The Econometrics of Networks 1st

Edition Áureo De Paula

Visit to download the full and correct content document:
https://ebookmeta.com/product/the-econometrics-of-networks-1st-edition-aureo-de-pa
ula/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Chess Explained The c3 Sicilian 1st Edition Sam Collins

https://ebookmeta.com/product/chess-explained-
the-c3-sicilian-1st-edition-sam-collins/

Starting Out The c3 Sicilian 1st Edition John Emms

https://ebookmeta.com/product/starting-out-the-c3-sicilian-1st-
edition-john-emms/

The Paper Issue 83 1st Edition Origamiusa

https://ebookmeta.com/product/the-paper-issue-83-1st-edition-
origamiusa/

The American Revolution 1774 83 2nd Edition Daniel

Marston

https://ebookmeta.com/product/the-american-
revolution-1774-83-2nd-edition-daniel-marston/
Essentials of Econometrics 5th Edition Damodar Gujarati

https://ebookmeta.com/product/essentials-of-econometrics-5th-
edition-damodar-gujarati/

The Econometrics of Multi-dimensional Panels 2nd

Edition Laszlo Matyas

https://ebookmeta.com/product/the-econometrics-of-multi-
dimensional-panels-2nd-edition-laszlo-matyas/

Cancer Metabolomics 2018 Paula Guedes De Pinho Editor

Márcia Carvalho Editor Joana Pinto Editor

https://ebookmeta.com/product/cancer-metabolomics-2018-paula-
guedes-de-pinho-editor-marcia-carvalho-editor-joana-pinto-editor/

The Passionate Heart 1st Edition Paula Quinn

https://ebookmeta.com/product/the-passionate-heart-1st-edition-
paula-quinn/

Physical Chemistry Thermodynamics Structure and Change

Peter Atkins Julio de Paula Tenth 2014 10th Edition
Peter Atkins

https://ebookmeta.com/product/physical-chemistry-thermodynamics-
structure-and-change-peter-atkins-julio-de-paula-tenth-2014-10th-
edition-peter-atkins/
THE ECONOMETRICS OF
NETWORKS
ADVANCES IN ECONOMETRICS
Series editors: Thomas B. Fomby, R. Carter Hill, Ivan
Jeliazkov, Juan Carlos Escanciano, Eric Hillebrand,
Daniel L. Millimet, Rodney Strachan
Previous Volumes
Volume 21 Modelling and Evaluating Treatment Effects in Econometrics – Edited by
Daniel L. Millimet, Jeffrey A.Smith and Edward Vytlacil
Volume 22 Econometrics and Risk Management – Edited by Jean-Pierre Fouque,
Thomas B. Fomby and Knut Solna
Volume 23 Bayesian Econometrics – Edited by Siddhartha Chib, Gary Koop, Bill
Griffiths and Dek Terrell
Volume 24 Measurement Error: Consequences, Applications and Solutions – Edited by
Jane Binner, David Edgerton and Thomas Elger
Volume 25 Nonparametric Econometric Methods – Edited by Qi Li and Jeffrey S.
Racine
Volume 26 Maximum Simulated Likelihood Methods and Applications – Edited by R.
Carter Hill and William Greene
Volume 27A Missing Data Methods: Cross-sectional Methods and Applications – Edited
by David M. Drukker
Volume 27B Missing Data Methods: Time-series Methods and Applications – Edited by
David M. Drukker
Volume 28 DSGE Models in Macroeconomics: Estimation, Evaluation and New
Developments – Edited by Nathan Balke, Fabio Canova, Fabio Milani and
Mark Wynne
Volume 29 Essays in Honor of Jerry Hausman – Edited by Badi H. Baltagi, Whitney
Newey, Hal White and R. Carter Hill
Volume 30 30th Anniversary Edition – Edited by Dek Terrell and Daniel Millmet
Volume 31 Structural Econometric Models – Edited by Eugene Choo and Matthew
Shum
Volume 32 VAR Models in Macroeconomics — New Developments and Applications:
Essays in Honor of Christopher A. Sims – Edited by Thomas B. Fomby,
Lutz Kilian and Anthony Murphy
Volume 33 Essays in Honor of Peter C. B. Phillips – Edited by Thomas B. Fomby,
Yoosoon Chang and Joon Y. Park
Volume 34 Bayesian Model Comparison – Edited by Ivan Jeliazkov and Dale J. Poirier
Volume 35 Dynamic Factor Models – Edited by Eric Hillebrand and Siem Jan Koopman
Volume 36 Essays in Honor of Aman Ullah – Edited by Gloria Gonzalez-Rivera, R.
Carter Hill and Tae-Hwy Lee
Volume 37 Spatial Econometrics – Edited by Badi H. Baltagi, James P. LeSage and R.
Kelley Pace
Volume 38 Regression Discontinuity Designs: Theory and Applications – Edited by
Matias D. Cattaneo and Juan Carlos Escanciano
Volume 39 The Econometrics of Complex Survey Data: Theory and Applications –
Edited by Kim P. Huynh, David T. Jacho-Chávez and Guatam Tripathi
Volume 40A Topics in Identification, Limited Dependent Variables, Partial Observability,
Experimentation, and Flexible Modelling Part A – Edited by Ivan Jeliazkov
and Justin L. Tobias
Volume 40B Topics in Identification, Limited Dependent Variables, Partial Observability,
Experimentation, and Flexible Modelling Part B – Edited by Ivan Jeliazkov
and Justin L. Tobias
Volume 41 Essays in Honor of Cheng Hsiao – Edited by Tong Li, M. Hashem Pesaran
and Dek Terrell
ADVANCES IN ECONOMETRICS VOLUME 42

THE ECONOMETRICS OF
NETWORKS
EDITED BY
ÁUREO DE PAULA
University College London, UK

ELIE TAMER
Harvard University, USA

MARCEL-CRISTIAN VOIA
University of Orléans, