25 0 5 10 15 20 25
B3′ (type 1) B3′ (type 1)
(e) 25
1 ESS:
equal
no ESS; sharing of
20 oscillations and decisions
partial segregation
15 3 possible
B2′ (type 1)
ESSs:
(i) equal
sharing of
10 decisions
2 possible ESSs:
(ii) type 1
(i) type 1 decides
decides
(ii) type 2 decides
5 (iii) type 2
decides
1 ESS: type
2 decides
0 5 10 15 20 25
B3′ (type 1)
Figure 1. ESSs depending on the grouping benefits to type 1 of being in a group of three (B30 , x -axis) or two (B20 , y-axis), respectively,
relative to the potential consensus costs; (a) both types with same grouping benefits to consensus cost ratio (i.e. B30 (type 1)ZB30 (type
2); B20 (type 1)ZB20 (type 2)), tZ0.5; (b) both types with same grouping benefits to consensus cost ratio, tZ0.1/0.9; (c) type 1 with
four times higher grouping benefits to consensus cost ratio (i.e. B30 (type 1)Z4$B30 (type 2); B20 (type 1)Z4$B20 (type 2)), tZ0.1;
(d ) type 1 with four times higher grouping benefits to consensus cost ratio, tZ0.5; (e) type 1 with four times higher grouping benefits
to consensus cost ratio, tZ0.9 (black areas, 1 ESS, all individuals play always uhigh, resulting in complete segregation; horizontally
striped areas, no ESS, individuals oscillate between playing umedium and uhigh, resulting in varying leading and partial segregation;
light grey areas, 1 (or 2) ESSs, the type with the highest potential consensus costs or which is more common plays always uhigh, the
other type plays always umedium, resulting partially in leading by first type and partially in segregation; dark grey areas, 1 ESS, the type
with the highest potential consensus costs or which is more common plays always uhigh, the other type plays always ulow, resulting in
leading of decisions by first type; vertically striped areas, 2 ESSs, one type plays always uhigh, the other type always ulow, resulting in
‘leading by type 1’ or ‘leading by type 2’; spotted areas, 1 ESS, all types play always umedium, resulting in equally shared decision
making; white areas, 3 ESSs, ‘leading by type 1’, ‘leading by type 2’ and ‘equally shared decision making’).

favour the evolution of equally shared versus less

relative to potential consensus costs
benefits of being ina smaller group shared decision making. The reason for this apparent
contradiction lies in the different nature of decisions
partial segregation
complete segregation equal sharing of decisions
about activity synchronization (which were the subject
of Conradt & Roper 2007) and decisions about
ze movement destinations. In decisions about the timing
p si
g rou of activities, compromises that can be reached in an
tim
a l size
o p r o up equally shared manner (e.g. the average or median of
ve al g
abo ptim preferred timings by group members) usually lower the
r o up e o
g ovw
babelo average consensus costs to individual group members
up
gro (Conradt & Roper 2003). This is generally true if the
leading by one type modality which the group decides upon (e.g. timing) is
continuous with respect to benefits/costs. In such a
benefits of being in a larger group case, a compromise can be viable. To give an example:
relative to potential consensus costs if you want to go to a restaurant at 19.00 hours, and I
Figure 2. Schematic summary of model predictions. The area want to go at 20.00 hours, going there at 19.30 hours
above the dotted line refers to groups that are above optimal might mean that both of us do not get inconvenienced
group size; the area below the dotted line refers to groups of too much. To compromise (and share a decision) on
suboptimal size. such a continuous modality will be particularly crucial
if potential consensus costs are high and grouping
consensus decisions rather than to leading by either benefits marginal, since group members should only
type (figure 1, white areas; see also section above). stay in a group if grouping benefits exceed consensus
costs. Thus, in such situations, one expects either
(v) Potential effect of t equally shared decisions or group fragmentation.
Qualitative results were fairly constant over a wide range Group decisions about timing, while not always, will
of values for t (compare figure 1a,b and figure 1c–e), usually fall into this category.
implying that an evolving t would not have led to By contrast, if the modality which the group decides
greatly different results, always assuming that ultimately upon (e.g. spatial targets) is disjunct with respect to
neither target type (e.g. sex, age or size class; Conradt & benefits/costs, then a compromise is unlikely to be
Roper 2007) would go completely extinct. viable. To give an example: if you are hungry and want
to go to a restaurant, and I need petrol and want to go
(d) Discussion to a garage, compromising and going halfway between
To date, hardly anything is known about the evolution the restaurant and the garage (where we could neither
of group decision-making strategies in animals. The get food nor petrol) does not make sense for either of
present study is therefore pioneering work, in which we us. While not all decisions on spatial targets are
introduce a novel technique (combining self-organizing necessarily ‘disjunct decisions’ in this way, the majority
system and game theory modelling) and explore a first are, most notably in patchy landscapes. Such circum-
application. Although we were only able to model a stances usually require ‘hard’ decisions, of an ‘either/or’
small group (group size three) owing to computational kind, to be made. Thus, groups in the kind of situation
limitations, our results provide a preliminary under- that we have considered here have only three viable
standing of the evolution of spatial decisions in groups options: move to target 1, move to target 2 or fragment.
in which there is no communication between members If grouping benefits are large relative to consensus
and in which individuals are unable to perceive or react costs, no group member wishes to risk group
directly to global group behaviour. We show that in fragmentation, so being too assertive with respect to
such groups, equally shared consensus decision mak- your own preference is risky and therefore disadvan-
ing, as well as leading by one ‘domineering’ target type, tageous. On the other hand, if nobody is highly
could in principle evolve in decisions about movement assertive it pays to be somewhat assertive, because
destinations. If grouping benefits were small relative to otherwise you would never move to your own preferred
potential consensus costs, groups either became target. Consequently, evolution favours a medium level
unstable such that individuals with different prefer- of assertiveness in all individuals, which leads in a self-
ences segregated into different subgroups (if group size organizing system to equally shared decision making
was above optimal group size), or leading by one type (Couzin et al. 2005). By contrast, if grouping benefits
evolved (if group size was not above optimal group size; are low relative to consensus costs, risking group
figure 2). The leading type was likely to be either the fragmentation rather than moving towards a less
type with the lower grouping benefits to potential preferred target becomes an evolutionary option.
consensus cost ratio, or the more common type in the Thus, it will pay some individuals, in particular those
overall population (although not necessarily within the which would face higher consensus costs if moving
group). As grouping benefits relative to potential towards the less preferred target (or those which stand
consensus costs increased, equal sharing of decisions to gain less grouping benefits), to be highly assertive. As
became more likely to evolve, in particular when group individuals of one target type evolve to be highly
size was above optimal group size. assertive, selection might then favour individuals of the
These results partially contradict those of Conradt & other target type to be unassertive in order to avoid
Roper (2007), who predicted that high potential group fragmentation. The result is that ‘leading by one
consensus costs and marginal grouping benefits should type’ evolves. If grouping benefits drop further,

individuals with different preferences segregate into them (‘social attraction’). However, the resulting
different subgroups. ‘social attraction vector’ might differ from the vector
An interesting implication of this is that if animals towards the individual’s preferred spatial target. There-
make simultaneous group decisions about continuous fore, the individual ‘balances’ attraction to its preferred
and disjunct modalities (e.g. about space and time), target against social attraction to neighbours with
they should apply different strategies to the two its own individual factor ui (‘degree of assertiveness’).
different aspects of the decision. How the details That is, an individual tries to move in the direction
of combined space-and-time decisions could look of social attractionCui$preferred target direction
requires further modelling. (where preferred target direction is the vector between
The model proves that equal sharing of decisions is the individual’s position and the preferred target
not a foregone conclusion, even in self-organizing position, scaled to unit length). Thus, if Nia(t)Z0
systems where individuals do not have an overview of, and Nir(t)O0, then
or could directly react to, global group behaviours
X
r N ðtÞ X
rN ðtÞ
(Camazine et al. 2003). On the contrary, we show that 1 c ðtÞK ci ðtÞ 1
di ðtÞ Z $ j C $ v ðtÞ
leading by one type can evolve, even if group members 4Nr ðtÞ jsi cj ðtÞK ci ðtÞ 4Nr ðtÞ jsi j
cannot communicate, and consequently have no
knowledge about, the preferences of other group gi K ci ðtÞ
C ui $ ; ðA 2Þ
members. The model offers testable predictions about 2$jgi K ci ðtÞj
when equal sharing or leading by one type should occur where Nir(t) is the number of neighbours within range
in group decisions about movement destinations. r; vi(t) is the unit vector of movement direction, for
The present model is for computational reasons individual i at time step t; and gi is the position vector of
restricted to groups of three. However, its results the preferred target of individual i. The first term
provide initial insight into an important type of group describes social attraction to neighbours, the second
decision (see Couzin & Krause 2003 for a review). Its term alignment with neighbours (first and second term
most important qualitative result (i.e. the principal together constitute the social attraction vector), and the
difference between decisions in continuous and dis- third term attraction to the preferred goal.
junct modalities, and thus, usually between timing and It follows that, if ui is large (and the individual is
spatial decisions) is likely to hold for larger group sizes, highly assertive), the individual tries to move at each
for the reasons given above. Whether other results also time step predominantly in the direction of its preferred
hold requires further investigation. As computer speeds target; if ui is small (and the individual is of low
increase, our approach will become feasible for studies assertiveness), it predominantly tries to move towards,
on larger groups. and align with, neighbours; and if ui is intermediate
L.C. was supported by a Royal Society University Research (medium assertiveness), the individual compromises
Fellowship, and would like to thank the Royal Society between its own target preference and moving towards,
particularly for its support during maternity leave and with and aligning with, neighbours.
respect to part-time working. Absence of neighbours. If there are no neighbours
within ranges a or r, the individual always tries to move
in the direction of its preferred target. Thus, if Nia(t)Z
APPENDIX A 0 and Nir(t)Z0, then
(a) Details of the self-organizing group
movement simulation model gi K ci ðtÞ
di ðtÞ Z : ðA 3Þ
(i) Individual local behaviour rules jgi K ci ðtÞj
Collision avoidance. For each individual, the highest
Note that none of these assumptions regarding local
priority is to avoid collision. Thus, if there are
behavioural rules requires a global overview of the
neighbours within a collision range a (aZ2; for
group, or a view of other group members further away
adequacy of parameter choice see Couzin et al. 2005),
than distance r.
the individual simply tries to turn away from those
neighbours. Thus, if Nia(t)O0, then
(ii) Self-organizing system simulations
X
N ia ðtÞ
1 c ðtÞK ci ðtÞ Individual group members start with a random move-
di ðtÞ ZK $ j ; ðA 1Þ
Nia ðtÞ jsi cj ðtÞK ci ðtÞ ment direction and a random position within a cohesive
group. From the preliminary movement direction di(t)
where Nia(t) is the number of neighbours within range (equations (A1)–(A3), above), we derive the new
a; di(t) is the preliminary movement direction; and ci(t) movement direction for each individual at the next
is the position vector, for individual i at time t. time step by adding an individual random directional
Balance between social attraction and preferred target error to di(t) (with standard deviation sZ0.01 radians;
direction. If there are no neighbours within range a (i.e. Couzin et al. 2005), and limiting the change in
if Nia(t)Z0), the individual’s movement is based on its movement direction to a maximum turning angle of
desire to maintain group cohesion as well as to move qDt (qDtZ0.2 radians; see Couzin et al. 2005 for details
towards its preferred target, as follows. In order to and justification of parameter choice). The
maintain group cohesion, the individual is attracted to new position of each individual is obtained by moving
neighbours within a range r (rZ30; for adequacy it one step ahead of DtZ0.1 (equivalent to a const-
of parameter choice see Couzin et al. 2005) so as to ant speed of 1 spatial unit/time unit) in the new
move towards them and align travel direction with movement direction.

(b) Game theory model and

(i) Criteria for the increases in the population of different
Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20
behavioural strategies
We determine the criteria under which different pure O Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 : ðA 7bÞ
behavioural strategies should increase or decrease in
the population, using the expected relative gains given Of course, the analogue is true for the members in the
in equation (3.2). Equation (3.2) can be written population of the opposite target preference. Thus, qlow
shorter as increases in the population if
A1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20
Relative gainsðfrlow ; rmedium ; rhigh g;
O B1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 ðA 8aÞ
fplow ; pmedium ; phigh ; qlow ; qmedium ; qhigh g; tÞ
and
Z rlow $At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20
A1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20
C rmedium $Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20
O C1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 ; ðA 8bÞ
C rhigh $Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 : ðA 4Þ
qmedium increases in the population if
These relative gains peak with rlowZ1 if B1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20
At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 is relative large (which
means that it is of advantage to play ulow). Similarly, they O A1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 ðA 9aÞ
peak with r medium if Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20
is relatively large (which means that it is of advantage to and
play umedium). Finally, they peak with rhigh if B1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20
Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 is relative large (which
means that it is of advantage to play uhigh). We assume O C1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 ðA 9bÞ
that relative gains translate into relative fitness increases
(Maynard-Smith 1989). It therefore follows that plow and qhigh increases in the population if
(i.e. the probability with which individuals of the focal’s C1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20
target preference play ulow) increases in the population O A1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 ðA 10aÞ
relative to other strategies, if relative gains reach a
maximum for rlowZ1. Similarly, pmedium increases in and
the population if relative gains reach a maximum for C1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20
rmediumZ1. Finally, phigh increases in the population if
relative gains reach a maximum for rhighZ1. In particular, O B1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 : ðA 10bÞ
plow increases in the population if
At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20

(ii) Simulations to find complex ESSs
O Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 ðA 5aÞ First, we chose a starting population strategy {plow,
pmedium, phigh, qlow, qmedium, qhigh}. Using equations (3.2)
and and (A 4), we calculated At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 ,
Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 , Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;
At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 qhigh ;B30 ;B20 , A1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 , B1Kt;qlow ;
qmedium ;qhigh ;plow ;pmedium ;phigh ;B30 ;B20 and C1Kt;qlow ;qmedium ;qhigh ;plow ;pmedium ;
O Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 ; ðA 5bÞ
phigh ;B30 ;B20 for this starting population strategy. According
pmedium increases in the population, if to the inequality relationships listed in the previous
section (equations (A 5a)–(A 10b)), we determined
Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 which pxx and qxx probabilities increase in the population.
We then added a random increase to the relevant
O At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 ðA 6aÞ probabilities in the population strategy (with a mean of
0.001 and a standard deviation of 0.0005, but capping
and the probabilities at 1). The relevant increase was
then deducted from the remaining non-increasing
Bt;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 probabilities (until they reach a minimum of 0) in such
a manner that the following holds true for the newly
O Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 ðA 6bÞ derived probabilities: plow 0 0
C pmedium C phigh0
Z 1 and
0 0 0
qlow C qmedium C qhigh Z 1. In this manner, we derive a
and phigh increases in the population, if new, evolved population strategy. We repeated the
process again using the newly derived population
Ct;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 strategy. The process was repeated until all the prob-
abilities stabilized and the derived population strategy
O At;plow ;pmedium ;phigh ;qlow ;qmedium ;qhigh ;B30 ;B20 ðA 7aÞ was an ESS, or until the system started to oscillate.

Our starting population strategies covered the whole Conradt, L. & Roper, T. J. 2003 Group decision-making
parameter space (in steps of 0.1) and used all in animals. Nature 421, 155–158. (doi:10.1038/nature
biologically possible combinations of plow2{0, 0.1, 01294)
0.2, ., 1); pmedium 2{0, 0.1, 0.2, . 1Kp low ); Conradt, L. & Roper, T. J. 2005 Consensus decision making
in animals. Trends Ecol. Evol. 20, 449–456. (doi:10.1016/
q low 2{0, 0.1, 0.2, . 1); and q medium2{0, 0.1,
j.tree.2005.05.008)
0.2, . 1Kqlow). Thus, by closely covering the whole Conradt, L. & Roper, T. J. 2007 Democracy in animals: the
range of potential starting population strategies, and by evolution of shared group decisions. Proc. R. Soc. B 274,
using small increases at each simulation step, we made 2317–2326. (doi:10.1098/rspb.2007.0186)
sure that we detected all potential ESSs in each system. Conradt, L., Couzin, I. D., Krause, J., Roper, T. J. In press.
The examined parameter range for proportion t was Leading according to need in self-organising groups.
0.1, 0.5 and 0.9; and that for relative grouping benefits Am. Nat.
B30 (type 1), B30 (type 2), B20 (type 1) and B20 (type 2): Couzin, I. D. & Krause, J. 2003 Self-organization and
0.5–25 (in steps small enough to cover all potentially collective behavior in vertebrates. Adv. Study Behav. 32,
different cases arising from the stability conditions in 1–75. (doi:10.1016/S0065-3454(03)01001-5)
table 2). It was assumed that the ratio of grouping Couzin, I. D., Krause, J., Franks, N. R. & Levin, S. A. 2005
Effective leadership and decision-making in animal groups
benefits relative to potential consensus costs (B30 and
on the move. Nature 433, 513–516. (doi:10.1038/
B20 ) were either the same for individuals that preferred nature03236)
targets 1 and 2 (i.e. B30 (type 1)ZB30 (type 2) and B20 Dostalkova, I. & Spinka, M. 2007 Synchronisation of
(type 1)ZB20 (type 2)), or they were four times higher behaviour in pairs: the role of communication and
for individuals that preferred target 1 than for those consequences in timing. Anim. Behav. 74, 1735–1742.
that preferred target 2 (i.e. B30 (type 1)Z4$B30 (type 2) (doi:10.1016/j.anbehav.2007.04.014)
and B20 (type 1)Z4 B20 (type 2)). Dyer, J. R. G., Johansson, A., Helbing, D., Couzin, I. D. &
Krause, J. 2009 Leadership, consensus decision making
and collective behaviour in humans. Phil. Trans. R. Soc. B
364, 781–789. (doi:10.1098/rstb.2008.0233)
Fischhoff, I. R., Sundaresan, S. R., Cordingley, J., Larkin,
REFERENCES H. M., Sellier, M. J. & Rubenstein, D. I. 2007 Social
Ame, J. M., Halloy, J., Rivault, C., Detrain, C. & relationships and reproductive state influence leadership
Deneubourg, J. L. 2006 Collegial decision making based roles in movements of plains zebra, Equus burchellii. Anim.
on social amplification leads to optimal group formation. Behav. 73, 825–831. (doi:10.1016/j.anbehav.2006.10.012)
Proc. Natl Acad. Sci. USA 103, 5835–5840. (doi:10.1073/ Franks, N. R., Dechaume-Moncharmont, F.-X., Hanmore,
pnas.0507877103) E. & Reynolds, J. K. 2009 Speed versus accuracy in
Austen-Smith, D. & Feddersen, T. J. 2009 Information decision-making ants: expediting politics and policy
aggregation and communication in committees. Phil. implementation. Phil. Trans. R. Soc. B 364, 845–852.
Trans. R. Soc. B 364, 763–769. (doi:10.1098/rstb.2008. (doi:10.1098/rstb.2008.0224)
0256) Gächter, S. & Herrmann, B. 2009 Reciprocity, culture and
Biro, D., Sumpter, D. J. T., Meade, J. & Guilford, T. 2006 human cooperation: previous insights and a new cross-
From compromise to leadership in pigeon homing. Curr. cultural experiment. Phil. Trans. R. Soc. B 364, 791–806.
Biol. 16, 2123–2128. (doi:10.1016/j.cub.2006.08.087) (doi:10.1098/rstb.2008.0275)
Black, J. M. 1988 Preflight signaling in swans—a mechanism Gompper, M. E. 1996 Sociality and asociality in white-nosed
for group cohesion and flock formation. Ethology 79, coatis (Nasua narica): foraging costs and benefits. Behav.
143–157. Ecol. 7, 254–263. (doi:10.1093/beheco/7.3.254)
Boinski, S. & Campbell, A. F. 1995 Use of trill vocalizations Hix, S., Noury, A. & Roland, G. 2009 Voting patterns
to coordinate troop movement among white-faced and alliance formation in the European Parliament. Phil.
capuchins—a 2nd field-test. Behaviour 132, 875–901. Trans. R. Soc. B 364, 821–831. (doi:10.1098/rstb.2008.
(doi:10.1163/156853995X00054) 0263)
Byrne, R. W. 2000 How monkeys find their way: leadership, Kerth, G., Ebert, C. & Schmidtke, C. 2006 Group decision
coordination and cognitive maps of African baboons. In making in fission–fusion societies: evidence from two-field
On the move (eds S. Boinski & P. A. Garber), pp. 491–518. experiments in Bechstein’s bats. Proc. R. Soc. B 273,
Chicago, IL: University of Chicago Press. 2785–2790. (doi:10.1098/rspb.2006.3647)
Camazine, S., Deneubourg, J. L., Franks, N. R., Sneyd, J., Krause, J. 1993 The relationship between foraging and shoal
Theraulaz, G. & Bonabeau, E. 2003 Self-organization in position in a mixed shoal of roach (Rutilus rutilus) and
biological systems. Princeton, NJ: Princeton University chub (Leuciscus-Cephalus)—a field-study. Oecologia 93,
Press. 356–359. (doi:10.1007/BF00317878)
Clutton-Brock, T. H., Guinness, F. E. & Albon, S. D. 1982 Krause, J. & Ruxton, G. D. 2002 Living in groups. Oxford,
Red deer: behaviour and ecology of two sexes. Chicago, IL: UK: Oxford University Press.
University of Chicago Press. Krause, J., Bumann, D. & Todt, D. 1992 Relationship
Codling, E. A., Pitchford, J. W. & Simpson, S. D. 2007 Group between the position preference and nutritional state of
navigation and the “many-wrongs principle” in models of individuals in schools of juvenile roach (Rutilus rutilus).
animal movement. Ecology 88, 1864–1870. (doi:10.1890/ Behav. Ecol. Sociobiol. 30, 177–180. (doi:10.1007/
06-0854.1) BF00166700)
Conradt, L. 1998 Could asynchrony in activity between the List, C. 2004 Democracy in animal groups: a political science
sexes cause intersexual social segregation in ruminants? perspective. Trends Ecol. Evol. 19, 168–169. (doi:10.1016/
Proc. R. Soc. B 265, 1359–1363. (doi:10.1098/rspb.1998. j.tree.2004.02.004)
0442) List, C., Elsholtz, C. & Seeley, T. D. 2009 Independence and
Conradt, L. & Roper, T. J. 2000 Activity synchrony and social interdependence in collective decision making: an agent-
cohesion: a fission–fusion model. Proc. R. Soc. B 267, based model of nest-site choice by honeybee swarms. Phil.
2213–2218. (doi:10.1098/rspb.2000.1171) Trans. R. Soc. B 364, 755–762. (doi:10.1098/rstb.2008.0277)

Lusseau, D. 2007 Evidence for social role in a dolphin social Seeley, T. D. & Buhrman, S. C. 1999 Group decision making
network. Evol. Ecol. 21, 357–366. (doi:10.1007/s10682- in swarms of honey bees. Behav. Ecol. Sociobiol. 45, 19–31.
006-9105-0) (doi:10.1007/s002650050536)
Maynard-Smith, J. 1989 Evolutionary genetics. Oxford, UK: Sellers, W. I., Hill, R. A. & Logan, B. S. 2007 An agent-based
Oxford University Press. model of group decision making in baboons. Phil. Trans.
Meunier, H., Leca, J. B., Deneubourg, J. L. & Petit, O. 2006 R. Soc. B 362, 1699–1710. (doi:10.1098/rstb.2007.2064)
Group movement decisions in capuchin monkeys: the Simons, A. M. 2004 Many wrongs: the advantage of group
utility of an experimental study and a mathematical model navigation. Trends Ecol. Evol. 19, 453–455. (doi:10.1016/
to explore the relationship between individual and j.tree.2004.07.001)
collective behaviours. Behaviour 143, 1511–1527. Skyrms, B. 2009 Evolution of signalling systems with
(doi:10.1163/156853906779366982) multiple senders and receivers. Phil. Trans. R. Soc. B
Milton, K. 2000 Quo vadis? Tactics of food search and group 364, 771–779. (doi:10.1098/rstb.2008.0258)
movements in primates and other animals. In On the move Stewart, K. J. & Harcourt, A. H. 1994 Gorillas vocalizations
(eds S. Boinski & P. A. Garber), pp. 375–418. Chicago, IL: during rest periods—signals of impending departure.
University of Chicago Press. Behaviour 130, 29–40. (doi:10.1163/156853994X00127)
Passino, K. M. & Seeley, T. D. 2006 Modeling and analysis of Stolba, A. 1979 Entscheidungsfindung in Verbanden von Papio
nest-site selection by honeybee swarms: the speed and hamadryas. Zuerich, Switzerland: Universitaet Zuerich.
accuracy trade-off. Behav. Ecol. Sociobiol. 59, 427–442. Sumpter, D. J. T. 2006 The principles of collective animal
(doi:10.1007/s00265-005-0067-y)
behaviour. Phil. Trans. R. Soc. B 361, 5–22. (doi:10.1098/
Prins, H. H. T. 1996 Ecology and behaviour of the African
rstb.2005.1733)
buffalo. London, UK: Chapman & Hall.
Sumpter, D. J. T. & Pratt, S. C. 2009 Quorum responses and
Rands, S. A., Cowlishaw, G., Pettifor, R. A., Rowcliffe, J. M.
consensus decision making. Phil. Trans. R. Soc. B 364,
& Johnstone, R. A. 2003 Spontaneous emergence of
743–753. (doi:10.1098/rstb.2008.0204)
leaders and followers in foraging pairs. Nature 423,
432–434. (doi:10.1038/nature01630) Trillmich, J., Fichtel, C. & Kappeler, P. M. 2004 Coordi-
Ruckstuhl, K. E. 1998 Foraging behaviour and sexual nation of group movements in wild Verreaux’s sifakas
segregation in bighorn sheep. Anim. Behav. 56, 99–106. (Propithecus verreauxi ). Behaviour 141, 1103–1120.
(doi:10.1006/anbe.1998.0745) (doi:10.1163/1568539042664579)
Ruckstuhl, K. E. 1999 To synchronise or not to synchronise: Visscher, P. K. & Seeley, T. D. 2007 Coordinating a group
a dilemma for young bighorn males? Behaviour 136, departure: who produces the piping signals on honeybee
805–818. (doi:10.1163/156853999501577) swarms? Behav. Ecol. Sociobiol. 61, 1615–1621. (doi:10.
Ruckstuhl, K. E. & Neuhaus, P. 2000 Sexual segregation in 1007/s00265-007-0393-3)
ungulates: a new approach. Behaviour 137, 361–377. Wallraff, H. G. 1978 Social interrelations involved in
(doi:10.1163/156853900502123) migratory orientation of birds—possible contribution of
Ruckstuhl, K. E. & Neuhaus, P. 2002 Sexual segregation in field studies. Oikos 30, 401–404. (doi:10.2307/3543490)
ungulates: a comparative test of three hypotheses. Biol. Ward, A. J. W., Sumpter, D. J. T., Couzin, J. D., Hart,
Rev. 77, 77–96. P. J. B. & Krause, J. 2008 Quorum decision making
Sapolsky, R. M. 2005 The influence of social hierarchy on facilitates information transfer in fish schools. Proc. Natl
primate health. Science 308, 648–652. (doi:10.1126/ Acad. Sci. USA 105, 6948–6953. (doi:10.1073/pnas.
science.1106477) 0710344105)

Phil. Trans. R. Soc. B (2009) 364, 821–831

doi:10.1098/rstb.2008.0263
Voting patterns and alliance formation in the

European Parliament
Simon Hix1,*, Abdul Noury2 and Gérard Roland3
1
London School of Economics and Political Science, Houghton Street, London WC2A, UK
2
University of Brussels, 1070 Brussels, Belgium
3
University of California, Berkeley, CA 94720-3880, USA
Members of the European Parliament (MEPs) have voluntarily formed transnational political groups
and invariably follow the voting instructions of these groups. This is intriguing as there are few obvious
incentives for doing so. Unlike national parties, for example, the political groups in the European
Parliament are not punished by the electorate if they are divided on key issues, as citizens know very little
about what goes on inside the European Parliament. This paper pieces together an explanation of why the
European political groups exist and why they have become so powerful by looking at the determinants of
group cohesion and by undertaking a spatial analysis of voting in the European Parliament. MEPs who
share preferences on a range of issues on the European Union policy agenda have an incentive to establish
a division-of-labour contract and to share the costs of collecting information. Once internal party policy
specialization and agenda setting has been established, MEPs have incentives to follow the voting
instructions of their group owing to the advantages of cohesion in a context of repeated voting.
Keywords: roll-call voting; coalition formation; legislative behaviour; political parties;
European Parliament
1. INTRODUCTION This is intriguing for several reasons. First, reforms

The European Parliament is a fascinating laboratory for of the EU treaty since the mid-1980s have transformed
the study of social behaviour, in general, and the behaviour the European Parliament from a purely consultative
of elected officials, in particular. The European Parlia- chamber into a powerful legislature, with the power
ment is a large and heterogeneous chamber, and has inter alia to amend and block most EU laws. How the
grown more heterogeneous over time. At the time of the MEPs vote now affects the lives of approximately 500
first European Parliament elections in June 1979, there million EU citizens. Not surprisingly, MEPs are now
were 410 Members of the European Parliament (MEPs) lobbied by governments, national politicians, interest
from 10 European Union (EU) member states and 51 groups and professional lobbyists. Yet, the European
different national political parties. Approximately 30 years political groups have not collapsed in the face of these
later, the European Parliament has 785 MEPs from 27 higher political stakes.
member states and over 170 national political parties. Second, the main political groups have become
Rather than sit as delegates of their member state more internally heterogeneous as a result of EU
or national party, the MEPs sit in ‘political groups’ enlargement and their expansion to bring in parties
according to their ideological preferences (table 1). with more varied policy preferences. This is particularly
Research suggests that these transnational party the case with the largest group in the current
alliances are more than convenient umbrella organiz- parliament, the European People’s Party (EPP),
ations. The political groups have become powerful which is a broad coalition of conservatives and other
actors, able to marshal their troops in support of the parties on the centre-right.
policy positions of the group leaders (Raunio 1997; Third, unlike most party ‘factions’ in democratic
Kreppel 2002; Hix et al. 2005). For example, Noury & parliaments, the political groups in the European
Roland (2002) estimate that if one only knows which Parliament have few powers to force their members to
political group an MEP belongs to, one could correctly follow voting instructions. National parties rather than
predict her voting behaviour 90 per cent of the time, the European political groups control the process of
while if one knows only which member state an MEP selecting candidates in the European Parliament
belongs to, one could correctly predict her voting elections. And, there are no personal electoral incen-
behaviour only 10 per cent of the time. In other words, tives for MEPs to vote with their European political
a British Labour MEP is far more likely to vote with a groups, as the European Parliament elections are
French or German Social Democrat than with a British generally fought by national parties on the performance
Conservative or British Liberal Democrat. of national governments and not on the performance of
the groups in the European Parliament (van der Eijk &
Franklin 1996; Hix & Marsh 2007).
* Author for correspondence (s.hix@lse.ac.uk). So, why do MEPs vote with their European political
One contribution of 11 to a Theme Issue ‘Group decision making in groups? To answer this question, the next section looks
humans and animals’. at the standard explanations in political science of why

822 S. Hix et al. Voting patterns in the European Parliament
Table 1. Political groups in the European Parliament in January 2008. (The ideology column describes the general policy
positions of the national members’ parties in each European political group.)
political group ideology number of MEPs percentage of MEPs
European People’s Party–European Democrats conservatives 289 36.8

Party of European Socialists socialists 215 27.4
Alliance of Liberals and Democrats for Europe liberals 101 12.9
Union for Europe of the Nations nationals 44 5.6
Greens–European Free Alliance greens 42 5.4
European United Left–Nordic Green Left radical left 41 5.2
Independence and Democracy anti-Europeans 24 3.1
non-attached members independents 29 3.7
total 785 100.0
parties form and act cohesively in parliaments. Section internal to parliamentary politics rather than mainly
3 then investigates the determinants of political group external, relating to the electoral process.
cohesion in the European Parliament, while §4 employs The first such explanation is that legislative parties
a spatial analysis to look at individual MEP voting reduce the volatility of policy and thus increase the
behaviour. Section 5 then relates the results of these predictability of parliamentary politics. In any session
analyses to the theoretical ideas. Finally, §6 contains a of a parliament, there are a high number of items on
brief conclusion. the agenda. Also, each one of these items could
have multiple sub-issues within it with various policy
implications. If an issue is multidimensional, there may
2. THEORIES OF PARTIES IN PARLIAMENTS
be no stable policy equilibrium if anyone is free to
Why do political parties form and act collectively in
propose amendments (McKelvey 1976; Riker 1980).
parliaments? The answer might, at first, seem relatively
trivial for anyone with a basic understanding of In other words, in the absence of some form of
democratic politics. A politician has a much better collective organization, legislative politics is potentially
chance of being elected if he or she is a member of a highly chaotic and uncertain. Legislative parties create
political party, as the party (usually) provides campaign voting cohesion among their members, which strongly
resources (people and money), and a party label reduces the dimensionality of voting and also increases
efficiently communicates a large amount of information stability and predictability.
about the retrospective and prospective policy prefer- By establishing an organization (a party) with other
ences of the politicians. Then, once elected, the politicians with similar policy preferences across a
politician has an incentive to maintain the coherence range of issues, a politician has a much greater chance
of the party ‘brand’, for example by following the voting of achieving her policy goals in the legislative process,
instructions of the party leadership. If the party is and at a lower cost. It is very costly for a politician to
divided in the parliament, it will send mixed signals to carefully consider her position on each issue and then
the electorate about what policies the party will pursue cast her vote sincerely. The politician would have to
in the future (cf. Schlesinger 1984; Cox 1987; Cox & collect a huge amount of information to be able to work
McCubbins 1993; Aldrich 1995). Elected representa- out her position on every item. Also, the politician is
tives have an incentive to follow party discipline owing unlikely to know before a vote whether she will be on
to the potential reputation loss and negative electoral the winning side, the losing side or pivotal. If she could
consequences, which might ensue otherwise. persuade some colleagues to vote the same way as she
This seems straightforward enough. However, why does, her chances of being on the winning side are likely
do politicians decide to form party organizations to increase. The group of politicians could agree to
inside legislatures independently of the process of delegate to the person among them who has the most
democratic elections? The first modern parties expertise on a particular issue to instruct the others
emerged in parliaments in the early nineteenth century, how to vote. The group could then be confident that
long before universal suffrage and mass electoral they will all vote the right way and that their chances of
politics ( LaPalombara & Weiner 1966). As the being on the winning side would be much higher.
Federalists and Republicans in the early US Con- However, two members of the group might claim to
gresses, pre-democratic ‘parliamentary parties’ in be experts on the same issue yet disagree on how the
Europe included the Whigs and Tories in the British group should vote and so issue competing instructions.
House of Commons, the National Liberals and Which way should the others vote now? It immediately
Progressives in the German Reichstag and the Con- becomes apparent that since the behaviour is repeated,
servative and Left (Venstre) parties in the Scandinavian a formal organization is required, with a leadership
parliaments (Duverger 1954 [1951]). Similarly, in the structure and rules governing how these sorts of issues
European Parliament, the political groups formed and will be resolved.
began to act cohesively before the first elections in A solution is a division-of-labour contract, where
1979, and have continued to develop despite almost no back-benchers provide labour and capital (such as
reward or punishment for their activities at the ballot policy expertise), and leaders distribute roles to back-
box. In other words, an explanation of parliamentary benchers (such as positions in committees), commu-
parties must be based on incentives that are largely nicate party positions (issue voting instructions) and

Voting patterns in the European Parliament S. Hix et al. 823
enforce the terms of the party organization contract interests of the member state governments or national
(such as expelling members who do not following the political parties.
instructions) (cf. Kiewiet & McCubbins 1991; Cox & There are three types of votes in the European
McCubbins 1993). Even without external electoral Parliament: a ‘show of hands’ vote; ‘electronic votes’,
incentives, then, party organizations inside parliaments where MEPs press the yes, no or abstain buttons on
reduce transaction costs for politicians and allow their desks and the vote outcome is announced but how
politicians to overcome collective action problems. In each MEP voted is not recorded; and ‘roll-call votes’,
this sense, party organizations in parliaments are where the vote choice of each MEP is reported in the
similar to firms in the labour market (Coase 1937). minutes. Under the European Parliament’s rules, only
However, when observing a high level of party line certain votes are required to be taken by roll-call.
voting, it is often difficult to identify the precise effect of However, a ‘political group’ or a fifth of all MEPs can
the party division-of-labour organization indepen- request any vote to be taken by roll-call. In practice,
dently of the effect of legislators’ preferences. As roughly a third of votes are by roll-call.
Krehbiel (1993) explains: To measure party cohesion in roll-call votes, we use
the Rice (1928) index defined as follows:
In casting apparently partisan votes, do individual
legislators vote with fellow party members in spite of jYi K Ni j
their disagreement about the policy in question, or do they AIi Z ;
Y i C Ni
vote with fellow party members because of their agreement
about the policy question? In the former case,.partisan where Yi denotes the number of yes votes expressed by
behavior may well result in a collective choice that differs group i on a given vote and Ni is the number of no votes.
from that which would occur in the absence of partisan The index equals 1 if all the members of a group vote
behavior. In the latter case, however,.the apparent together and equals 0 if the members are equally
explanatory power of the variable, party, may be divided between these voting options.
attributed solely to its being a good measure of If there is not much variation in the overall majority
preferences. (1993: 238, italics in original). size, party cohesion in one period can easily be
compared with that in another period. However, if
Specifically, if all the members of a party have the
there are large variations in the overall majority size,
same preferences when faced with a choice between a
measuring the absolute level of cohesion is deceptive, as
status quo policy and an alternative proposal, then if
parties will be measured as more cohesive if votes are
the members have the same information about the
‘lopsided’ than if votes are evenly split. The cohesion of
policy implications of the two choices, all the members
the European Parliament as a whole has varied
will decide to vote the same way, whether or not their
considerably since 1979: rising between 1979 and
party has instructed them to do so and threatened to
1987 and then declining until 2004. So, to compare
enforce this instruction.
group cohesion in different periods, Hix et al. (2005,
Nevertheless, even with heterogeneous policy pre-
2007) calculated a ‘relative cohesion’ index by dividing
ferences among its members, a party organization may
the basic cohesion index of a group in a particular
shape policy outcomes by controlling what issues get to
period by the cohesion score for the European
the floor of the chamber (Cox et al. 2000; Amorim Neto
Parliament as a whole in the same period. The resulting
et al. 2003; Cox & McCubbins 2005). If a party leader
score was then divided by 2, so that the relative
has a monopoly on agenda setting, she will only make
cohesion scores were predominantly between 0 and 1.
proposals that will lead to policy outcomes that are
First, looking at the average relative cohesion scores of
closer to her ideal policy than the current policy status
the European political groups (figure 1), transnational
quo. One side effect of this power is that a party leader
party cohesion in the European Parliament is relatively
does not need to work hard to enforce party cohesion,
high and has risen since the late 1980s. By way of
since, if her party is divided on an issue, it is unlikely
comparison, the Democrats and Republicans in the US
that the issue would be put to the chamber. Some bills
Congress have voting cohesion scores around 0.80 while
may be passed with the support of parties that do not
most parties in national parliaments in Europe score
hold agenda-setting power. On these bills, leaders will
above 0.90. Second, as the figure also shows, the
not need their party to act cohesively if there is a
European political groups vote in a more cohesive way
sufficient majority in the chamber in favour of the
than do national groups of MEPs, and the gap between
proposal. In general, though if control of the agenda is
voting as European political groups and voting as national
concentrated in the hands of one party or a coalition of
groups of MEPs has widened since the mid-1980s.
parties, the members of the party (or parties) holding
To understand why party cohesion has risen and
this power are likely to vote the same way on most bills.
what might explain variances between the groups, two
types of analysis can be undertaken (cf. Hix et al.
2007). First, a time-series analysis, looking at changes
3. DETERMINANTS OF POLITICAL GROUP in political group cohesion between different periods.
COHESION IN THE EUROPEAN PARLIAMENT This allows us to look at the effects of the changing
How far do these theories explain the organization and powers of the European Parliament, the enlargement of
development of the political groups in the European the EU, the changing sizes of the political groups and
Parliament? To answer this question, a good starting the internal ideological and national make-up of the
point is to look at how MEPs vote, and how far their groups. Second, a cross-sectional analysis, looking at
voting behaviour is shaped by the policy positions of variances in political group cohesion in all votes in a
the European political groups, as opposed to the given period. This allows us to look at the effects of

1.3 procedure (when the European Parliament is weaker

average of political than the Commission and the EU governments).
1.2 groups Finally, we included eight dummy variables corre-
1.1 sponding to policy areas including economic, environ-
mental, social, external, agricultural, institutional,
cohesion
1.0 internal and budgetary votes.

The results indicate that pivotality and group
0.9
cohesion are positively correlated regardless of the
0.8 average of ideological position of a political group (see table A1 in
national groups appendix A). However, the groups are less cohesive
0.7 when participation is higher. This suggests that the
0.6 groups are less cohesive on high-stakes issues as it is
1979 1984 1989 1994 1999 reasonable to assume that participation is higher on
European Parliament important votes. We also find that cohesion is higher on
whole bills rather than on individual amendments,
Figure 1. Party and national voting cohesion in the European which suggests that on complex high-dimension issues
Parliament in the first five directly elected European MEPs are more likely to follow the instructions of their
Parliaments. The figure shows the average cohesion of the
leaders. For some variables, our analysis leads to mixed
political groups and national groups of MEPs in all the roll-
call votes in each European Parliament between 1979 and results: being on the winning side increase cohesion
2004. The figure was calculated from the data used in Hix only for socialists, conservatives and nationalists. Other
et al. (2007). variables, such as being an agenda setter or requesting a
roll-call are not statistically significant, again indepen-
agenda setting, pivotality, political group size and the dently from MEPs’ ideological preferences.
EU’s legislative procedures. By contrast, a time-series analysis of political group
To illustrate the cross-sectional method, we estimate cohesion suggests that larger political groups are on
a statistical model of the determinants of the relative average more cohesive than smaller political groups,
cohesion scores of the six main groups on each roll- and that the political groups have voted more
call vote in the fifth European Parliament. Three cohesively since the major increase in the powers of
sets of factors are likely to have influenced party the European Parliament in the Maastricht Treaty
cohesion: (i) roll-call vote characteristics, (ii) political (cf. Hix et al. 2005). Meanwhile, ideologically centrist
group characteristics, and (iii) the power of the groups are neither more nor less cohesive than
European Parliament. more extreme groups, and variations in internal
Regarding roll-call vote characteristics, we included ideological or national heterogeneity between the
the following variables: absolute majority; final reading; political groups do not have a significant effect on the
legislative vote; participation; and whole bill. We expect voting cohesion of the groups. Similarly, as the results
cohesion to be higher when an absolute majority of for the second model show, at the political group level,
MEPs is required to pass a bill, when participation is as a group grows bigger its voting cohesion increases,
higher (indicating that the issue is more important for although not significantly so. Nevertheless, while
the MEPs), when a vote is on the final reading of a bill increased ideological heterogeneity does not decrease
and when a vote is on a complex high-dimension issue the voting cohesion of a group, increased national
(for example on a whole bill rather than on an heterogeneity does.
individual amendment). By contrast, we expect cohe-
sion to be lower on a legislative vote (as opposed to a
European Parliament own resolution), when the 4. SPATIAL ANALYSIS OF VOTING IN THE
proposal has been put forward by the Commission EUROPEAN PARLIAMENT
rather than initiated internally by the MEPs. With the new availability of parliamentary voting data
Four dummy variables capture political group on the internet, the growth of computer power and the
characteristics: whether the group was an agenda setter development of new scaling technologies, there has
(if the group proposed the bill or the amendment); been an explosion of research in the last decade on
whether the group requested the roll-call vote; whether the geometric scaling of parliamentary votes (e.g.
the group was pivotal in the vote (if the result of a vote Londregan 2000; Voeten 2000; Schonhardt-Bailey
would have been different if the group had voted on the 2003; Clinton et al. 2004; Morgenstern 2004;
other side); and finally whether the group was on the Rosenthal & Voeten 2004; Poole 2005; Spirling &
winning side. Here, we expect all these factors to McLean 2007). One of the most popular scaling
increase group cohesion. methods, which was developed by Keith Poole and
Regarding the power of the European Parliament, Howard Rosenthal for studying voting in the US
two variables are included: (i) a dummy variable for Congress, is known as NOMINATE (from NOMINAl
votes under co-decision procedure, where political Three-Step Estimation) (Poole & Rosenthal 1985,
stakes are greater as the European Parliament has 1997, pp. 233–51). W-NOMINATE (the non-dynamic
more powers and (ii) a dummy variable for consultation version of the method) estimates the ‘ideal point’ of
procedure. Here we expect cohesion to be higher under each individual parliamentarian as follows.
the co-decision procedure (when the European Parlia- Let s denote the number of policy dimensions
ment has equal legislative power with the EU (kZ1,2, ., s), p denote the number of parliamentar-
governments) and lower under the consultation ians (iZ1,2, ., p) and q denote the number of roll-call

(a) 1.0 (b) (c)
0.5
dimension 2
– 0.5
–1.0
–1.0 – 0.5 0 0.5 1.0 –1.0 –0.5 0 0.5 1.0 –1.0 –0.5 0 0.5 1.0
dimension 1 dimension 1 dimension 1
Figure 2. Revealed ideal point estimates of MEPs in the three European Parliaments (a) the first elected European Parliament,
1979–1984, (b) the third elected European Parliament, 1989–1994, (c) the fifth elected European Parliament, 1999–2004.
These figures are the result of applying the W-NOMINATE geometric scaling metric to all the roll-call votes in each of these
parliaments. Each dot is the estimated location of an MEP on the two main dimensions of voting in a particular European
Parliament. The method is an inductive scaling technique, and as such does not provide an a priori substantive meaning of the
dimensions. The colours in the figures indicate the political groups: red dots, the social democrats; blue dots, the conservatives
and Christian democrats; yellow dots, the liberals; dark red dots, the radical left; light blue dots, the British conservatives and
their allies (who then joined the EPP); navy blue dots, the national conservatives; green dots, the greens and regionalists; purple
dots, the extreme right; pink dots, the anti-Europeans; grey dots, the non-attached MEPs.
votes ( jZ1,2, ., q). Let parliamentarian i’s ideal point weakness of such inductive scaling methods is that
be xi , which is a vector of length s. The ideal point for a they cannot provide a substantive interpretation of the
given parliamentarian is fixed on any given dimension. content of the dimensions. The relative location of the
Call zjy the policy outcome of dimension s, where y political groups in the maps suggests that the first
refers to the policy outcomes associated with a yes vote. dimension in the European Parliament is the left–right,
W-NOMINATE then assumes that parliamentarian i, as the parties seem to be ordered from left to right
who votes sincerely, has a utility function over outcome exactly as one would expect. Meanwhile, the second
y on vote j of dimension might be related to pro-/anti-European
positions, with the pro-EU groups at the top of the
Uijy Z uijy C 3ijy Z bexp d 2ijy C 3ijy ; figures and the more anti-EU groups at the bottom.
Nevertheless, the substantive meaning of the
where u ijy is the deterministic portion of the utility dimensions can be investigated more systematically
function; 3ijy is the stochastic (idiosyncratic or error) by looking at the exogenous determinants the
portion; and the dijy term is the Euclidean distance W-NOMINATE scores. Here, the dependent variable
between xi and zjy. The coefficient b is a constant, is the average score of each national party’s group of
which acts as a signal-to-noise ratio: as b increases, the MEPs on a dimension in a European Parliament. The
deterministic element of the function increases relative unit of analysis is a national party’s group of MEPs, as
to the stochastic element and perfect spatial voting opposed to individual MEPs, because reliable exogen-
results, and as b decreases, voting becomes more ous measures of the policy positions of actors in each
random. The usefulness of outcome n on vote j is European Parliament only exist for national parties,
defined simply by substituting n for y where zjn is and because national parties have a powerful influence
defined accordingly. The stochastic term 3ijy is assumed on the behaviour of their MEPs via the control of
to have an extreme value distribution. The constructed candidate selection in European elections.
likelihood function is then maximized (using a standard There are two main types of independent variables:
algorithm such as Gauss–Newton method) to obtain (i) policy positions of national parties and (ii) the
the parameters of the model: the dimensions of the institutional interests of national parties (whether their
political space and a ‘score’ for each parliamentarian on party is in national government in a particular period or
each recovered dimension. whether their party has a European Commissioner).
Figure 2 shows the results of applying The former influence the policy preferences of the
W-NOMINATE to the roll-call votes in the first, MEPs and the latter influence what types of issues
second and fifth elected European Parliaments. As in MEPs are voting on, as national governments (in the
most other democratic parliaments, voting in the EU Council) and the European Commission are
European Parliament is predominantly one-dimen- the external agenda setters in the EU legislative
sional, and increasingly so. The W-NOMINATE scores process. The national party policy positions are
on dimension 1 correctly predict approximately 85 per taken from political scientists’ estimations of
cent of votes in the first elected European Parliament parties’ positions on two dimensions: left–right, and
and approximately 90 per cent in the fifth elected pro/anti-European integration, scaled between 0 (most
European Parliament, while the scores on dimension 2 left/anti-European) and 1 (most right/pro-European)
only predict an additional 6 per cent in the first (Marks & Steenbergen 2004). For the institutional
Parliament and 2 per cent in the fifth. variables, two dummy variables are included: the
But, what is the meaning of the first and second first coded 1 if a national party was in government
dimensions estimated by W-NOMINATE? One for the majority of a particular European Parliament,

(a) (b)
left–right position of national party
EU integration position of national party
national party is in government
national party has a commissioner

0 0.5 1.0 1.5 2.0 –0.5 0 0.5 1.0
Figure 3. Determinants of national party voting in the European Parliament. The figure plots the coefficients from four OLS
regression models of the average W-NOMINATE score of a national party’s MEPs in a particular European Parliament
((a) dimension 1 scores, (b) dimension 2 scores). Dummy variables for each European Parliament are included in all models.
There are 352 observations (national parties) in each model. The figure plots the coefficients for each variable and the 95%
confidence interval. The full results are reported in table A2 in appendix A. (a) Grey circles, model 1 (adj R 2Z0.696); black
circles, model 2 (model 1C EP group fixed effects; adj. R 2Z0.890); (b) grey circles, model 3 (adj R 2Z0.290); black circles,
model 4 (model 3C EP group fixed effects; adj. R 2Z0.632).
0 otherwise; and the second coded 1 if a national of the second dimension, while MEPs from parties in
party had an EU Commissioner during a particular opposition are located towards the bottom. However,
European Parliament, 0 otherwise. Finally, to examine these institutional interests are not significant once
whether these factors shape MEP behaviour within the party dummies are introduced, which reveals that these
political groups as well as between them, separate institutional interests do not produce voting conflicts
models with dummy variables for each political group within the political groups. Again, as with the first
are estimated. dimension, a large proportion of the variance is
The results are summarized in figure 3 and reported explained by the location of the political groups.
more fully in table A2 in appendix A. As suggested by
the spatial maps (figure 2), MEP scores on the first
5. THE EFFECT OF REPEATED INTERACTIONS
dimension are strongly explained by exogenous left–
AND LOPSIDED VOTES
right policy positions of national parties. In fact, 1 s.d.
So, MEPs increasingly vote along transnational politi-
change along the left–right policy dimension corre-
cal lines rather than national lines, and these European
sponds with a 78 per cent standard deviation change on
political groups are strong determinants of MEP
the first W-NOMINATE dimension. As model 2 behaviour on the two main dimensions of politics in
shows, left–right policy positions also explain variations the European Parliament. What is intriguing is that the
in MEP voting behaviour within the European political European political groups have few powers to enforce
groups. In other words, a national party that has a party line voting. Unlike most national parties, the
policy position to the left (right) of the average member European political groups do not have the power to
of a European political group will be revealed to vote prevent MEPs from standing at the next election if they
slightly to the left (right) of the average member of their do not follow voting instructions. The groups can
group. By contrast, institutional interests of national influence the allocation of office rights inside the
parties do not correlate with scores on the main European Parliament, such as committee positions
dimension of voting. And, once one controls for and rapporteurships (MEPs who prepare a report on the
European party positions, European integration pre- bill and propose amendments). However, even this
ferences are not relevant explanatory factors on the power is shared with national parties, as committee
main dimension of voting in the European Parliament. positions and rapporteurships are allocated to national
Meanwhile, EU integration preferences matter more parties, who then decided which of their MEPs will
than left–right preferences on the second dimension of hold what offices. In other words, the emergence and
politics in the European Parliament. A 1 s.d. change cohesion of the European political groups is largely a
along the European integration policy dimension result of the voluntary and strategic actions of the
corresponds with a 37 per cent standard deviation MEPs and their national parties.
change on the second W-NOMINATE dimension, As the theory of agenda control would suggest, the
whereas a 1 s.d. change on the left–right dimension leaderships of the political groups could promote
corresponds with only a 10 per cent standard deviation collective group action by preventing issues from
change on the second W-NOMINATE dimension. getting to the floor on which the group is internally
Also, MEPs from national parties in government and divided. However, unlike the US Congress, the parties
who have Commissioners are located towards the top in the European Parliament do not control the

(a) 0.30 PES group MEPs (b) 15

EPP group MEPs PES group MEPs
0.25
10
density
0.20
EPP group MEPs
0.15
5
0.10
0.05
0
0 2 4 6 8 10 –0.50 –0.25 0 0.25 0.50
MEP left–right self-placement MEP W-NOMINATE score dimension1
Figure 4. Comparison of posited preferences and revealed behaviour of MEPs in the 2004–2009 European Parliament. (a) A
kernel density plot of the preferences of socialists (PES) and conservatives (EPP) in the sixth elected European Parliament
(posited preferences of MEPs), as measured by the self-placement of the MEPs on a 10-point left–right scale, where 0 equals the
furthest left and 10 equals the furthest right. The data come from a survey of the MEPs that was conducted between March and
June 2006 (see Farrell et al. 2006; and the European Parliament Research Group website: http://www.lse.ac.uk/collections/
EPRG/survey.htm). (b) A kernel density plot of the preferences of the same PES and EPP MEPs in the sixth elected European
Parliament (revealed behaviour of MEPs), as measured by applying W-NOMINATE to all roll-call votes between July 2004 and
December 2006. The correlation between the first dimension W-NOMINATE scores and the left–right self-placement scores in
this period is 0.70.
legislative agenda. In the EU legislative process, the working time directive. A key provision of the
European Commission has the right of initiative and legislation was the removal of a British opt-out from
the EU Council can amend legislation. This means that the existing rules governing the maximum number of
the political groups in the European Parliament are hours certain employees could work. The British
often forced to vote on issues proposed by others. And, Labour government wanted to keep the opt-out, but
even when an issue arrives in the European Parliament, had lost the vote in the EU Council. The Labour
agenda-control rights inside the chamber (such as leadership put pressure on the 19 British Labour MEPs
committee chairs and rapporteurships) are allocated to vote against the directive in the European Parlia-
between the groups on a broadly proportional basis ment. In pure policy-preference terms, the British
rather than monopolized by one particular group. This Labour MEPs agreed with the position of their party
means that a group leadership cannot prevent other leaders in London. However, the national party
groups proposing amendments or requesting a roll-call delegation of MEPs decided to vote with the other
vote on an issue on which the group might be split. In members of the PES group in support of the directive.
fact, the groups vote less cohesively on amendments They did this for two reasons. First, there was a large
proposed by other groups and in roll-call votes called by majority in favour of the legislation (the vote passed by
other groups (Hix et al. 2007, ch. 6). Hence, if a group 378 to 262), so a vote against the bill would have been
or a coalition of groups could control the agenda, they purely symbolic. Second, voting against the PES on
would vote even more cohesively than they do. such an important piece of legislation would have
Nevertheless, the spatial analysis of voting suggests
threatened the future influence of the British Labour
that the European political groups are based on the
MEPs, as the other members of their group would be
preferences of MEPs and national parties about
reluctant to allow Labour MEPs to be rapporteurships or
the direction of social and economic policies at the
committee chairmen in the future.
European level, as captured by the traditional ‘left–
In other words, national parties rather than the
right’ dimension of politics. However, the analysis of
European political groups ultimately control the
group cohesion suggests that declining preference
homogeneity inside the largest groups (as a result of MEPs. However, owing to the sheer volume of
expanding group membership) did not have a negative decisions that have to be made in the European
effect on voting cohesion. How can these results be Parliament, and the potential for unstable and chaotic
reconciled? One possible interpretation is illustrated in outcomes, national parties have an incentive to join a
figure 4. A significant proportion of MEPs in the two political group that broadly shares their policy pre-
largest political groups in the fifth elected European ferences on a range of issues that are likely to come up
Parliament have centrist policy preferences, such as the on the EU agenda, and to delegate organizational and
British Labour MEPs in the Party of European leadership powers to the European political groups, to
Socialists (PES) and the Benelux Christian Democrat share information, allocate agenda-setting rights
MEPs in the EPP (figure 4a). In their voting behaviour, between the member parties and set the broad
however, these centrist MEPs tend to vote with their guidelines of policies. The result is that national parties
European political group colleagues rather than with might be forced to vote against their policy preferences
the MEPs in the other group, who may have closer on some issues, but on average will vote according to
policy preferences. their policy preferences in the knowledge that they are
An anecdote helps to explain why this happens. In more likely to achieve these preferences as their
May 2005, the European Parliament voted on the colleagues in the group will be voting the same way.

Table 2. Vote on the third reading of the Takeover Directives on 4 July 2001. (The political groups and member states are sorted
from the most to the least in favour of the directive in terms of who they voted on the bill. Because the vote was a tie, the directive
was rejected. Using the Rice index to calculate voting cohesion in the European Parliament, the average cohesion score of the
political groups in the vote was 0.56 (not counting the non-attached MEPs), while the average cohesion score of the member
states was 0.63.)
yes no abstain cohesion score
political group liberals (ELDR) 45 0 0 1.00

nationals (UEN) 16 0 0 1.00
non-attached MEPs 16 4 6 0.60
anti-Europeans (EDD) 9 5 3 0.29
social democrats (PES) 80 84 0 0.02
conservatives (EPP-ED) 98 119 4 0.10
greens and regionalists (G/EFA) 8 31 2 0.59
radical left (EUL/NGL) 1 30 7 0.94
member state Denmark 13 0 1 1.00
United Kingdom 72 6 0 0.85
Sweden 19 0 3 1.00
Portugal 19 1 2 0.90
Ireland 10 2 0 0.67
Luxembourg 5 1 0 0.67
Finland 11 2 1 0.69
France 45 26 10 0.27
Italy 32 36 3 0.06
Spain 26 31 1 0.09
The Netherlands 9 22 0 0.42
Belgium 5 16 0 0.52
Austria 4 14 1 0.56
Greece 2 21 0 0.83
Germany 1 95 0 0.98
total 273 273 22
This internal group bargain is self-policing for two The two largest groups collapsed because the
main reasons. First, legislative policy making involves outcome of the vote was highly uncertain—it ended
repeated interactions. So, if a national party breaks in a tied vote, which meant that the legislative was
away from the group on a key issue, they may gain in rejected. Unlike the story of the British Labour MEPs
that particular vote but would risk losing out in the and the working time directive, in such a close vote,
future. Second, in most votes, the outcome is easily every vote decision by a national party was potentially
predictable because the positions of the political groups pivotal rather than purely symbolic. Also, once several
are usually clear before a vote is taken. If a vote is likely national parties signalled that they would vote against
to go a particular way, a national party is unlikely to be the position of their group, the threat of the group
pivotal if they vote against their European political withholding the future benefits declined, because the
group, yet may risk future benefits by doing so. Hence, group leadership would find it difficult to withhold
national parties voluntarily choose to vote with their rapporteurships or committee chairs from a large
European political groups approximately 90 per cent of number of national parties. In this situation, the
the time. national parties were willing to risk the future benefits
However, if a key vote is likely to be close, this to try to secure a short-term policy gain.
second calculation changes. For example, this hap- Interestingly, voting behaviour on close votes as
pened in a high-stakes vote in July 2001 on the takeover opposed to lopsided votes seems to be opposite in the
directive (table 2). This issue—the harmonization of European Parliament and the US Congress. Snyder &
national rules on hostile takeover bids—was certainly Groseclose (2000), for example, looked at lopsided
ideological, with the MEPs on right broadly in favour of votes in the US House of Representatives to measure
proposed new rules and the MEPs on the left broadly the ideological preferences of Congressmen, on the
opposed. As a result, the groups to the right of the EPP assumption that the party caucuses will work hard to
voted cohesively in favour, while the groups to the left enforce party discipline in tight vote, and Congressmen
of the PES voted cohesively against. However, the two will follow instructions in these votes so as not to
largest groups were split down the middle along undermine the electoral appeal of the party. In lopsided
national lines, with the Danish, British, Swedish, votes, by contrast, Congressmen would be freer to vote
Portuguese, Irish, Luxembourg and Finnish MEPs according to their personal policy preferences. In the
overwhelmingly supporting the legislation and the European Parliament, by contrast, where the European
Dutch, Belgian, Austrian, Greek and German MEPs Parliament elections have very little to do with the
overwhelmingly opposing it. performance of the European political groups, the

European political groups cannot enforce cohesion in information. Once this division of labour has been
close votes. One might think that MEPs, like US established, which involves policy specialization and
Congressmen, would be also be free to vote how they the division of agenda-setting powers inside the
like in a lopsided vote, as the group knows the political group, MEPs are likely to follow voting
outcome. However, an MEP has an interest to keep instructions from their political group leaders. They
her policy preferences private in these votes, as do this because they reasonably expect that these
revealing a difference of opinion with her colleagues instructions are the positions they would come to if
on an important policy issue might lead her they had the time and resources to work out their
colleagues not to trust her to hold a key agenda- position on the complex legislative issue. On some
setting office on a future issue. Hence, the fluidity issues, an MEP may receive information (e.g. from
of the allocation of agenda-setting rights in the interest groups) that their preferences in a particu-
European Parliament, via rapporteurships and com- lar issue are different from those being commu-
mittee assignments, as opposed to the more stable nicated by their political group, and the frequency
and monopolized system of agenda-setting rights in of this conflicting information has probably
the US Congress, provides an incentive for the increased as the European Parliament’s power has
MEPs to vote collectively in lopsided votes.
increased. However, an MEP is often unlikely to
follow these alternative instructions because the
chances of being pivotal in a vote are usually small,
6. CONCLUSION and because voting against her group would send a
The most persuasive explanation of the formation negative signal about the preferences of the MEP
and operation of the political groups in the relative to the group. If the other members of a
European Parliament is that they are vehicles for group believe the MEP has variant preferences from
the promotion of the policy preferences of MEPs them, then they will cease to trust the information
and their national political parties. MEPs who share the MEP is sharing on the issues on which she has
preferences on a range of issues on the EU policy specialized knowledge and will prefer to allocate
agenda have an incentive to establish a division- internal party agenda-setting rights to the members
of-labour contract, to share the costs of collecting who are closer to the average party member.
APPENDIX A
Table A1. Determinants of political group cohesion. (The dependent variable is the relative Rice index for a political group in a
vote. Robust t-statistics in parentheses. p%0.05, p%0.01. Policy area variables (including economic, social, agricultural,
institutional, internal and budgetary variables) are included in the estimations but their coefficients are not reported.)
PES EPP-ED ELDR UEN G/EFA EUL/NGL
majority size K0.289 K0.367 K0.220 K0.375 K0.392 K0.372

(15.63) (19.49) (14.15) (17.00) (24.28) (19.97)
absolute majority K0.014 0.030 K0.034 0.013 K0.006 0.018
(1.69) (3.11) (3.84) (1.14) (0.79) (2.03)
final reading 0.012 0.031 0.017 0.007 0.014 0.016
(1.10) (2.64) (1.66) (0.47) (1.49) (1.23)
whole bill 0.020 0.030 0.000 0.019 0.022 0.023
(2.94) (4.30) (0.02) (2.08) (3.36) (3.15)
legislative K0.080 K0.073 K0.007 0.042 0.030 K0.047
(2.95) (1.90) (0.22) (0.91) (0.65) (0.86)
agenda setter K0.005 K0.006 K0.009 K0.014 K0.002 0.015
(0.99) (0.99) (1.28) (0.40) (0.20) (1.35)
party calling RCV 0.017 0.012 0.003 0.017 0.007 0.012
(2.12) (2.05) (0.44) (1.24) (1.41) (1.46)
party is winner K0.036 K0.027 0.005 0.028 K0.002 0.006
(5.54) (4.01) (0.93) (4.12) (0.39) (1.15)
party is pivotal 0.098 0.043 0.181 0.176 0.232 0.227
(17.11) (7.79) (26.16) (9.61) (38.71) (26.76)
participation K0.060 K0.067 K0.044 K0.142 K0.068 K0.080
(4.07) (4.10) (3.13) (6.43) (4.84) (4.68)
coKdecision 0.056 0.037 K0.004 K0.025 K0.023 0.060
(2.11) (0.97) (0.14) (0.55) (0.49) (1.10)
consultation 0.044 K0.001 K0.016 K0.030 K0.076 0.003
(1.70) (0.03) (0.55) (0.67) (1.64) (0.06)
constant 0.880 0.927 0.772 0.887 0.962 0.913
(44.70) (45.74) (42.36) (33.90) (54.30) (42.88)
observations 5163 5163 5160 5082 5159 5159
R-squared 0.20 0.14 0.27 0.10 0.35 0.25


Table A2. Determinants of national party voting in the European Parliament. (The dependent variable is the average position of a national party’s group of MEPs on dimension 1 or 2 in a
particular European Parliament (5-year period). The positions of the MEPs on the dimensions are produced by applying the W-NOMINATE method of ideal point estimation to all the roll-
call votes in each European Parliament. The parameters are estimated by linear OLS regression. Standard errors in parentheses. p%0.05, p%0.01. Dummy variables for each European
We are grateful to the editors and three anonymous referees
0.863
0.927
1.000
1.000
for their very helpful comments on an earlier version of
max
this paper.
1
K0.903
K0.953
0.000
0.000
variable summary statistics
min REFERENCES
Aldrich, J. H. 1995 Why parties? The origin and transformation
0
0.490 of political parties in America. Chicago, IL: The University
0.449
0.229
0.277
0.486
0.407
of Chicago Press.
s.d.
Amorim Neto, O., Cox, G. W. & McCubbins, M. D. 2003

Agenda power in Brazil’s Câmara dos Deputados,
1989–98. World Polit. 55, 550–578. (doi:10.1353/wp.
K0.055
2003.0023)
0.018
0.503
0.716
0.381
0.216
mean
Clinton, J., Jackman, S. & Rivers, D. 2004 The statistical

analysis of roll call data. Am. Polit. Sci. Rev. 98, 355 – 370.
Coase, R. H. 1937 The nature of the firm. Economica 4,
386 – 405. (doi:10.1111/j.1468-0335.1937.tb00002.x)
K0.457 (0.103)
Cox, G. W. 1987 The efficient secret: the cabinet and the

K0.257 (0.104)
development of political parties in Victorian England.

0.174 (0.081)
0.036 (0.034)
0.017 (0.042)
dim.1 scores:
dim.2 scores:
Cambridge, MA: Cambridge University Press.

Cox, G. W. & McCubbins, M. D. 1993 Legislative leviathan:
party government in the house. Berkeley, CA: University of
0.632
Yes
California Press.
352
No
(4)
Cox, G. W. & McCubbins, M. D. 2005 Setting the agenda:

responsible party government in the US house of representa-
tives. Cambridge, MA: Cambridge University Press.
dimension 2 scores
K0.575 (0.095)
Cox, G. W., Masuyama, M. & McCubbins, M. D. 2000 Agenda

K0.206 (0.089)
0.601 (0.079)
0.129 (0.047)
0.253 (0.053)
power in the Japanese house of representatives. Jpn J. Polit.

Sci. 1, 1–21. (doi:10.1017/S1468109900000116)
Duverger, M. 1954 [1951] Political parties: their organization
and activities in the modern state. New York, NY: Wiley.
0.290
Farrell, D., Hix, S., Johnson, M. & Scully, R. 2006 EPRG

352
No
No
(3)
2000 and 2006 MEP surveys dataset. London, UK: London

School of Economics and Political Science.
Hix, S. & Marsh, M. 2007 Punishment or Protest? Under-
standing European Parliament elections. J. Polit. 69,
K0.317 (0.062)
K0.033 (0.049)
K0.006 (0.025)
0.826 (0.062)
495 –510. (doi:10.1111/j.1468-2508.2007.00546.x)

0.019 (0.020)
Hix, S., Noury, A. & Roland, G. 2005 Power to the parties:

cohesion and competition in the European Parliament,
1979–2001. Br. J. Polit. Sci. 35, 209 –234. (doi:10.1017/
0.890
Yes
S0007123405000128)
352
No
(2)
Hix, S., Noury, A. & Roland, G. 2007 Democratic politics in the

European Parliament. Cambridge, MA: Cambridge
University Press.
dimension 1 scores
Kiewiet, D. R. & McCubbins, M. D. 1991 The logic of

K0.893 (0.062)
1.686 (0.053)
0.193 (0.056)
delegation: congressional parties and the appropriations process.

0.052 (0.033)
0.064 (0.038)
Chicago, IL: University of Chicago Press.

Krehbiel, K. 1993 Where’s the party? Br. J. Polit. Sci. 23,
235 –266.
0.696
Kreppel, A. 2002 The European Parliament and supranational

352
No
No
(1)
party system: a study in institutional development.

score of a national party’s group of MEPs in
dependent variable: average W-NOMINATE
LaPalombara, J. & Weiner, M. 1966 The origin and

Parliament are included but not reported.)
development of political parties. In Political parties and

EU integration position of national party
political development (eds J. LaPalombara & M. Weiner),

pp. 3–42. Princeton, NJ: Princeton University Press.
left–right position of national party
national party has a commissioner
Londregan, J. B. 2000 Legislative institutions and ideology in

national party is in government
Chile. Cambridge, MA: Cambridge University Press.

Marks, G. & Steenbergen, M. 2004. Marks/Steenbergen party
political group fixed effects
member state fixed effects
dataset. See http://www.unc.edu/wgwmarks/data.htm.

a European Parliament
McKelvey, R. D. 1976 Intransitivities in multidimensional

voting models and some implications for agenda control.
adjusted R-squared
J. Econ. Theory 12, 472– 482. (doi:10.1016/0022-0531

(76)90040-5)
observations
Morgenstern, S. 2004 Patterns of legislative politics.

constant
Noury, A. & Roland, G. 2002 European Parliament: should it

have more power? Econ. Policy 17, 279 – 319. (doi:10.
1111/1468-0327.t01-1-00090)

Poole, K. T. 2005 Spatial models of parliamentary voting. Schlesinger, J. A. 1984 On the theory of party organization.
Cambridge, MA: Cambridge University Press. J. Polit. 46, 369 – 400. (doi:10.2307/2130967)
Poole, K. T. & Rosenthal, H. 1985 A spatial model for Schonhardt-Bailey, C. 2003 Ideology, party and interests in
legislative roll call analysis. Am. J. Polit. Sci. 29, 357– 384. the British Parliament of 1841–47. Br. J. Polit. Sci. 33,
(doi:10.2307/2111172) 581–605. (doi:10.1017/S0007123403000267)
Poole, K. T. & Rosenthal, H. 1997 Congress: a political– Snyder Jr, J. M. & Groseclose, T. 2000 Estimating party
economic history of roll call voting. Oxford, UK: Oxford influence in congressional roll-call voting. Am. J. Polit. Sci.
University Press. 44, 193 –211. (doi:10.2307/2669305)
Raunio, T. 1997 The European perspective: transnational party Spirling, A. & McLean, I. 2007 UK OC OK? Interpreting
groups in the 1989–1994 European Parliament. London, optimal classification scores for the UK house of
UK: Ashgate.
commons. Polit. Anal. 15, 85 – 96. (doi:10.1093/pan/
Rice, S. A. 1928 Quantitative methods in politics. New York,
NY: Knopf. mpl009)
Riker, W. H. 1980 Implications for the disequilibrium of van der Eijk, C. & Franklin, M. (eds) 1996 Choosing
majority rule for the study of institutions. Am. Polit. Sci. Europe? The European electorate and national politics
Rev. 74, 432– 446. (doi:10.2307/1960638) in the face of union, Ann Arbor, MI: University of
Rosenthal, H. & Voeten, E. 2004 Analyzing roll calls with Michigan Press.
perfect spatial voting. Am. J. Polit. Sci. 48, 620–632. Voeten, E. 2000 Clashes of the assembly. Int. Organ. 54,
(doi:10.1111/j.0092-5853.2004.00091.x) 185 –214. (doi:10.1162/002081800551154)

Phil. Trans. R. Soc. B (2009) 364, 833–843

doi:10.1098/rstb.2008.0259
Behavioural social choice: a status report

Michel Regenwetter1,*, Bernard Grofman2, Anna Popova1, William Messner1,
Clintin P. Davis-Stober1 and Daniel R. Cavagnaro3
1
University of Illinois at Urbana-Champaign, Champaign, IL 61820, USA
2
University of California at Irvine, Irvine, CA 92697, USA
3
Ohio State University, Columbus, OH 43210, USA
Behavioural social choice has been proposed as a social choice parallel to seminal developments in
other decision sciences, such as behavioural decision theory, behavioural economics, behavioural finance
and behavioural game theory. Behavioural paradigms compare how rational actors should make certain
types of decisions with how real decision makers behave empirically. We highlight that important
theoretical predictions in social choice theory change dramatically under even minute violations of
standard assumptions. Empirical data violate those critical assumptions. We argue that the nature of
preference distributions in electorates is ultimately an empirical question, which social choice theory
has often neglected. We also emphasize important insights for research on decision making by
individuals. When researchers aggregate individual choice behaviour in laboratory experiments to
report summary statistics, they are implicitly applying social choice rules. Thus, they should be aware
of the potential for aggregation paradoxes. We hypothesize that such problems may substantially mar
the conclusions of a number of (sometimes seminal) papers in behavioural decision research.
Keywords: behavioral social choice; decision making; voting paradoxes
1. RECONCILING THE SEGREGATED DECISION important exception is the fact that social choice
SCIENCES theory has systematically incorporated rational utility
The decision sciences are currently segregated into theory as a theoretical primitive. Another noteworthy
nearly disparate research areas. On the one hand, exception is the literature on justice and fair division
researchers study individual choice; that is, decision (Balinski & Young 1982; Schokkaert & Lagrou 1983;
making at the level of the individual decision maker. On Kahneman et al. 1986; Brams & Taylor 1996;
the other hand, another research community studies Schokkaert & Devooght 2003; Konow in press). As
social choice; that is, aggregate decision making at the Regenwetter et al. (2007a) have emphasized, major
level of groups or societies, especially in the form of progress in the decision sciences may hinge on the
voting. These two research communities, individual ability of the various disparate communities to integrate
and social choice researchers, by and large associate their collective wisdoms and develop new synergies.
with different scientific societies and publish in The need for a unified framework to the decision
different journals. Another important distinction is sciences is indicated by the large dotted box in figure 1.
that between normative, i.e. rational theories of choice, Several major movements have arisen, which
which satisfy certain theoretically motivated optimality respond to the largely normative tone of prior theory
criteria, and behavioural, i.e. descriptive theories that in the decision sciences. Behavioural counterparts of
describe or explain empirically observed choice normative theories (e.g. behavioural decision research,
behaviour. Figure 1 shows these conceptual distinc- experimental economics, behavioural finance, beha-
tions along two major axes. Different research para- vioural game theory and behavioural social choice,
digms fall into different sections of the implied 2!2 see also the glossary at the end of the paper) have
table. While many paradigms in the decision sciences a decades-long tradition of contrasting normative
do not fit squarely in a single spot, we have nonetheless proposals, such as expected utility theory, against
attempted to place several important paradigms empirical human choice data.
in their most pertinent locations in the table. For We provide a status report on behavioural social
instance, utility theory, such as expected utility theory choice research, and discuss the facilitating role this
(von Neumann & Morgenstern 1947; Savage 1954), is paradigm can play in establishing a broader and
the normative theory of rational individual decision more unified decision sciences research programme.
making under uncertainty or risk. The paper is organized as follows. In §2, we review
past work showing that the famous Condorcet paradox
There has been limited cross-fertilization between
of majority cycles may have limited behavioural
individual and social choice research areas. The most
support. In §3, we discuss the fact that behavioural
decision research in the individual choice domain
* Author for correspondence (regenwet@illinois.edu). routinely uses aggregation and therefore must
One contribution of 11 to a Theme Issue ‘Group decision making in become attuned to social choice paradoxes in order to
humans and animals’. avoid artefacts caused by unsound data aggregation.

834 M. Regenwetter et al. Behavioural social choice
normative and descriptive

all (weakly or) linearly ordered actor preferences
prescriptive theory and among some set of objects are taken to be equally
theory empirical data likely. This literature asks how often we must expect to
find intransitive social preferences and concludes that
the paradox should be ubiquitous (DeMeyer & Plott
utility theory and
Heuristics and biases 1970; Gehrlein & Fishburn 1976; Riker 1982; Gehrlein
individual decision analysis
1983; Lepelley 1993; Jones et al. 1995; Van Deemen
choice behavioural decision research 1999; Mueller 2003).
behavioural economics The second major literature on the Condorcet
paradox provides theorems that show sufficient con-
behavioural social choice
ditions to avoid cycles and reveals that these conditions
finance, game theory, appear to be highly restrictive (Sen 1966, 1970). Two
mechanism design and fads, stock bubbles major themes of both these literatures are (i) the
collective social choice theory theoretical prediction that majority elections should be
and/or plagued with cycles and (ii) the broadly advertised
experimental economics
interactive policy recommendation (Shepsle & Bonchek 1997) to
choice behavioural game theory beware the use of majority rule in real elections because
a majority winner is unlikely to even exist.
Figure 1. A unified framework to the decision sciences.
Regenwetter & Grofman (1998a) and Regenwetter
et al. (2002b) found virtually no empirical evidence for
the Condorcet paradox in survey or ballot data.
Section 4 provides a new result on the sampling Therefore, in a series of publications (Regenwetter &
properties of social choice rules when dealing with Grofman 1998b; Regenwetter et al. 2002a,c, 2003;
the preference distributions other than the symmetric Tsetlin & Regenwetter 2003; Tsetlin et al. 2003) that
distributions we have coined ‘cultures of indifference’ culminated in a Cambridge University Press book
(see glossary for a definition). Section 5 illustrates a ‘Behavioral social choice’ (Regenwetter et al. 2006),
behavioural social choice analysis. In §6, we propose a members of the present team of authors, and others,
research paradigm that expands the notions of re-examined the arguments leading to the belief that
‘Condorcet efficiency’ and ‘Borda efficiency’ (see the Condorcet paradox should be an inevitable
glossary for definitions). concomitant of any majority rule voting process.
They showed that the existing results, while math-
ematically correct, were nonetheless misleading. For
2. THE EMPIRICAL RARITY OF THE CONDORCET example, they found that simulation results were based
PARADOX on ‘knife-edge’ theoretical assumptions, where even
Social choice theory (Arrow 1951; Black 1958; Sen minuscule deviations from the theoretical assumptions
1970; Gehrlein & Fishburn 1976; Riker 1982; lead to dramatic changes in predictions. Similarly, they
Tangiane 1991; Saari 1995; Mueller 2003) has had as found that the theoretical sufficiency conditions for
its principal concerns the axiomatic structure (i.e. the avoiding the Condorcet paradox primarily tell us what
abstract mathematical properties) of voting rules and ‘cannot be ruled out under all possible circumstances’
social welfare functions, including impossibility results rather than providing realistic evaluations of the threat
(see glossary for definitions). A major concern in social posed by the Condorcet paradox. Regenwetter et al.
choice theory has been the problem of intransitivity, i.e. (2003) stated abstract and yet empirically plausible
a cyclical situation where there exist three choice sufficient conditions to avoid the paradox, and found
alternatives such that the first is socially preferred to empirical evidence in survey and ballot data that the
the second, the second is socially preferred to the third, conditions they had identified as sufficient to avoid the
yet, the third is socially preferred to the first. Condorcet paradox were satisfied (or sufficiently nearly
Intransitive cycles are often labelled a ‘social choice satisfied). Recent experimental work on deliberative
paradox’. Social choice theory has generated many polls suggests potential explanations of how delibera-
estimates of the degree of intransitivity created by/ tive democracy may avoid the Condorcet paradox (List
inherent in the aggregation of individual preferences et al. 2007).
into collective decisions. At the heart of much of this
work is the Condorcet paradox of cyclical majorities
(see glossary for a simple example). In a cyclical
3. THE IMPORTANCE OF BEHAVIOURAL AND
majority, no matter which candidate is elected,
NORMATIVE SOCIAL CHOICE THEORY FOR
a majority of voters will be disappointed because they
BEHAVIOURAL INDIVIDUAL DECISION
would prefer someone else to be the chosen candidate.
RESEARCH
Perhaps even more importantly, cyclical majorities
We now turn to the role that social choice theory should
seem to cast into doubt the very notion of meaningful
(but currently does not) play in individual behavioural
majority decision making (Riker 1982).
decision research. Consider the following three binary
There are two important literatures dealing directly
non-negative gambles:
with the Condorcet paradox. The first is based on
analytic or simulation results that look at theoretical (
90% chance to win $73;
distributions. The most common assumption is the gamble A : ð3:1Þ
impartial culture (see glossary), a distribution in which 10% chance to win $90:50;

Behavioural social choice M. Regenwetter et al. 835
Table 1. Three examples of CPT predictions for gambles A, B, C.
parameter values V(A) V(B) V(C) preference order
CPT(0.32, 0.98) gZ0.32, aZ0.98 69.37 69.00 68.83 ABC

CPT(0.56, 0.97) gZ0.56, aZ0.97 67.06 87.90 69.95 BCA
CPT(0.67, 0.50) gZ0.67, aZ0.50 8.71 7.80 8.80 CAB
(
85% chance to win $20; choice of parameters g and a, we will not be able to
gamble B : ð3:2Þ accommodate the empirical cycle using equation (3.6).
15% chance to win $385;
This is because (3.6) implies transitive preferences.
(
85% chance to win $65; Thus, it seems as if CPT could not explain the
gamble C : ð3:3Þ hypothetical cycle in our example.
15% chance to win $130: A prominent recently proposed decision heuristic
Imagine that individual decision makers in a laboratory tackles this problem. The priority heuristic (PH, see
experiment make pairwise choices among these glossary) of Brandstätter et al. (2006) theorizes that
gambles. Suppose that most participants choose decision makers compare gambles via a process that
A over B, most choose B over C and yet, most also induces a lexicographic interval order (see glossary for a
choose C over A. This kind of cyclical pattern of definition). In a nutshell, decision makers sequentially
choices poses a challenge to standard decision theories (‘lexicographically’) consider three attributes (the
and it has motivated recent prominent developments of so-called ‘reasons’) and an aspiration level for each
heuristic decision theories, i.e. theories of decision reason. They visit the reasons in a specific order and
making by ‘computationally simple rules of thumb’. stop their decision process whenever an aspiration level
For illustrative purposes consider cumulative pro- is met for the given reason currently under consider-
spect theory (CPT, see glossary; Tversky & Kahneman ation (Brandstätter et al. 2006). For the three gambles
1992; Wakker & Tversky 1993). For non-negative above, the PH predicts that the decision maker chooses
gambles such as A, B, C in (3.1)–(3.3), CPT trans- A over B (by reason 1). The PH also predicts that the
forms each probability p of an outcome via a probability decision maker will choose B over C (by reason 3).
weighting function w, say, However, the PH predicts that the decision maker
chooses C over A (by reason 3). Clearly, the PH
pg accounts for 100 per cent (all three) of the pairwise
wðpÞ Z ; for some 0! g% 1; ð3:4Þ
ðpg C ð1K pÞg Þ1=g majority choices.
In the terminology of Brandstätter et al. (2006), the
and each gamble outcome x via a utility function v, say,
PH is able to capture perfectly ‘the process’ by which
of the form
the decision makers arrived at their final choices. Using
vðxÞ Z x a ; for some 0! a% 1: ð3:5Þ a similar approach, Brandstätter et al. (2006) argued
that the PH is superior to several leading decision
The probability weighting function w, depending on theories because it models the cognitive process of
the value of g, inflates low probabilities (i.e. predicts decision making and, compared with these competing
risk-seeking behaviour when probabilities are low, theories, it correctly predicts the largest number of
such as in a lottery) and deflates high probabilities modal pairwise choices in several datasets from the
(i.e. predicts risk avoidance when probabilities are literature (Kahneman & Tversky 1979; Tversky &
high). The utility function v, depending on a, inflates Kahneman 1992; Lopes & Oden 1999; I. Erev et al.
relatively small gains and deflates relatively large gains.
2002, unpublished manuscript). Because the modal
The biasing of probabilities and utilities in CPT is
choice among a pair of gambles is also the majority
based on a large empirical literature that has reported
choice, these conclusions are, in fact, based on
cognitive biases and limitations in humans when
descriptive analyses within a pairwise majority aggrega-
dealing with choice under conditions of risk
tion approach.
or uncertainty.
Now, let us return to the imaginary decision
For a binary non-negative gamble f, writing p1 for
experiment. Before, we reported the data in the usual
the probability of winning the smaller amount x 1, and
aggregated fashion that one often finds in the
p2 for the probability of winning the larger amount x 2,
behavioural decision literature that studies individual
let p2 Z wð p2 Þ and p1 Z 1Kwð p2 Þ. In CPT, for binary
choice. However, if we consider the data in more detail,
non-negative gambles g, h,
an interesting new picture emerges. Suppose the data
g is preferred to h5 V ð gÞO V ðhÞ; came from three decision makers (DM 1, DM 2 and
X 2 ð3:6Þ DM 3) who made the combinations of choices shown
where V ð f Þ Z pi vðxi Þ; f 2 f g; hg: in table 2. Note that each decision maker acted
i Z1
in accordance with CPT using (3.6) and using one of
The two functions w and v depend on two parameters, the parameter choices in table 1. Most importantly, not
g and a. Table 1 gives examples of parameter values, a single decision maker chose in accordance with the
implied values V( f ) for the above three prospects A, B PH. But majority aggregation, a popular method for
and C, and the implied preference order among summarizing choice data ( Tversky 1969; Kahneman &
gambles, from the best to the worst. Regardless of the Tversky 1979; Tversky & Kahneman 1981, 1986;

Table 2. Three decision makers acting by CPT, with a two-third majority supporting PH.
two-third majority
gamble pair DM 1 DM 2 DM 3 choice
A versus B A B A A
B versus C B B C B
C versus A A C C C
compatible theory CPT(0.32, 0.98) CPT(0.56, 0.97) CPT(0.67, 0.50) priority heuristic
Birnbaum 2004; Brandstätter et al. 2006), obscures In this section, we show that the sampling
this fact. By majority, the PH alone is able to approaches to studying the behaviour of scoring rules
accommodate 100 per cent of the (majority choice) in large electorates are extremely dependent on
‘data’, while CPT, the theory according to which we underlying theoretical assumptions. This provides
computed each choice, accommodates at best two- additional motivation as to why behavioural analyses
thirds of majority choices. We are facing a Condorcet of social choice procedures are critical for our under-
paradox, where the pattern of majority choices does not standing of their real-world performance. Recall that
match the choice pattern of even one individual the most famous distributional assumption in social
decision maker. choice theory is the impartial culture assumption,
Any aggregation of choice data could create artefacts according to which an electorate can be considered to
in the analysis of decision-making behaviour. Aggrega- be a random sample with replacement from a uniform
tion, especially by majority, is common in individual distribution over linear (or weak) orders on the set of
behavioural decision-making research, including in candidates. According to this assumption, if one
seminal papers (Tversky 1969; Kahneman & Tversky randomly samples a voter from the population and if
1979; Tversky & Kahneman 1981, 1986; Brandstätter there are n candidates, then all of the n! possible orders
et al. 2006). This means that much past research in of candidates are equally likely to match the preference
behavioural decision research is susceptible to aggrega- of that voter.
tion paradoxes such as the Condorcet paradox. This is Consider the impartial culture from a statistical view
a reason why behavioural decision researchers should point. When computing Condorcet’s majority rule, or
systematically incorporate social choice theoretical any scoring rule (see glossary for definitions), such as
considerations into their work. plurality or Borda, at the distribution level of the
impartial culture, we obtain a perfect tie among all
candidates. Yet, random samples of any size, if they
4. SAMPLE SOCIAL CHOICE OUTCOMES AS contain an odd number of voters, will reproduce that
ESTIMATORS OF POPULATION SOCIAL majority tie among any two candidates with probability
CHOICE OUTCOMES zero, when voters are assumed to have linear-order
We now proceed to a new, yet simple, result in social preferences. Condorcet’s majority relationship is not a
choice theory. We explain the behaviour of scoring rules consistent estimator of the population majority
in samples from nearly any conceivable kind of culture relationship when samples may originate from knife-
(population distribution) with the only caveat that it edge distributions (see glossary for a definition) similar
deviates in one crucial way from the cultures that have in nature to the impartial culture.
dominated the discussion historically. Consider any social choice method, such as
While statistics has played a major role in social Condorcet’s majority rule, or a scoring rule, aka
choice theory in the guise of sampling distributions positional voting method. A culture of indifference (with
derived from various theoretical distributions, such as regard to that procedure) is a probability distribution over
the impartial culture (see glossary for the definition), preference relationships (linear orders, weak orders
inferential statistics does not seem to be used system- (WOs) and partial orders) with the property that one or
atically. Traditionally, social choice theorists have rarely more pairs of candidates are tied at the distribution
considered the need to draw statistical inferences from level, according to that social choice method. When
empirical data about underlying population properties considering majority rule, for example, a culture of
of social choice functions in a given electorate. Yet, ever indifference is any probability distribution over binary
since the close call in the 2000 US presidential election, relationships of any kind, such that for at least some
it has become clear that published ballot counts are distinct pair of candidates, A, B, the total probability of
not a deterministic function of the distribution of those preference relationships in which A is preferred to
preferences in a population. Rather there are many B equals the total probability of those preference
probabilistic components that affect turnout, ballot relationships in which B is preferred to A. Regenwetter
casting and ballot counting. Two of the present authors et al. (2006) have shown that the majority rule
have studied this notion for more than a decade. outcomes of large electorates (drawn from an under-
Regenwetter et al. (2006) and the component papers lying theoretical culture) will converge (with increasing
that were published earlier, as well as Regenwetter & size of the electorate) to the majority preferences in the
Tsetlin (2004) and Regenwetter & Rykhlevskaia underlying culture as long as that culture is not a culture
(2007) have promoted the need to consider social of indifference. The purpose of this section is to show
choice data from an inferential statistical point of view. the analogous result for scoring rules.

Cultures of indifference are knife-edge assumptions Proof. Since the random variables ðX A i ÞnZ1;.;N are
that lead to chronically paradoxical behaviour (in the i.i.d. with finite mean and variance, the weak law of
statistical and social choice sense) of social choice rules large numbers implies that S An converges in probability
in random samples. Unfortunately, much work on the to SA. Hence, for any e, dO0 there exists Nd;e 2 N
likelihood of voting paradoxes has hinged on cultures of such that
indifference. This has created a common perception
that voting paradoxes are extremely likely. We consider Pr jS An K SA jO e ! d; for nO Nd;e ;
this a profound and far-reaching artefact of unrealistic
theoretical modelling assumptions. for any candidate A 2 S:
Let S be a finite set of candidates. We define a
probability space (B, F , P), where B is a family of order Pick any pair of candidates A, B2S. Since we assume
relationships on S (e.g. linear orders, WOs or, say, that there are no ties in the population social order,
asymmetric and acyclic binary relationships), F Z 2S we assume without loss of generality that SAOSB. Pick
the power set of S and P : F 1 ½0; 1 is a culture, dO0 and eO0, such that e! jSA K SB j. Then
and, in particular, a population probability distribution A
n ! Pr jS n KSA j C jS n KSB jO jSA KSB j ;
Pr S Bn R S A B
over B. A scoring rule takes any preference relation
from B and gives a numerical score to each candidate in ! Pr jS An KSA jO 3 AND
S. Regenwetter & Rykhlevskaia (2007) have developed
jS Bn KSB jO ðjSA KSB jK3Þ ;
general scoring rules for a general class of binary
relationships, using the notion of a generalized rank of % Pr jS An KSA jO 3 ! d for nO Nd;3 :
Regenwetter & Rykhlevskaia (2004).
More formally, a scoring rule is a set of functions Thus, the sample social order converges to the
f fA : A 2 Sg, where fA : B/ R and fA(R) is the score population social order. &
given to candidate A for the preference relationship
In summary, first, we have seen that, in cultures of
R. For a given scoring rule, we define the random
indifference, social choice rules display highly irregular
variable X A i to be the score of candidate A for the ith behaviour in the sense that they need not be consistent
draw in a sequence of independent and identically estimators of culture social orders. Yet, second, in
distributed (i.i.d.) draws from a population with cultures that are not cultures of indifference (i.e. nearly
distribution P. That is, for R2B, any culture one could think of ), large electorates will
XA display exactly the same social choice behaviour as has
i ðRÞ Z fA ðRÞ:
been theoretically assumed at the level of the culture to
The sample score for candidate A is the average score of begin with. This holds for majority rule (Regenwetter
A in a sample of size n. It is denoted by et al. 2006) and, as we have proved here, for scoring
X n rules. Note also that we have allowed individual
XA
SA
n Z i
: preferences to be binary relationships of any kind, not
i Z1
n just weak or linear orders.
For a sample of size n, whenever S A B
n O S n , we say that
A is socially ordered ahead of B in the sample, by the
given scoring rule. The resulting order by any given 5. EIGHT BEHAVIOURAL SOCIAL CHOICE
scoring rule for a given sample is called the sample social ANALYSES
order according to that scoring rule. The population score We now expand recent developments in behavioural
of a candidate A is the expectation of X A i . We write social choice along the lines of Regenwetter et al.
(2007b). Those authors analysed four large sets of
X jSj
SA Z E X A
i Z fA ðRi ÞPðRi Þ: empirical ballot data from the 1998–2001 annual
i Z1 presidential election ballots of the American Psycho-
logical Association (APA), that were collected under
Whenever SAOSB, we say that A is socially ordered
the Hare system. Each election featured on the order of
ahead of B in the population, by the given scoring rule.
20 000 voters and a rather politicized electorate.
The order prescribed by a scoring rule for the entire
Regenwetter et al. (2007b) analysed these ballots
population is called the population social order of that
using a series of different models and using bootstrap
scoring rule.
methods for statistical inference.
The following result shows that if the population
Table 3 illustrates some key points of such a
social order has no ties, then S A n converges in
behavioural social choice analysis, similar to that of
probability to SA as n grows arbitrarily large. Hence,
Regenwetter et al. (2007b). First, since social choice
SAn is a consistent estimator of SA. We conclude, as
outcomes can depend very heavily on theoretical
a consequence, that the sample social order converges
assumptions, we carry out empirical analysis using at
to (is a consistent estimator of ) the population
least two sets of fundamentally different assumptions
social order whenever the population is not a culture
about the nature of preferences and about the vote
of indifference.
casting process. Second, we evaluate the statistical
Theorem 4.1. If the population social order of a given replicability of our findings. Third, we contrast the
scoring rule has no ties, then the sample social order of that famed theoretical incompatibility of competing social
scoring rule converges to the population social order as the choice procedures with a high degree of agreement
sample size increases. among methods in empirical data.

Table 3. Behavioural social choice analysis of 1998–2005 APA presidential election data, each with five candidates and tens of
thousands of voters. (We report results from the WO model (WO) and the Zwicker model (ZW). The WO results for 1998–2001
have appeared in Regenwetter et al. (2007b), all others are new. Social orders with bootstrapped confidence R98% are in italics.
The order 31524 indicates that candidate 3 is the winner, followed by candidates 1, 5 and 2, whereas 4 is the loser.)
1998 1999 2000 2001 2002 2003 2004 2005
Condorcet
WO 32145 (0.9) 43215 (0.6) 52134 53124 (0.7) 54213 (0.7) 42531 41532 21453
ZW 32415 (0.8) 43215 (0.8) 52134 51324 54123 (0.9) 42531 41532 21453 (0.7)
Borda
WO 32145 43125 (0.9) 52134 53124 54213 42531 41532 21453 (0.9)
ZW 32415 (0.6) 43215 (0.9) 52134 51324 54123 42531 41532 21453 (0.8)
plurality
WO 35124 (0.7) 43152 (0.8) 53214 (0.7) 53124 54213 (0.8) 42351 41352 12543
ZW 31524 43125 (0.5) 52314 (0.9) 53124 54213 42351 41352 12543 (0.02)
Regenwetter et al. (2007b) included the WO analysis the empirical absence of a cycle. Third, note that
of the 1998–2001 data. The WO model assumes that all plurality, which uses the least information in the
ranked candidates are preferred to all non-ranked ballots, comes with sometimes extremely low statistical
candidates on the ballot, and that the voter is indifferent confidence, i.e. even small changes in the ballot
among all candidates s/he does not rank. In the distribution can affect the social order. Fourth,
table, we have added four new datasets (2002–2005) table 3 suggests that there is a fair degree of agreement
and a new model which we call the Zwicker model among the three voting rules. In particular, in nearly
(Dr W. Zwicker 2006, personal communication). every case where the three social orders can be
The ballot data are partial ranking counts. Zwicker estimated with high confidence, they yield identical
suggested interpreting the data as follows: count A as winners and identical losers. This stands in direct
strictly preferred to B if and only if both options have contrast with the literature that predicts very sub-
been ranked and A has been ranked as preferable to B. stantial disagreements among the three rules. Besides
The Zwicker model (ZW) does not assume any the absence of a cycle, this is another important
preference among any pair of candidates of which one divergence from common wisdom in social choice
or both were left unranked. Thus, ZW translates partial theory. We discuss this more directly next, with a
rankings into strict partial orders. For scoring rules, the special focus on the agreement about the winner.
general results of (Regenwetter & Rykhlevskaia 2004,
2007) allow us to assign appropriate scores to all 6. GENERALIZATIONS OF CONDORCET
candidates from every ballot, even to those candidates EFFICIENCY AND BORDA EFFICIENCY
that have not been listed in the voter’s partial ranking. Building on research about the Condorcet paradox, a
We have drawn 10 000 random samples, with highly sophisticated, and often quite technical, literature is
replacement, of sample size equal to the original ballot concerned with the Borda and Condorcet efficiencies of
count, from a hypothetical population distribution voting methods. For instance, the Condorcet efficiency
(culture) estimated via either the WO model or the of a voting method is the conditional probability
Zwicker model. In each sampled set of ballots, we have that the election winner matches the Condorcet
computed the Condorcet, Borda and plurality out- winner in a random sample of ballots (from some
comes. This is a non-parametric bootstrap of the theoretical distribution), provided that there exists a
confidence we can have in the empirical social welfare Condorcet winner. More generally, this literature studies
outcomes under the three rules. The bootstrap is a way the interrelationship among social choice rules
to simulate possible sources of uncertainty in election (Chamberlin & Cohen 1978; Gehrlein & Fishburn
outcomes, such as unreliabilities in turnout, ballot 1978; Riker 1982; Merrill 1984, 1985; Bordley 1985;
casting and ballot counting. Intuitively speaking, it Gehrlein 1985, 1992; Merrill & Nagel 1987; Nurmi 1988,
shows how sensitive the final tally is to small 1992; Adams 1997; Gehrlein & Lepelley 2000; Merlin
perturbations in the ballot distribution. et al. 2000; Mueller 2003). A large part of this literature
For each dataset, and for each model, we report concentrates on cultures of indifference. This literature
the modal social welfare order, as well as its approxi- predicts that many standard and competing voting
mate bootstrapped confidence. When the confidence procedures disagree with one another a substantial part
exceeds 98 per cent, then we leave out the value, and of the time. A related theoretical and empirical literature
simply display that social order in italics. First, in all studying variants of the Condorcet jury theorem
eight elections, and independently of the model, we (Grofman 1981; Miller 1986; List & Goodin 2001),
avoid the Condorcet paradox with confidence near however, avoids cultures of indifference. The empirical
100 per cent (the table omits some details). Second, we literature that compares social choice procedures against
find some degree of model dependence regarding the each other (Yaari & Bar-Hillel 1984; Felsenthal et al. 1986,
exact nature of the social orders. For example, the 1993; Rapoport et al. 1988; Leining 1993; Felsenthal &
Condorcet order by WO and ZW differ in 1998, 2001 Machover 1995; Hastie & Kameda 2005; Tideman 2006)
and 2002, due to tight pairwise margins for plurality. is small, by and large it avoids considerations of statistical
Nonetheless, this model dependence has no bearing on inference, and it usually considers sparse datasets.

Table 4. Probability of agreement on the winner by Condorcet, Borda and plurality. (As benchmarks, we report simulation based
probabilities for the impartial cultures over WOs on five candidates (WO5) and for a uniform distribution on partial rankings of
five candidates (PR5). The remaining quantities are the bootstrapped (simulated sampling) probabilities using the 1998–2005
APA presidential election data. Confidences above 98% are in italics.)
benchmark empirical data based
model WO5 PR5 1998 1999 2000 2001 2002 2003 2004 2005
agreement among Condorcet and Borda winners

WO 0.68 0.65 O0.99 0.99 O0.99 O0.99 O0.99 O0.99 O0.99 0.96
ZW 0.54 O0.99 0.89 O0.99 O0.99 O0.99 O0.99 O0.99 0.89
agreement among Condorcet and plurality winners
WO 0.51 0.43 O0.99 0.79 O0.99 O0.99 O0.99 O0.99 O0.99 0.02
ZW 0.39 O0.99 0.76 O0.99 O0.99 O0.99 O0.99 O0.99 0.7
agreement among Borda and plurality winners
WO 0.59 0.51 O0.99 0.79 O0.99 O0.99 O0.99 O0.99 O0.99 0.05
ZW 0.56 O0.99 0.93 O.99 O0.99 O0.99 O0.99 O0.99 0.67
agreement among Condorcet, Borda and plurality winners
WO 0.46 0.38 O0.99 0.79 O0.99 O0.99 O0.99 O0.99 O0.99 0.02
ZW 0.34 O0.99 0.84 O0.99 O0.99 O0.99 O0.99 O0.99 0.63
Arrow’s (1951) famous impossibility theorem can cultures of indifference, large samples will have high
be interpreted to mean that any choice of a consensus agreement among social choice rules if and only if the
method comes at the cost of giving up principles that cultures (population distributions) themselves have
underlie other, competing, and mathematically not social orders that agree. This is a direct consequence of
universally compatible voting methods. Saari (1994, the fact that the Condorcet procedure and all scoring
1995, 1999) has shown that one can create distri- rules are consistent estimators of the corresponding
butions that yield virtually any combination of population social orders, whenever we are not drawing
differences in results across voting methods. He has from a culture of indifference. Clearly, we face another
developed an algorithm to specify such distributions situation where theoretical predictions are direct
precisely. We propose a straightforward extension of the functions of the underlying theoretical assumptions:
study of Condorcet efficiency: what is the probability in in cultures other than cultures of indifference, if the
random samples from known cultures, and what is the culture features agreement among Condorcet and/or
inferred (e.g. bootstrapped) population confidence scoring rules, then large samples will replicate that
based on empirical data, that any two or more social agreement with probability converging to 1. Even
choice procedures, e.g. Condorcet and some scoring though margins may be small in some empirical
rules, agree on (i) the winner, (ii) the loser, (iii) the elections that involve heavy campaigning, we do not
entire social order? Which social choice rules appear to consider cultures of indifference to be realistic rep-
be in heavy empirical disagreement, and which appear resentations of real-world electorates.
to be highly consistent in most empirical settings? What Our findings show how crucial empirical work will
be in untangling the puzzle surrounding the question of
characteristics of the empirical distribution appear to
agreement or disagreement among social choice
drive the agreement and/or disagreement among
procedures. Axiomatic theory is, by and large, mute
competing social welfare functions?
about population distributions (i.e. about what are or
While much more work is needed to support any
are not suitable assumptions to make about underlying
general claims, we have some early indication that this
cultures that generate ballot frequencies). The nature
approach will reveal more puzzles about social choice.
of preference distributions in electorates is ultimately
Table 4 shows, as benchmarks, our simulated agree- an empirical question. We hope that this paper will
ment probabilities for the impartial culture on WOs encourage the social choice community to augment
over five candidates (WO5) and for the uniform their traditionally normative theoretical work with a
distribution on partial rankings of five candidates behavioural analysis component. However, we would
(PR5). These benchmarks suggest that one should like to emphasize that empirical work should be carried
not have high hopes of two or even all three among out in a fashion that is statistically sound. Published
Condorcet, Borda and plurality yielding the same ballot counts are not deterministic functions of the
winner for five candidates. The table compares these underlying preference distributions and must be
benchmarks to agreement probabilities we derived by subjected to adequate statistical inferential methods.
bootstrapping from the empirical ballot data. Our
analysis in table 4 suggests that in six out of the eight
ballots-based distributions, the corresponding prob- 7. CONCLUSION AND DISCUSSION
abilities virtually equal 100 per cent. Behavioural social choice research has collected
A large proportion of the theoretical literature is evidence that Condorcet cycles are surprisingly rare
based on cultures of indifference, where sample social in empirical survey and ballot data. This work has
choice functions are not consistent estimators of the leveraged statistical inference as a major methodo-
population social orders. Once we move away from logical tool. Behavioural social choice theory has also

highlighted the role of model dependence, i.e. the fact ballot counts. Ultimately, it falls upon empirical
that conclusions about social choice procedures can researchers to discover the properties of real-world
hinge on the theoretical assumptions that enter distributions of preferences in real populations, and to
theorems, simulations or statistical analyses of empiri- characterize the conditions under which competing
cal data. Nonetheless, while inferred distributions of social choice rules agree or disagree with each other.
preferences in electorates often depend on theoretical Here, we have highlighted a behavioural social
assumptions in the analysis, the empirical absence of choice approach to understanding the empirical and
majority cycles has been extremely robust across a statistical properties of preference aggregation and
range of modelling assumptions. The same holds for voting methods. We think that this paradigm can be
the agreement among competing voting methods. usefully extended to other domains. For example, there
In this paper, we have reviewed how various is a growing literature on statistical properties of belief
branches of the decision sciences are nearly completely aggregation methods that builds on the Condorcet jury
disparate. Furthermore, we firmly believe that major theorem in much the same way that early social choice
advances could be possible if the different ‘constitu- work built on the Condorcet criterion for aggregating
encies’ of the decision sciences were consolidated. For preferences (Black 1958; Grofman et al. 1983). Much
example, we have highlighted how the standard early work in this area uses very strong assumptions
research paradigm of individual behavioural decision about statistical independence, while some recent work
research routinely relies on social choice aggregation uses Nash–Bayesian assumptions that seem to us
of individual choice data, often without regard to behaviourally implausible ( List & Goodin 2001;
possible social choice paradoxes. While it is too early to Dryzek & List 2003).
tell, this practice could permeate the empirical Science often proceeds in a two-stage process where
literature with artefacts. empirical work and theoretical work go hand-in-glove,
In the social choice domain, we have discussed a each inspiring the other in an upwardly spiralling
new but straightforward result about the statistical ladder of knowledge. This is our hope for behavioural
nature of scoring rules. When sampling from a culture social choice.
that is not a culture of indifference, (sufficiently) large
electorates’ social order, by any scoring rule, will This material is based upon work supported by the Air Force
match (with probability arbitrarily close to one) the Office of Scientific Research, Cognition and Decision Program,
social order by the same scoring rule found in the under Award no. FA9550-05-1-0356 entitled ‘Testing
Transitivity and Related Axioms of Preference for Individuals
underlying culture. Our previous work has shown that and Small Groups’ (to M.R., PI ) and by the National
the same is true for the Condorcet criterion. This has Institute of Mental Health under Training Grant Award no.
major implications for the famed disagreement among PHS 2 T32 MH014257 entitled ‘Quantitative methods for
social choice rules. Whether Condorcet and/or various behavioral research’ (to M.R., PI ). Grofman’s contributions
scoring rules, such as Borda, agree or disagree with to this research were supported by the Institute for
each other in large samples will completely depend on Mathematical Behavioral Sciences and the Center for the
the assumptions made about the underlying culture. Study of Democracy at the University of California, Irvine.
D. Cavagnaro carried out this work while a NIH postdoctoral
In cultures of indifference, which have received a
trainee at the University of Illinois at Urbana-Champaign.
disproportionate amount of attention, all of these Any opinions, findings and conclusions or recommendations
scoring rules display pathological sampling behaviours expressed in this publication are those of the authors and do
because, in this case, the sample social orders are not not necessarily reflect the views of the Air Force Office of
consistent estimators of the population social orders. Scientific Research, the National Institute of Mental Health,
For example, with just two candidates and an the Insitute for Mathematical Behavioral Sciences, or the
impartial culture, the social order in the impartial Center for the Study of Democracy.
culture is a two-way tie by Condorcet and by every
scoring rule. Yet, for samples of any size and for
asymmetric and weakly complete pairwise individual REFERENCES
preferences (i.e. when each decision maker strictly Adams, J. 1997 Condorcet efficiency and the behavioral
prefers one or the other among the two choice model of the vote. J. Polit. 59, 1252–1263. (doi:10.2307/
alternatives), if the sample size is odd, the probability 2998600)
that the sample social order matches the social order Arrow, K. J. 1951 Social choice and individual values. New York,
of the underlying population is zero, regardless of NY: Wiley.
sample size. Balinski, M. L. & Young, H. P. 1982 Fair representation:
Behavioural social choice analyses have now meeting the idea of one man, one vote. New Haven, CT: Yale
revealed in a number of datasets that competing social University Press.
orders appear to be in nearly perfect agreement with Birnbaum, M. H. 2004 Causes of Allais common conse-
quence paradoxes: an experimental dissection. J. Math.
each other, often with high statistical confidence. This
Psychol. 48, 87–106. (doi:10.1016/j.jmp.2004.01.001)
suggests that realistic cultures should not be cultures of
Black, D. 1958 The theory of committees and elections.
indifference, and that the theoretical literature may Cambridge, UK: Cambridge University Press.
promote overly pessimistic views about the likelihood Bordley, R. 1985 A precise method for evaluating election
of consensus among consensus methods. Axiomatics schemes. Public Choice 46, 113–123. (doi:10.1007/
highlight that competing methods cannot universally BF00179734)
agree with each other. Simulation results, as we have Brams, S. J. & Taylor, A. D. 1996 Fair division: from cake
shown, will completely hinge on the assumptions made cutting to dispute resolution. New York, NY: Cambridge
about the generating distribution that underlies the University Press.

Brandstätter, E., Gigerenzer, G. & Hertwig, R. 2006 The List, C., Luskin, R., Fishkin, J. & McLean, I. 2007
priority heuristic: making choices without trade-offs. Deliberation, single-peakedness, and the possibility of
Psychol. Rev. 113, 409–432. (doi:10.1037/0033-295X. meaningful democracy: evidence from deliberative polls.
113.2.409) Working paper, London School of Economics and
Chamberlin, J. & Cohen, M. 1978 Toward applicable social Stanford University.
choice theory—a comparison of social choice functions Lopes, L. L. & Oden, G. C. 1999 The role of aspiration level
under spatial model assumptions. Am. Polit. Sci. Rev. 72, in risky choice: a comparison of cumulative prospect
1341–1356. (doi:10.2307/1954543) theory and SP/A theory. J. Math. Psychol. 43, 286–313.
DeMeyer, F. & Plott, C. R. 1970 The probability of a cyclical (doi:10.1006/jmps.1999.1259)
majority. Econometrica 38, 345–354. (doi:10.2307/ Merlin, V., Tataru, M. & Valognes, F. 2000 On the
1913015) probability that all decision rules select the same winner.
Dryzek, J. & List, C. 2003 Social choice theory and J. Math. Econ. 33, 183–207. (doi:10.1016/S0304-
deliberative democracy: a reconciliation. Br. J. Polit. Sci. 4068(99)00012-9)
33, 1–28. (doi:10.1017/S0007123403000012) Merrill, S. I. 1984 A comparison of efficiency of multi-
Felsenthal, D. S. & Machover, M. 1995 Who ought to be candidate electoral systems. Am. J. Polit. Sci. 28, 23–48.
elected and who is actually elected—an empirical investi- (doi:10.2307/2110786)
gation of 92 elections under 3 procedures. Elect. Stud. 14, Merrill, S. I. 1985 A statistical model for Condorcet
143–169. (doi:10.1016/0261-3794(94)00019-6) efficiency based on simulation under spatial model
Felsenthal, D. S., Maoz, Z. & Rapoport, A. 1986 Comparing assumptions. Public Choice 47, 389–401. (doi:10.1007/
voting systems in genuine elections: approval voting vs BF00127534)
selection-plurality. Soc. Behav. 1, 41–53. Merrill, S. I. & Nagel, J. 1987 The effect of approval balloting
Felsenthal, D. S., Maoz, Z. & Rapoport, A. 1993 An on strategic voting under alternative decision rules. Am.
empirical evaluation of 6 voting procedures—do they Polit. Sci. Rev. 81, 509–524. (doi:10.2307/1961964)
really make any difference? Br. J. Polit. Sci. 23, 1–27. Miller, N. R. 1986 Information, electorates, and democracy:
Gehrlein, W. V. 1983 Concorcet’s paradox. Theory Decis. 15, some extensions and interpretations of the Condorcet jury
161–197. (doi:10.1007/BF00143070) theorem. In Information pooling and group decision making
Gehrlein, W. V. 1985 Condorcet efficiency of constant scoring (eds B. Grofman & G. Owen), pp. 173–192. Greenwich,
rules for large electorates. Econ. Lett. 19, 13–15. (doi:10. UK: JAI Press.
1016/0165-1765(85)90093-X) Mueller, D. C. 2003 Public choice III. Cambridge, UK:
Cambridge University Press.
Gehrlein, W. V. 1992 Condorcet efficiency of simple voting
Nurmi, H. 1988 Discrepancies in the outcomes resulting
rules for large electorates. Econ. Lett. 40, 61–66. (doi:10.
from different voting schemes. Theory Decis. 25, 193–208.
1016/0165-1765(92)90245-T)
(doi:10.1007/BF00134159)
Gehrlein, W. V. & Fishburn, P. C. 1976 The probability of the
Nurmi, H. 1992 An assessment of voting system simulations.
paradox of voting: a computable solution. J. Econ. Theory
Public Choice 73, 459–487. (doi:10.1007/BF01789562)
13, 14–25. (doi:10.1016/0022-0531(76)90063-6)
Rapoport, A., Felsenthal, D. & Maoz, Z. 1988 Proportional
Gehrlein, W. V. & Fishburn, P. C. 1978 Probabilities of
representation: an empirical evaluation of single-stage,
election outcomes for large electorates. J. Econ. Theory 19,
non-ranked voting procedures. Public Choice 59, 151–165.
38–49. (doi:10.1016/0022-0531(78)90054-6)
(doi:10.1007/BF00054451)
Gehrlein, W. V. & Lepelley, D. 2000 The probability that all
Regenwetter, M. & Grofman, B. 1998a Approval voting,
weighted scoring rules elect the same winner. Econ. Lett. Borda winners and Condorcet winners: evidence from
66, 191–197. (doi:10.1016/S0165-1765(99)00224-4) seven elections. Manag. Sci. 44, 520–533. (doi:10.1287/
Grofman, B. 1981 Mathematical models of juror and jury mnsc.44.4.520)
decision making: the state of the art. The trial processes, Regenwetter, M. & Grofman, B. 1998b Choosing subsets:
vol. II, pp. 305–351. New York, NY: Plenum Press. a size-independent probabilistic model and the quest for a
Grofman, B., Owen, G. & Feld, S. L. 1983 Thirteen social welfare ordering. Soc. Choice Welfare 15, 423–443.
theorems in search of the truth. Theory Decis. 15, (doi:10.1007/s003550050115)
261–278. (doi:10.1007/BF00125672) Regenwetter, M. & Rykhlevskaia, E. 2004 On the (numeri-
Hastie, R. & Kameda, T. 2005 The robust beauty of majority cal) ranking associated with any finite binary relation.
rules in group decisions. Psychol. Rev. 112, 494–508. J. Math. Psychol. 48, 239–246. (doi:10.1016/j.jmp.2004.
(doi:10.1037/0033-295X.112.2.494) 03.003)
Jones, B., Radcliff, B., Taber, C. & Timpone, R. 1995 Regenwetter, M. & Rykhlevskaia, E. 2007 A general concept
Condorcet winners and the paradox of voting: probability of scoring rules: general definitions, statistical inference,
calculations for weak preference orders. Am. Polit. Sci. and empirical illustrations. Soc. Choice Welfare 29,
Rev. 89, 137–144. (doi:10.2307/2083080) 211–228. (doi:10.1007/s00355-006-0204-1)
Kahneman, D. & Tversky, A. 1979 Prospect theory: an Regenwetter, M. & Tsetlin, I. 2004 Approval voting and
analysis of decision under risk. Econometrica 47, 263–291. positional voting methods: inference, relationship,
(doi:10.2307/1914185) examples. Soc. Choice Welfare 22, 539–566. (doi:10.1007/
Kahneman, D., Knetsch, J. L. & Thaler, R. 1986 Fairness as a s00355-003-0232-z)
constraint on profit seeking entitlements in the market. Regenwetter, M., Adams, J. & Grofman, B. 2002a On the
Am. Econ. Rev. 76, 728–741. (sample) Condorcet efficiency of majority rule: an
Konow, J. In press. Is fairness in the eye of the beholder? An alternative view of majority cycles and social homogeneity.
impartial spectator analysis of justice. Soc. Choice Welfare. Theory Decis. 53, 153–186. (doi:10.1023/A:1021215
Leining, W. 1993 The fatal vote: Bonn versus Berlin. 903030)
Finanzarchiv 1, 1–20. Regenwetter, M., Grofman, B. & Marley, A. A. J. 2002b On
Lepelley, D. 1993 Concorcet’s paradox. Theory Decis. 15, the model dependence of majority preferences recon-
161–197. structed from ballot or survey data. Math. Soc. Sci.
List, C. & Goodin, R. E. 2001 Epistemic democracy: (Special Issue on Random utility theory and probabilistic
generalizing the Condorcet jury theorem. J. Polit. Philos. measurement theory) 43, 451–466. (doi:10.1016/S0165-
9, 277–306. (doi:10.1111/1467-9760.00128) 4896(02)00022-7)

Regenwetter, M., Grofman, B. & Marley, A. A. J. 2002c Van Deemen, A. 1999 The probability of the paradox of
A general concept of majority rule. Math. Soc. Sci. voting for weak preference orderings. Soc. Choice Welfare
(Special Issue on Random utility theory and probabilistic 16, 171–182. (doi:10.1007/s003550050138)
measurement theory) 43, 405–428. (doi:10.1016/S0165- von Neumann, J. & Morgenstern, O. 1947 Theory of games
4896(02)00021-5) and economic behavior. Princeton, NJ: Princeton University
Regenwetter, M., Marley, A. A. A. J. & Grofman, B. 2003 Press.
General concepts of value restriction and preference Wakker, P. P. & Tversky, A. 1993 An axiomatization of
majority. Soc. Choice Welfare 21, 149–173. (doi:10.1007/ cumulative prospect theory. J. Risk Uncertain. 7, 147–176.
s00355-003-0210-5) (doi:10.1007/BF01065812)
Regenwetter, M., Grofman, B., Marley, A. & Tsetlin, I. 2006 Yaari, M. E. & Bar-Hillel, M. 1984 On dividing justly. Soc.
Behavioral social choice. Cambridge, UK: Cambridge Choice Welfare 1, 1–24. (doi:10.1007/BF00297056)
University Press.
Regenwetter, M., Ho, M.-H. & Tsetlin, I. 2007a Sophis-
ticated approval voting, ignorance priors, and plurality
heuristics: a behavioral social choice analysis in a GLOSSARY
Thurstonian framework. Psychol. Rev. 114, 994–1014.
Behavioural economics, behavioural finance and behavioural
(doi:10.1037/0033-295X.114.4.994)
game theory: these fields differ from their traditional
Regenwetter, M., Kim, A., Kantor, A. & Ho, M.-H. 2007b
counterparts by integrating realistic psychological factors,
The unexpected empirical consensus among consensus
such as bounded rationality and cognitive biases, into
methods. Psychol. Sci. 18, 559–656. (doi:10.1111/j.1467-
their models.
9280.2007.01950.x)
Riker, W. H. 1982 Liberalism v. populism. San Fransisco, CA: Borda efficiency: a voting method’s Borda efficiency is the
W. H. Freeman and Co. likelihood that the election winner matches the Borda
Saari, D. G. 1994 Geometry of voting. New York, NY: (scoring rule) winner, when one exists, in a random
Springer. sample of ballots (typically from an impartial culture).
Saari, D. G. 1995 Basic geometry of voting. Berlin, Germany; Borda winner: a candidate who receives the highest score total
New York, NY: Springer. in a contest where each voter ballot consists of a complete
Saari, D. G. 1999 Explaining all three-alternative voting ranking of the candidates, and where each candidate
outcomes. J. Econ. Theory 87, 313–355. (doi:10.1006/jeth. scores as many points from a given ballot as there are
1999.2541) candidates with a worse rank position on that ballot. The
Savage, L. J. 1954 The foundations of statistics. New York, NY: Borda score has also been extended to situations where the
Wiley. ballots provide binary relationships (e.g. partial orders),
Schokkaert, E. & Devooght, K. 2003 Responsibility-sensitive not necessarily full rankings of the candidates (Regenwetter &
fair compensation in different cultures. Soc. Choice Welfare Rykhlevskaia 2007).
21, 207–242. (doi:10.1007/s00355-003-0257-3) Cognitive bias: a pattern of deviation in judgement that occurs
Schokkaert, E. & Lagrou, L. 1983 An empirical approach to in certain situations. This deviation is relative to people
distributive justice. J. Public Econ. 21, 33–52. (doi:10. outside of the particular situation. For example, a framing
1016/0047-2727(83)90072-5) bias occurs when a change in the wording of a survey
Sen, A. K. 1966 A possibility theorem on majority decisions. question leads to a systematic change in the responses to
Econometrica 34, 491–499. (doi:10.2307/1909947) that question.
Sen, A. K. 1970 Collective choice and social welfare. San Condorcet efficiency: a voting method’s Condorcet efficiency is
Fransisco, CA: Holden-Day. the likelihood that the election winner matches the
Shepsle, K. & Bonchek, M. 1997 Analyzing politics: Condorcet (majority) winner, when one exists, in a
rationality, behavior and institutions. New York, NY: random sample of ballots (typically from an impartial
Norton. culture).
Tangiane, A. S. 1991 Aggregation and representation of Condorcet paradox: also called the ‘paradox of cyclical
preferences: introduction to mathematical theory of democracy. majorities’, the Condorcet paradox occurs when group
Berlin, Germany: Springer. preferences are not transitive even though individual
Tideman, N. 2006 Collective decisions and voting. Williston,
preferences are transitive. The simplest example involves
VT: Ashgate.
three voters and three alternatives. Label the alternatives
Tsetlin, I. & Regenwetter, M. 2003 On the probability
A, B and C, and assume that voter 1 has preference order
of correct or incorrect majority preference relations.
ABC, voter 2 has preference order BCA and voter 3 has
Soc. Choice Welfare 20, 283–306. (doi:10.1007/s003550
preference order CAB. Then majorities prefer A to B, and
200182)
Tsetlin, I., Regenwetter, M. & Grofman, B. 2003 The B to C and C to A.
impartial culture maximizes the probability of majority Condorcet winner: a candidate who beats each other candidate
cycles. Soc. Choice Welfare 21, 387–398. (doi:10.1007/ in pairwise contest. A Condorcet winner need not always
s00355-003-0269-z) exist; see ‘Condorcet paradox’.
Tversky, A. 1969 Intransitivity of preferences. Psychol. Rev. Culture of indifference: with regard to a given voting procedure,
76, 31–48. (doi:10.1037/h0026750) a culture of indifference is a probability distribution over
Tversky, A. & Kahneman, D. 1981 The framing of decisions preference relationships (linear orders, WOs, partial
and the psychology of choice. Science 211, 453–458. orders or other binary relationships) with the property
(doi:10.1126/science.7455683) that one or more pairs of candidates are tied, at the
Tversky, A. & Kahneman, D. 1986 Rational choice and the distribution level, according to that voting method.
framing of decisions. J. Bus. 59, 251–278. (doi:10.1086/ Cumulative prospect theory: prospect theory is a descriptive
296365) model of decision making between risky prospects. It was
Tversky, A. & Kahneman, D. 1992 Advances in prospect developed by Kahneman & Tversky (1979) as a realistic
theory: cumulative representation of uncertainty. J. Risk alternative to expected utility theory. Cumulative prospect
Uncertain. 5, 297–323. (doi:10.1007/BF00122574) theory (CPT ) was put forth by Tversky & Kahneman (1992)

as a theoretical improvement of the original prospect theory. The order of elements is determined by sequentially
The main observations underlying CPTare that people have comparing each attribute. The first attribute in which two
different risk attitudes towards gains and losses, caring more elements differ determines the order between those two
about potential losses than potential gains, and that people elements. The name comes from its generalization of the
tend to overweight extreme but unlikely outcomes. ordering of words in a dictionary. A word (i.e. a sequence
Descriptive theory: a theory that describes how people actually of letters) a1, a2, ., ak comes before another word b1,
behave in some situations. b2, ., bk if and only if ai comes before bi in the alphabet,
Experimental economics: the use of laboratory experiments to where i is the first index for which ai is different from bi.
test the empirical validity of various economic theories. Mechanism theory: the study of how the rules of a game
Game theory: a branch of applied mathematics that aims to influence the behaviour of the game’s participants, under
capture decision-making behaviour in situations where an certain assumptions of rationality. Mechanism design is
individual’s success in making a choice depends on the the application of mechanism theory for the purpose of
choices made by others. designing a set of rules to encourage a specified pattern of
Heuristics: simple, efficient rules that have been proposed to behaviour, such as truth telling or participation.
explain how people make decisions, come to judgements, Normative theory: a theory that describes how rational agents
or solve problems, typically when facing complex pro- should behave in some situations.
blems or incomplete information. Heuristics often involve Positional voting method (aka, scoring rule): a voting method in
educated guesses, rules-of-thumb and common sense. which candidates are awarded points based on their
Impartial culture: a population consisting of equal numbers of position in the ranking on each ballot, and the candidate
voters holding each possible linear or WO preference over with the most total points is the winner. The Borda count
candidates. An impartial culture implies a complete is an example of a positional voting method. Positional
majority tie among all candidates at the population level, voting methods have been extended to situations where
i.e. it is an example of a culture of indifference. the ballots provide binary relationships (e.g. partial
Impossibility results: a class of theorems showing that no social orders), not necessarily full rankings of the candidates
choice rule can simultaneously satisfy a given set of (Regenwetter & Rykhlevskaia 2007).
reasonable criteria. See, Arrow’s (1951) famous ‘impossi- Priority heuristic: descriptive model of decision making put
bility theorem’. forth by Brandstätter et al. (2006). The model posits a
Knife-edge distribution: refers to the situation in which small decision-making heuristic in which decision makers who
disturbances to an assumed probability distribution in a face a choice between two gambles will sequentially
model can yield dramatically different analytical and consider three attributes (minimum gain, probability of
simulation results. minimum gain, maximum gain) on which they compare
Lexicographic interval order: a lexicographic order with the the gambles. The choice among two choice gambles is
additional property that the categories of each attribute are terminated as soon as an attribute yields an unequivocal
interval ordered. That is to say, for each attribute, there is a preference, beyond some pre-specified aspiration level,
mapping from the set of categories of that attribute to a set among the two choice alternatives.
of closed intervals on the real line with the following Social choice theory: the study of rules for aggregating
property: category A is preferred to category B whenever individual preferences to form a collective preference.
the left-hand endpoint of A’s interval is strictly to the right Transitivity: in mathematics, a binary relationship is said to be
of the right-hand endpoint of B’s interval. This type of transitive if whenever an element x is related to an element
order relationship is noteworthy because it can generate an y, and y is in turn related to an element z, then x is also
intransitive preference order from intuitively reasonable related to z. In social science, transitivity is often applied
transitive preferences within each separate attribute. to preferences. To have transitive preferences, a person or
Lexicographic order: a preference ordering applied to a set of group who prefers A over B and B over C, must also prefer
elements differing on two or more ordered attributes. A over C.

Phil. Trans. R. Soc. B (2009) 364, 845–852

doi:10.1098/rstb.2008.0224
Speed versus accuracy in decision-making ants:

expediting politics and policy implementation
Nigel R. Franks*, François-Xavier Dechaume-Moncharmont†,
Emma Hanmore and Jocelyn K. Reynolds
School of Biological Sciences, University of Bristol, Woodland Road, Bristol BS8 1UG, UK
Compromises between speed and accuracy are seemingly inevitable in decision-making when accuracy
depends on time-consuming information gathering. In collective decision-making, such compromises
are especially likely because information is shared to determine corporate policy. This political process
will also take time. Speed–accuracy trade-offs occur among house-hunting rock ants, Temnothorax
albipennis. A key aspect of their decision-making is quorum sensing in a potential new nest. Finding a
sufficient number of nest-mates, i.e. a quorum threshold (QT), in a potential nest site indicates that many
ants find it suitable. Quorum sensing collates information. However, the QT is also used as a switch,
from recruitment of nest-mates to their new home by slow tandem running, to recruitment by carrying,
which is three times faster. Although tandem running is slow, it effectively enables one successful ant to
lead and teach another the route between the nests. Tandem running creates positive feedback; more and
more ants are shown the way, as tandem followers become, in turn, tandem leaders. The resulting corps
of trained ants can then quickly carry their nest-mates; but carried ants do not learn the route. Therefore,
the QT seems to set both the amount of information gathered and the speed of the emigration. Low QTs
might cause more errors and a slower emigration—the worst possible outcome. This possible paradox of
quick decisions leading to slow implementation might be resolved if the ants could deploy another
positive-feedback recruitment process when they have used a low QT. Reverse tandem runs occur after
carrying has begun and lead ants back from the new nest to the old one. Here we show experimentally
that reverse tandem runs can bring lost scouts into an active role in emigrations and can help to maintain
high-speed emigrations. Thus, in rock ants, although quick decision-making and rapid implementation
of choices are initially in opposition, a third recruitment method can restore rapid implementation after a
snap decision. This work reveals a principle of widespread importance: the dynamics of collective
decision-making (i.e. the politics) and the dynamics of policy implementation are sometimes intertwined,
and only by analysing the mechanisms of both can we understand certain forms of adaptive organization.
Keywords: context-dependent behaviour; collective behaviour; group movement; recruitment;
emigration
1. INTRODUCTION in house-hunting ants. For ants choosing new homes,

From first-hand experience, and from first principles, faster decisions are more error-prone (Marshall et al.
one can deduce a mutual antagonism between speed 2006; Pratt & Sumpter 2006). However, these earlier
and accuracy in many decisions. Accuracy may require studies do not tell the full story. Indeed, they leave a
extra information, but gathering, processing and central issue unresolved.
sharing information take time. Compromises between Paradoxically, in house-hunting ants, faster
speed and accuracy have been demonstrated in decisions could lead to slower emigrations, and nothing
decision-making by humans (e.g. Edwards 1965; would be gained.
Vitevitch 2002), monkeys (e.g. Roitman & Shadlen The reason for this potential paradox is that house-
2002) and bees (Chittka et al. 2003). They are also a hunting ants use quorum sensing in such a way that two
characteristic of ‘anytime algorithms’ in computing. key processes are inseparable. Essentially, quorum
These provide an early answer when interrupted but at sensing collates individual decisions into collective
the cost of accuracy (Dean & Boddy 1988). ones (i.e. corporate policy is determined politically);
Franks et al. (2002, 2003a), respectively, first but quorum sensing may also limit the number of ants
predicted and then experimentally demonstrated that know how to get into the new nest and can take an
speed–accuracy trade-offs in collective decision-making active role in expediting the emigration. Thus if the
ants use a low quorum threshold (QT) in an
emergency, the decision may be faster, and less well
* Author for correspondence (nigel.franks@bristol.ac.uk). informed, and the emigration may be slower. There-
†
Present address: Evolutionary Ecology, UMR CNRS 5561
BioGéoSciences, Université de Bourgogne, 6 Boulevard Gabriel, fore, by making a quicker decision, the ants might get
21000 Dijon, France. the worst of all possible worlds—a poor choice and a
One contribution of 11 to a Theme Issue ‘Group decision making in slow emigration, in which vulnerable members of the
humans and animals’. colony remain exposed for longer. Thus, because the

846 N. R. Franks et al. Politics and policy in house-hunting ants
QT sets a limit on both information gathering and is three times quicker than leading a tandem run.
training, the use of a low threshold in an emergency However, transported ants do not learn the route, as
could turn a crisis into a chronic refugee problem. To they are being carried along and, thus, have not been
understand this more fully, we will need to explain the trained to take an active role in the emigration.
natural history of this decision-making system. Moreover, because the QT sets a limit to tandem
When a colony needs to find a new home, certain running and may limit the ants that know the route
active scouts leave their vulnerable nest-mates and start between the old and new nests, it sets a limit on the
to search for something suitable. The majority of a speed of the emigration. Low QTs mean faster and
colony’s workers, however, remain at the old nest potentially less well-informed decisions and potentially
guarding the brood. Such workers take a passive role in slower emigrations, and, thus, greater exposure and
the emigration but have a crucial role in colony risk to colony members. Hence, lowering the QT in an
maintenance. When active scouts strike lucky, they emergency would seem to make things worse. The
assess many nest variables (Mallon & Franks 2000; notion of ‘any port in a storm’ implies a quick decision
Mugford et al. 2001; Franks et al. 2003b, 2005, and rapid implementation to minimize risk. However,
2006a,b, 2007a), and, all else being equal, they recruit unless the ants have an additional mechanism to restore
nest-mates more quickly to a high-quality nest than to a emigration speed, their use of low QTs in harsh
low-quality one (Mallon et al. 2001). Hesitation over conditions would appear to be maladaptive.
poor nests will favour selection of better ones. Initial Reverse tandem runs (RTRs) might resolve this
recruitment is typically by forward tandem running paradox. RTRs are the third form of recruitment
(Pratt et al. 2002). In this process, an ant that has found exhibited during emigrations and occur from the new
a suitable new nest, and knows an efficient route to it, nest back to the old one. First, as described earlier, the
leads a single nest-mate worker, i.e. an active scout that ants begin with FTRs; second, they switch to carrying
is still at the old nest, from the old nest forwards to the nest-mates; and then finally, as carrying continues, they
new one. Such tandem running is slow—but the ant typically begin RTRs. Earlier work by Pratt et al. (2002)
following the tandem leader can learn the route to the has shown that RTRs occur sporadically during the
new nest and later can either lead tandem runs itself or transport phase; 99 per cent of reverse tandem leaders
carry nest-mates to the new nest. Tandem running also recruited from the old nest; 84 per cent of RTRs
trains naive ants, who train others, thereby creating a ended with the leader picking up a nest-mate (adult or
positive-feedback information cascade. Moreover, such brood) at the old nest; 73 per cent of RTR followers
tandem running was the first case in animal behaviour were also, at some stage in the emigration, active
to be shown to meet all the criteria of a strict definition recruiters from the old nest; and 53 per cent of RTR
of teaching ( Franks & Richardson 2006). More followers had previously engaged in such recruitment
recently, Richardson et al. (2007) have even shown before following an RTR. Therefore, RTRs may, in
that the teaching ant, i.e. the tandem leader, engages in part, serve to reactivate recruiters and it seems likely
three forms of evaluation, namely (i) the amount she that all participants in FTRs and RTRs, both leaders
has already invested, (ii) the quality of the goal and (iii) and followers, are from the active group of workers.
the rate of progress of the tandem. Forward tandem Here, we report, for the first time, the results of
runs (FTRs), i.e. those from the old nest to the new new experiments to examine the role of reverse
one, continue until tandem leaders encounter a certain tandem runs and to determine whether they can
number of nest-mates in the new nest—the QT. The resolve the paradox of faster decisions potentially
size of the QT varies with circumstances. In benign causing slower emigrations.
conditions in the laboratory, it is typically between There has been extensive mathematical modelling of
approximately 10 and 20 workers (Pratt et al. 2002), the house-hunting algorithms of Temnothorax albipennis
whereas, the average colony has approximately 110 (Pratt et al. 2002, 2005; Marshall et al. 2006; Planqué
workers in total ( Franks et al. 2006a). In harsh et al. 2006, 2007; Pratt & Sumpter 2006). The
conditions, QTs are lower ( Franks et al. 2003a). most recent of these models by Planqué et al. (2007)
Moreover, if the ants have no need to move at all focused on the possible role of RTRs and predicted that
because their existing nest is fully functional but they they might serve to restore the speed of emigrations
have encountered something much better, they use a when the ants have used a low QT and have made
very high QT (Dornhaus et al. 2004). Moderate to rather few FTRs.
large QTs are sufficiently high that it is unlikely that one Accordingly, here we test two alternative hypotheses
ant alone would have created that quorum by leading for the function of RTRs. Hypothesis 1 is the simplest
very large numbers of nest-mates to the new nest in explanation for RTRs, namely that RTRs compensate
successive tandem runs (Pratt et al. 2002). Therefore, for a disruption of FTRs. Thus, if FTRs have been
high QT’s imply that many ants consider the new nest unsuccessful for whatever reason, the ants may
suitable ( Franks et al. 2002). In this way, quorum compensate by leading more RTRs.
sensing serves as a method of collating the separate Hypothesis 2 postulates that RTRs are related to
evaluations of many individual ants (Pratt 2005). scout dispersal (e.g. in emergencies) and the associated
Quorum sensing is a wonderful device in collective rarity of candidates to follow FTRs in the vicinity of the
decision-making. old nest. In an extreme emergency, scouts may disperse
Once a QT has been satisfied, the active ants switch far and wide from the old nest in search of any port in a
from tandem running to carrying ‘passive’ nest-mates: storm. Hence, when one does find a suitable new nest,
either other adults (brood-care workers and the queen) it may later find few, if any, active ants at the old nest to
or brood (eggs, larvae or pupae). Carrying a nest-mate recruit with an FTR. Therefore, such a pioneering

Politics and policy in house-hunting ants N. R. Franks et al. 847
old nest acetate bridge new nest 48 hours to settle into this environment before the experiment
was performed so that they could learn landmarks (Pratt et al.
2001; McLeman et al. 2002) that remained throughout the
experimental period.
Emigrations were induced by removing the upper slide
from the old nest to make it uninhabitable and an acetate
bridge was introduced to link the two Petri dishes (figure 1).
The bridge was made out of a rectangle of acetate 19!6 cm,
which was transparent to deter colonies from nesting beneath
its arch. The bridge was fixed to each dish by a strip of
old arena new arena double-sided tape, such that it was flush to the dishes with no
sticky surfaces exposed. To prevent the ants from escaping,
Figure 1. Experimental arenas used to control dispersion of the bridge did not touch the sides of the dishes. This design
the ants. When the original nest has been destroyed, the provided a clear division between the two Petri dish arenas.
second arena with the new nest site can be connected with an The new Petri dish was terra incognita for the ants. This design
acetate bridge. also created a substantial separation of the two nests and the
relatively narrow bridge also served further to challenge the
scout might do the next best thing and begin recruiting ants’ navigational abilities (Pratt et al. 2001; McLeman et al.
by transporting a passive nest-mate to the new nest. 2002). We favoured this design because it reduced the
Meanwhile, other scouts may have finally stumbled possibility of excited ants quickly finding the new nest by
upon the same new nest by circuitous routes and may chance and should thereby have reduced the number of ants
be available to be led in RTRs. These slow scouts may that had initially discovered the new nest. This should have
be partly disoriented by the time-consuming and increased the potential benefit of recruitment behaviours.
The colonies were allocated to give a similar distribution of
haphazard path they have taken. It might then benefit
colony sizes in the control (10 colonies) and treatment (10
the colony and their own inclusive fitness if they are
colonies) groups. Colonies from the control group were
shown, through following an RTR, a quicker path
allowed to emigrate undisturbed. By contrast, in the
between the old and new nests so that they can help
treatment group, we disrupted every FTR (that did not
expedite the emigration by carrying nest-mates to the
escape our attention). We wanted to do this without removing
new nest. any ants or causing them to panic. Our procedure consisted of
artificially stimulating the hind legs and abdomen of each
2. MATERIAL AND METHODS tandem run leader with an eyelash mounted on a cocktail
Twenty queen-right colonies of T. albipennis were collected in stick. We did this when a suitably large gap had naturally
October 2005 from south Dorset in the United Kingdom. occurred between the leader and the follower. Such large gaps
Colonies were housed in artificial nests consisting of a are common (Möglich 1978; Franks et al. 2002; Franks &
cardboard perimeter sandwiched between two microscope Richardson 2006). Suitable stimulation with an eyelash is
slides 75 mm!50 mm, forming nesting cavities with internal sufficient to encourage the leader to continue in the absence
dimensions 49 mm!30 mm!1.5 mm with an entrance of its true follower (Möglich et al. 1974). We continued to
tunnel 2 mm wide and 5 mm long. The artificial nests were tickle the leader with the eyelash, a few times per second, until
located in large, square Petri dishes 22 cm!22 cm!2.2 cm, the leader and the follower were approximately 2 cm apart so
which acted as foraging arenas. Fluon-coated walls prevented that they lost one another and the tandem run did not reform.
the ants from escaping and a closed lid preserved humidity in FTRs were disrupted as soon as they were observed, i.e.
the dish. Except during experiments, colonies had access to almost always within a few centimetres of the old nest.
Drosophila, honey solution and water ad libitum. Throughout The numbers of successful forward and reverse tandem
the study, colonies were kept on a low-vibration bench. runs were recorded by two investigators until 20 min after the
Worker populations ranged from 49 to 311 and brood last brood item was moved into the new nest. We considered a
populations from 78 to 287 (for information on colony successful tandem run to be the one that progressed to within
sizes, see Franks et al. 2006a). There were no significant a few millimetres of its target (i.e. the new or old nest). Some
differences in the median sizes of colonies used in the two tandem runs aborted naturally during travel from one nest to
treatments and two controls described below (Kruskal–Wallis the other either because the tandem was broken by a collision
test HZ0.35, d.f.Z3, pZ0.951). with nest-mates or because the follower became lost. Thus,
we also recorded the number of partial tandem runs. Other
(a) Experiment 1: do reverse tandem runs components of the emigration were also recorded such as the
compensate for a disruption of FTRs? time that elapsed between the opening of the old nest and the
A new Petri dish was placed abutting the one containing the queen and last brood item being moved into the new nest. To
colony (figure 1). The new Petri dish contained a new nest record the numbers of ants in the old and new nests and
site that was identical to the original one except that it had a hence the dynamics of the emigration, pictures of the old and
removable cardboard cover to make it dark and hence more new nests were taken immediately before and after the
attractive than the original one ( Franks et al. 2003b) to removal of the upper microscope slide of the original nest and
encourage an emigration ( Franks et al. 2003b). The two every 10 min thereafter.
nests, the old and the new one, were thus placed 33.5 cm During each emigration, we observed a striking initial
apart, at opposite ends of the old and new dishes (figure 1). drop in worker numbers during the first 20 min of the
Two digital cameras (Nikon, Coolpix 889) were mounted on emigration followed by a plateau phase (figure 2a, see also
stands 18 cm above each nest. Each camera’s field of view was figure 3a for comparable data from experiment 2). This
adjusted to encompass the entire nest interior and the typical pattern was associated with the initial exodus of scouts
cardboard perimeters. Colonies were allowed at least from the old nest during the emigration. Thus, the number of

(a) 160 (a) 160
140 140
120 120
no. of adults in the old nest
no. of adults in the old nest

100 100
80 80
60 60
40 40
20 20
0 0
(b) 160 (b) 160
140 140
no. of adults in the new nest
no. of adults in the new nest

120 120
100 100
80 80
60 60
40 40
20 20
0 25 50 75 100 125 150 0 25 50 75 100 125 150

time (min) time (min)
Figure 2. Experiment 1. FTRs disrupted in the treatment but Figure 3. Experiment 2. In the treatment, scouts were
not in the control. Mean numbers of adults in the (a) old and released in the arena with the new nest. In the control, scouts
(b) new nests for colonies from the control (grey line) or were released in the arena with the old nest. Mean numbers of
treatment (black line) group as a function of time since the individuals in the (a) old and (b) new nests for colonies from
old nest was destroyed. The error bars represent the standard the control (grey line) or treatment (black line) group as a
error to the mean. FTRs expedite emigrations. function of time since the old nest was destroyed. The error
bars represent standard errors. In the treatment group, the
scouts was estimated simply as the total number of ants in
ants used many reverse tandem runs. In the control group,
the colony minus the number of ants still in the old nest the ants used many FTRs. RTRs can expedite emigrations.
after 20 min.
Colonies were induced to emigrate as before, however, the

(b) Experiment 2: is the abundance of reverse tandem
runs influenced by the location of scouts and a bridge was not introduced immediately. Colonies were
reduction in the availability of potential tandem allowed 5 min to relocate brood and settle after the nest
recruits at the old nest? had been opened. After this time and for a further 15 min
Several weeks after experiment 1 (so that the ants were not period, ants scouting at least 2 cm from the nest perimeter
influenced by their prior experience—see Langridge et al. were assumed to be active scouts and were aspirated from the
2004), the same colonies were emigrated for a second time. old dish and isolated in the separate small Petri dish corral.
Half of the treatment colonies in the first experiment became Twenty minutes after starting the experiment, scouts
controls in the second, while half of the control colonies in belonging to control colonies were then gently transferred
the first experiment became treatments in the second. The from the temporary arena into the centre of the old Petri dish,
experimental set-up was similar to experiment 1, with i.e. the one that housed the old nest, while those belonging to
the addition of temporary arenas that served as corrals into treatment colonies were gently deposited into the centre of
which active scouts (as defined below) were imprisoned at the new Petri dish, i.e. the one housing the potential new nest
the beginning of each emigration. Each temporary arena site. Scouts in the treatments should have been much more
consisted of small (10 cm!10 cm!1.9 cm) Petri dishes, disorientated than those in the controls. Colonies were
situated to one side of the large Petri dish that housed the subsequently left for 20 min, so transferred scouts could
original nest. recover from the disturbance and investigate the particular

large Petri dish to which they had been transferred. The 30 (a) 30 (b)
bridge (figure 1) was then introduced and all subsequent *** n.s.
no. of successful forward tandem runs
no. of successful reverse tandem runs

emigration behaviour was recorded as in experiment 1.
25 25
(c) Statistical analysis
Effects of the treatment, colony size or number of scouts on 20 20
the numbers of tandem runs were analysed using general
linear model analyses of variance. Data on the dynamics of 15 15
the emigration were analysed by means of Cox’s (1972)
proportional hazards models. Cox analysis is a standard
10 10
technique for the analysis of censored survival data (Collett
1994; Dechaume-Moncharmont et al. 2003) and ecological
and behavioural data (for example Dechaume-Moncharmont 5 5
et al. 2005).
0 0
control treatment control treatment
3. RESULTS
Figure 4. Experiment 1. Box-plots of the number of
(a) Experiment 1: do reverse tandem runs
successful (a) forward and (b) reverse tandem runs in control
compensate for a disruption of FTRs? and treatment groups (10 colonies in each group). Horizontal
(i) Number of tandem runs lines within boxes are medians, boxes show inter-quartile
Disruption of FTRs during experiment 1 prevented ranges and whiskers show entire range (excluding the outliers
almost every FTR from being successful. Thus, represented by circles). The squares indicate means.

significantly more successful FTRs were recorded for p!0.001.
the control colonies than for the treatment colonies
(ANOVA: F1,18Z19.47, p!0.001; figure 4a). However, 30 (a) 30 (b)
** *
no. of successful forward tandem runs
there was no significant difference between the total
no. of successful reverse tandem runs

numbers of FTRs (including partial and successful 25 25
tandem runs) initiated by ants from control colonies or
from treatment colonies before disruption by the 20 20
experimenters (F1,18Z3.02, pZ0.100). This suggests
that the ants in the treatment group neither tried harder
15 15
to recruit by FTRs in response to low levels of success,
nor did they give up on their attempts to lead FTRs.
Furthermore, no significant difference was found 10 10
between either the number of successful RTRs
(F1,18Z2.29, pZ0.149; figure 4b) or the total number 5 5
of RTRs initiated (F1,18Z2.39, pZ0.140) by control and
treatment colonies. Colony size had no significant effect 0 0
on the number of successful FTRs (F1,18Z1.97, control treatment control treatment
pZ0.178); however, it did have a significant positive Figure 5. Experiment 2. Box-plots of the number of
effect on the number of successful RTRs (F1,18Z22.23, successful (a) forward and (b) reverse tandem runs in control
p!0.001). The number of scouts also had no significant and treatment groups (10 colonies in each group). Interpre-
effect on the number of successful FTRs (F1,16Z2.05, tation of box-plots and symbols as in the legend of figure 4.

pZ0.173) or the number of successful RTRs p!0.01, p!0.05.
(F1,16Z0.44, pZ0.517).
colonies in experiment 2 (in which the scouts were
placed in the arena containing the new nest), than by
(ii) Dynamics of colony emigration
control colonies (in which the scouts were returned
We recorded the number of ants in both nests (figure 2).
to the arena containing the old nest; F1,17Z5.85,
The treatment significantly decreased the probability of
pZ0.028; figure 5b).
ants both leaving the old nest (c21Z35.43, p!0.0001)
There was no significant effect of the number of
and entering the new nest (c21Z101.05, p!10K5).
scouts corralled on the number of successful FTRs
There was a significant positive effect of colony size
(F1,15Z0.91, pZ0.357) or successful RTRs (F1,15Z2.73,
on the probability of ants leaving the old nest
pZ0.121). Colony size had no significant effect on the
(c21Z33.05, p!0.0001) and entering the new one
number of successful FTRs (F1,17Z1.69, pZ0.212) and
(c21Z13.08, p!0.001).
had a significant positive effect on the number of
successful RTRs (F1,17Z4.79, pZ0.044).
(b) Experiment 2: is the abundance of reverse
tandem runs influenced by the location of scouts
and a reduction in the availability of potential (ii) Dynamics of colony emigration
tandem recruits at the old nest? There was a significant positive relationship between
(i) Number of tandem runs colony size and the number of active ants that left the
Successful FTRs were performed more often by old nest in the period between 5 and 20 min after the
control colonies compared with treatment colonies old nest had been opened. The relationship is best
(F1,17Z9.41, pZ0.007; figure 5a). Successful RTRs described by active scoutsZ5.95C0.3795 (colony size)
were performed significantly more often by treatment (r 2Z0.77, p!0.001).

160 4. DISCUSSION
The results reported here are the first systematic
140 manipulation of the relative abundance of FTRs and
reverse tandem runs (RTRs). Moreover, they not only
120 explain why these house-hunting ants recruit in both
directions but they also solve the prima facie paradox
100 of quick decisions being of no value if they lead to
active scouts
slow implementation of choices.

80 The results of experiment 1 show that even a complete
absence of successful FTRs did not lead to more RTRs.
60 This refutes the hypothesis that RTRs directly compen-
sate for too few successful FTRs. However, RTRs were
40
significantly more abundant when the scouts were
experimentally transferred (in experiment 2) to the
vicinity of the new nest but did not ‘know’ how they got
20
there (figure 5b). Accordingly, there were many more
RTRs when the most active ants were more abundant
0
50 100 150 200 250 300 near the new nest than at the old one (figure 5b). All of
colony size (total workers) these results taken together strongly support the
hypothesis that FTRs may occur when suitable recruits
Figure 6. Experiment 2. The relationship between colony size are available near the old nest and that RTRs may occur
and the number of active ants that left the old nest in the
when suitable recruits are more available in the vicinity
period between 5 and 20 min after the old nest had been
opened. The relationship is best described by active scoutsZ
of the new nest. In both experiments, larger colonies
5.95C0.3795 (colony size) (r 2Z0.77, p!0.001). The produced significantly more RTRs. This is consistent
central line is the fitted regression line; also shown are the with earlier results (Franks et al. 2006a). Larger colonies
95% confidence limits for this line and for the dataset. One have more active workers (figure 6) so they may have a
outlier has been removed. larger pool of potential RTR followers and leaders. In
general, our observations suggest that tandem runs are
initiated by tandem leaders rather than by lost scouts
These active ants were the ones collected and ‘asking for directions’. However, it obviously takes two to
initially imprisoned. This procedure was identical in tandem and an FTR or RTR will only occur if both
the treatment and control so the data were pooled. parties are willing and able actively to participate.
Active scouts, in colonies of all sizes, are approximately Overall, our results also strongly suggest that RTRs
40 per cent of a colony’s total workforce (figure 6). can rescue the fast dynamics of emigrations (figure 3),
The treatment significantly increased the probability just as predicted by Planqué et al. (2007).
of ants both leaving the old nest (c21Z15.65, p!0.001, Our focus has been on how colonies maximize their
figure 3a) and entering the new nest (c21Z4.96, emigration rates, as indicated by the gradients of the
pZ0.026, figure 3b). lines in figures 2 and 3, and hence how they minimize
There was a significant positive effect of colony size the average exposure of nest-mates to extranodal
on the probability of ants leaving the old nest hazards. Note that total emigration times can be both
(c21Z21.08, p!0.001) but not on the probability highly variable and arguably not very meaningful if a
of entering the new one (c21Z1.65, pZ0.19). The colony takes several hours finally to retrieve the last
number of scouts had a significant effect on the brood item from the old nest.
In the light of these new results, we will now
probability of both leaving the old nest (c21Z28.76,
summarize our current understanding of this decision-
p!0.001) and entering the new one (c21Z5.88,
making/choice implementation system.
pZ0.015). The number of ants in the new nest in
In benign conditions, scouts slowly disperse from
experiment 2 (figure 3) shows different trends to those
the old nest in search of the best available one. When
found in experiment 1 (figure 2). There was no
they have found something suitable, they return to the
increase at all in the number of ants in the new nest at old nest and can find many candidates among the active
the beginning of the emigration because the scouts scout population that can be led in FTRs. Such
were kept in the temporary corral. Moreover, as the recruitment builds a large quorum in a suitable nest.
ants from the treatment group were introduced into Hence, by the time the ants switch to recruitment by
the new large Petri dish arena, they found the new nest carrying, there is an abundance of active ants that have
more rapidly than the ants from the control group, been taught an efficient emigration route between the
which had been reintroduced to the original ‘old’ old and new nests and can take an active role in
arena. Thus, the population in the new nest from the carrying their more passive nest-mates quickly to the
treatment group increased 20 min earlier than the colony’s new home. In such circumstances, we expect
corresponding populations in the control colonies. and find rather few reverse tandem runs. For example,
When this lag of 20 min was removed by transposing the most benign conditions occur when the old nest
the data from the control emigrations forwards remains intact and colonies are moving to improve.
20 min, the treatment was found to have no significant Under such circumstances, the ants use very high QTs,
effect on the dynamics of new nest colonization large numbers of FTRs and very few reverse tandem
(c21Z1.64, pZ0.20). runs (Dornhaus et al. 2004).

By contrast, in harsh conditions many more scouts assume to be benign conditions, we sometimes
quickly leave the old nest (N. R. Franks 2003, personal observed more RTRs than FTRs and both may be
observations; see also Pratt & Sumpter (2006) for a very variable (Dornhaus et al. 2004; Franks et al.
similar finding for a closely related species) dispersing 2006a; E. A. Langridge 2000, personal communication;
rapidly far and wide looking, as it were, for any ‘port in A. B. Sendova-Franks 1998, unpublished data). We now
a storm’, i.e. any moderately suitable new nest. When a suspect that this is most often the case when, for
scout finds something acceptable, and she returns to technical reasons such as filming emigrations, the old
the old nest, very few scouts are available to be led in nest and new one were very close together and in a
FTRs. Therefore, in an emergency, returning scouts do small arena. This would have made the new nest quick
the next most useful task; they carry passive nest-mates and easy to discover by independent scouts. Hence a
(nest workers, the brood and even eventually the high quorum could be quickly met with very few FTRs.
queen) to the new nest. (We define quorum sensing This interpretation is consistent with the data from
as complete when carrying begins. Therefore, by experiment 2. In the treatment, the large numbers of
default, in emergency emigrations when the ants start scouts displaced by hand to the new arena found the
carrying very soon, they have de facto typically used new nest quickly (figure 3b) and this may have created
very small or non-existent QTs.) Carrying passive nest- an artificial quorum that helped to suppress FTRs
mates continues but when a scout has deposited its load (figure 5a; see also Pratt et al. 2002). In addition, if the
at the new nest, it attempts to lead RTRs from there new nest is extremely easy to find, many ants may find it
(Pratt et al. 2002). This becomes increasingly easy as so easily, and directly, that both FTRs and RTRs may
other scouts independently discover the new nest and be scarce because no ant needs to be taught the route
become available to be led in an RTR. Why should (Franks & Richardson 2006; Richardson et al. 2007).
these initially independent scouts be prepared to follow Therefore, in sum, we suggest that reverse tandem
RTRs? One obvious explanation is that the longer they runs enable colonies to implement quick decisions
have taken to find the new nest the less direct will be the rapidly. This interpretation is also compatible with
route they have taken to it from the old nest. Tandem certain RTRs also serving to reactivate scouts who had
runs teach effective and direct routes—but even these
played an active recruitment role earlier in the
seem to get better through time (Pratt et al. 2005;
emigration (Pratt et al. 2002).
Franks & Richardson 2006).
These ants recruit in both directions when they
In an emergency emigration (i) FTRs can be rare,
emigrate, so that they can both decide quickly and
(ii) QTs appear to be low or non-existent, (iii) decision-
emigrate quickly. They are thus very likely to have
making is quicker and more error-prone simply
speed–accuracy trade-offs not just in the initial decision-
because it is less consensual and less well informed,
making stages of their house hunting—but these should
(iv) carrying occurs sooner, and then (v) the ants may
lead large numbers of reverse tandem runs to bring also feed forward into global compromises between
lost or disorientated scouts into an active role in the accuracy of choice and speed of implementation.
emigration. (For results supporting (i–iv), see Franks Temnothorax albipennis also have an additional
et al. 2003b, (v) applies to data reported in this paper.) mechanism for expediting nest choice and emigrations
In the best of all possible (benign) worlds, FTRs in an emergency. They reconnoitre for potential nest
might yield both more discerning collective decisions sites before they have any need to emigrate and they
(because more individuals will contribute to nest remember poor ones with landmarks and mark them
evaluation) and faster emigrations if the positive feed- with pheromones. This presumably enables them to
back, associated with FTRs (which recruit further focus their search elsewhere for better nests if and when
recruiters and so on), ‘kicks-in’ quickly. However, in they become beset by the emergency of homelessness
harsh conditions, the ants use quick individual decisions (Franks et al. 2007b).
and start emigrations directly by carrying because no Our results therefore support the view that colony-
active ant is available to be led in a FTR. Carrying is level emigration behaviour emerges from interactions
interspersed by reverse tandem runs (Planqué et al. among individuals following local, but sophisticated,
2006) that teach lost scouts to take an active role in the behavioural rules (Camazine et al. 2001; Pratt et al.
emigration (Richardson et al. 2007). Furthermore, any 2005). This indicates yet another example of a
positive feedback associated with RTRs may also begin sophisticated strategy that T. albipennis has evolved in
without delay (in the overall scheme of things, compared the process of efficiently finding, choosing and moving
with FTR in benign conditions) because in harsh in to a new home.
conditions latency periods associated with lengthy Our study shows that decision-making may mean
collective assessments will be minimized. very little in isolation. It is not enough just to make a
The emigration dynamics shown in figure 3a,b seem decision, either quickly or accurately. What is the value
to imply that RTR’s can lead to faster emigrations than of a decision if it is not implemented well? Recall the
FTR’s. This, however, is almost certainly an artefact scathing epitaph ‘He never said a foolish thing, nor ever
of scouts being transferred by hand to the new nest did a wise one’.
site. Nevertheless, these results do strongly suggest
N.R.F. wishes to thank the Biotechnology and Biological
that reverse tandem runs can restore much of the speed Sciences Research Council (BBSRC) for research grant
of emigrations. (E19832), which supported F.-X.D.-M. We also gratefully
These interpretations also explain why some of our acknowledge Elizabeth Langridge, Ana Sendova-Franks, Bob
earlier results might appear anomalous. In certain of Planqué, Nathalie Stroeymeyt, Elva Robinson and other
our experiments, in the laboratory under what we members of the Bristol ant lab for their helpful discussions.

REFERENCES Langridge, E. A., Franks, N. R. & Sendova-Franks, A. B.

Camazine, S., Deneubourg, J.-L., Franks, N. R., Sneyd, J., 2004 Improvement in collective performance with experi-
Theraulaz, G. & Bonabeau, E. 2001 Self-organization in ence in ants. Behav. Ecol. Sociobiol. 56, 523 –529. (doi:10.
biological systems. Princeton, NJ: Princeton University Press. 1007/s00265-004-0824-3)
Chittka, L., Dyer, A. G., Bock, F. & Dornhaus, A. 2003 Bees Mallon, E. & Franks, N. R. 2000 Ants estimate area using
trade off foraging speed for accuracy. Nature 424, 388 . Buffon’s needle. Proc. R. Soc. B 267, 765–770. (doi:10.
(doi:10.1038/424388a) 1098/rspb.2000.1069)
Collett, D. 1994 Modelling survival data in medical research. Mallon, E. B., Pratt, S. C. & Franks, N. R. 2001 Individual
London, UK: Chapman & Hall. and collective decision-making during nest site selection
Cox, D. R. 1972 Regression models and life tables (with by the ant Leptothorax albipennis. Behav. Ecol. Sociobiol. 50,
discussion). J. R. Stat. Soc. B 34, 187–220. 352– 359. (doi:10.1007/s002650100377)
Dean, T. L. & Boddy, M. 1988 An analysis of time-dependent Marshall, J. A. R., Dornhaus, A., Franks, N. R. & Kovacs, T.
planning. In Proc. 7th Nat. Conf. on Artificial Intelligence 2006 Noise, cost and speed-accuracy trade-offs: decision-
(eds T. M. Mitchell & R. G. Smith), pp. 49–54. Menlo Park, making in a decentralized system. J. R. Soc. Interface 3,
CA: AAAI Press. 243 –254. (doi:10.1098/rsif.2005.0075)
Dechaume-Moncharmont, F. X., Decourtye, A., Hennequet, McLeman, M. A., Pratt, S. C. & Franks, N. R. 2002
C., Pons, O. & Pham-Delègue, M.-H. 2003 Statistical Navigation using visual landmarks by the ant Leptothorax
analysis of honeybee survival after chronic exposure to
albipennis. Insectes Soc. 49, 203 –208. (doi:10.1007/
insecticides. Environ. Toxicol. Chem. 22, 3088 – 3094.
s00040-002-8302-2)
(doi:10.1897/02-578)
Möglich, M. 1978 Social organization of nest emigration in
Dechaume-Moncharmont, F. X., Azzouz, H., Pons, O. &
Leptothorax (Hym. Form.). Insectes Soc. 25, 205 –225.
Pham-Delègue, M. H. 2005 Soybean proteinase inhibitor
(doi:10.1007/BF02224742)
does not affect the foraging strategy of free flying honeybees.
Apidologie 36, 421– 430. (doi:10.1051/apido:2005031) Möglich, M., Maschwitz, U. & Hölldobler, B. 1974 Tandem
Dornhaus, A., Franks, N. R., Hawkins, R. M. & Shere, calling: a new kind of signal in ant communication. Science
H. N. S. 2004 Ants move to improve—colonies of 186, 1046 –1047. (doi:10.1126/science.186.4168.1046)
Leptothorax albipennis emigrate whenever they find a Mugford, S. T., Mallon, E. B. & Franks, N. R. 2001 The
superior nest site. Anim. Behav. 67, 959 – 963. (doi:10. accuracy of Buffon’s needle: a rule of thumb used by ants
1016/j.anbehav.2003.09.004) to estimate area. Behav. Ecol. 12, 655 –658. (doi:10.1093/
Edwards, W. 1965 Optimal strategies for seeking infor- beheco/12.6.655)
mation: models for statistics, choice reaction times, and Planqué, R., Dornhaus, A., Franks, N. R., Kovacs, T. &
human information processing. J. Math. Psychol. 2, Marshall, J. A. R. 2006 Weighting waiting in collective
312– 329. (doi:10.1016/0022-2496(65)90007-6) decision making. Behav. Ecol. Sociobiol. 61, 347– 356.
Franks, N. R. & Richardson, T. 2006 Teaching in tandem- (doi:10.1007/s0265-006-0263-4)
running ants. Nature 439, 153 . (doi:10.1038/439153a) Planqué, R., Dechaume-Moncharmont, F.-X., Franks, N. R.,
Franks, N. R., Pratt, S. C., Mallon, E. B., Britton, N. F. & Kovacs, T. & Marshall, J. A. R. 2007 Why do house-hunting
Sumpter, D. J. T. 2002 Information flow, opinion polling ants recruit in both directions? Naturwissenschaften 94,
and collective intelligence in house-hunting social insects. 911– 918. (doi:10.1007/s00114-007-0273-8)
Phil. Trans. R. Soc. B 357, 1567–1583. (doi:10.1098/rstb. Pratt, S. 2005 Quorum sensing by encounter rate in the ant
2002.1066) Temnothorax albipennis. Behav. Ecol. 16, 488 – 496. (doi:10.
Franks, N. R., Dornhaus, A., Fitzsimmons, J. P. & Stevens, 1093/beheco/ari020)
M. 2003a Speed versus accuracy in collective decision Pratt, S. C. & Sumpter, D. J. T. 2006 A tunable algorithm for
making. Proc. R. Soc. B 270, 2457–2463. (doi:10.1098/ collective decision-making. Proc. Natl Acad. Sci. USA 103,
rspb.2003.2527) 15 906 –15 910. (doi:10.1073/pnas.0604801103)
Franks, N. R., Mallon, E. B., Bray, H. E., Hamilton, M. J. & Pratt, S. C., Brooks, S. E. & Franks, N. R. 2001 The use of
Mischler, T. C. 2003b Strategies for choosing among edges in visual navigation by the ant Leptothorax albipennis.
alternatives with different attributes: exemplified by Ethology 107, 1125–1136. (doi:10.1046/j.1439-0310.2001.
house-hunting ants. Anim. Behav. 65, 215 –223. (doi:10. 00749.x)
1006/anbe.2002.2032) Pratt, S. C., Mallon, E. B., Sumpter, D. J. T. & Franks, N. R.
Franks, N. R., Hooper, J. W., Webb, C. & Dornhaus, A. 2005 2002 Quorum sensing, recruitment, and collective
Tomb evaders: house-hunting hygiene in ants. Biol. Lett. 1,
decision-making during colony emigration by the ant
190–192. (doi:10.1098/rspb.2005.0302)
Leptothorax albipennis. Behav. Ecol. Sociobiol. 52, 117–127.
Franks, N. R., Dornhaus, A., Best, C. S. & Jones, E. L. 2006a
(doi:10.1007/s00265-002-0487-x)
Decision-making by small & large house-hunting ant
Pratt, S. C., Sumpter, D. J. T., Mallon, E. B. & Franks, N. R.
colonies: one size fits all. Anim. Behav. 72, 611–616.
2005 An agent-based model of collective nest choice by
(doi:10.1016/j.anbehav.2005.11.019)
Franks, N. R., Dornhaus, A., Metherell, B. G., Nelson, T. R., the ant Temnothorax albipennis. Anim. Behav. 70,
Lanfear, S. A. J. & Symes, W. S. 2006b Not everything that 1023 –1036. (doi:10.1016/j.anbehav.2005.01.022)
counts can be counted: ants use multiple metrics for a Richardson, T. O., Sleeman, P. A., McNamara, J. M.,
single nest trait. Proc. R. Soc. B 273, 165 –169. (doi:10. Houston, A. I. & Franks, N. R. 2007 Teaching with
1098/rspb.2005.3312) evaluation in ants. Curr. Biol. 17, 1520–1526. (doi:10.
Franks, N. R., Dornhaus, A., Hitchcock, G., Guillem, R., 1016/j.cub.2007.08.032)
Hooper, J. & Webb, C. 2007a Avoidance of conspecific Roitman, J. D. & Shadlen, M. N. 2002 Response of neurons
colonies during nest choice by ants. Anim. Behav. 73, in the lateral intraparietal area during a combined visual
525 –534. (doi:10.1016/j.anbehav.2006.05.020) discrimination reaction time task. J. Neurosci. 22,
Franks, N. R., Hooper, J. W., Dornhaus, A., Aukett, P. J., 9475 – 9489.
Hayward, A. L. & Berghoff, S. M. 2007b Reconnaissance Vitevitch, M. S. 2002 Influence of onset density on spoken
and latent learning in ants. Proc. R. Soc. B. 274, word recognition. J. Exp. Psychol. Hum. Percep. Perform.
1505 –1509. (doi:10.1098/rspb.2007.0138) 28, 270–278. (doi:10.1037/0096-1523.28.2.270)

RSTB_364_1518.qxp 1/27/09 8:21 PM Page 1
Phil. Trans. R. Soc. B | vol. 364 no. 1518 pp. 717–852 | 27 Mar 2009
ISSN 0962-8436
volume 364
27 March 2009 number 1518

volume 364 . number 1518 . pages 717–852
pages 717–852
Group decision making in humans and
animals In this issue
Papers of a Theme Issue compiled and edited by Larissa Conradt and Christian List Group decision making
Introduction
in humans and animals
L. Conradt & C. List Papers of a Theme Issue compiled and edited by
Articles Larissa Conradt and Christian List
D. J. T. Sumpter & S. C. Pratt
Independence and interdependence in collective decision making: an agent-based model of
nest-site choice by honeybee swarms 755
C. List, C. Elsholtz & T. D. Seeley
D. Austen-Smith & T. J. Feddersen
B. Skyrms

J. R. G. Dyer, A. Johansson, D. Helbing, I. D. Couzin & J. Krause
Reciprocity, culture and human cooperation: previous insights and a new cross-cultural experiment 791
S. Gächter & B. Herrmann
L. Conradt & T. J. Roper
S. Hix, A. Noury & G. Roland
M. Regenwetter, B. Grofman, A. Popova, W. Messner, C. P. Davis-Stober & D. R. Cavagnaro
Speed versus accuracy in decision-making ants: expediting politics and policy implementation 845
N. R. Franks, F.-X. Dechaume-Moncharmont, E. Hanmore & J. K. Reynolds
The world’s longest running science journal

Founded in 1660, the Royal Society
is the independent scientific academy
of the UK, dedicated to promoting
excellence in science
rstb.royalsocietypublishing.org
Registered Charity No 207043 Published in Great Britain by the Royal Society, 27 March 2009
6–9 Carlton House Terrace, London SW1Y 5AG

Larissa Conradt, Christian List - Group Decision Making in Humans and Animals (Philosophical Transactions of The Royal Society Series B)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Larissa Conradt, Christian List - Group Decision Making in Humans and Animals (Philosophical Transactions of The Royal Society Series B)

Uploaded by

Copyright:

Available Formats

RSTB_364_1518.

qxp 1/27/09 8:21 PM Page 1

27 March 2009 number 1518

Group decision making in humans and animals

The world’s longest running science journal

GUIDANCE FOR AUTHORS

SUBSCRIPTIONS and technology

Independence and interdependence in collective decision making: an agent-based 755

Information aggregation and communication in committees 763

Evolution of signalling systems with multiple senders and receivers 771

Leadership, consensus decision making and collective behaviour in humans 781

Conflicts of interest and the evolution of decision sharing 807

Voting patterns and alliance formation in the European Parliament 821

Behavioural social choice: a status report 833

Phil. Trans. R. Soc. B (2009) 364, 719–742

Group decisions in humans and animals: a survey

1. GENERAL BACKGROUND mole rats). Group decision making is just as important

719 This journal is q 2008 The Royal Society

720 L. Conradt & C. List Introduction

3. KEY CONCEPTS FOR THE ANALYSIS OF

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 721

Table 1. Examples of different categories of group decisions.

social sciences natural sciences

groups with global overview

Phil. Trans. R. Soc. B (2009)

722 L. Conradt & C. List Introduction

Box 1. Aggregation rules in choices between two options.

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 723

decision making object of group decision

unshared forage patch choice in primates (King et al. 2008)

Phil. Trans. R. Soc. B (2009)

724 L. Conradt & C. List Introduction

Box 2. Equilibria in interactive/combined decisions.

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 725

Phil. Trans. R. Soc. B (2009)

726 L. Conradt & C. List Introduction

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 727

Phil. Trans. R. Soc. B (2009)

728 L. Conradt & C. List Introduction

Box 3. Informational differences between group members.

1. Unshared decisions versus equally shared decisions

2. Difference in information between group members

3. Influence of the decision threshold on decision accuracy

(a) Skewed likelihood of a particular option to be the ‘best’

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 729

(b) Benefits are also skewed

The expected benefits with a majority threshold are

(c) The fully general result

Phil. Trans. R. Soc. B (2009)

730 L. Conradt & C. List Introduction

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 731

Box 4. Interest differences between group members.

Phil. Trans. R. Soc. B (2009)

732 L. Conradt & C. List Introduction

Phil. Trans. R. Soc. B (2009)

Introduction L. Conradt & C. List 733

Phil. Trans. R. Soc. B (2009)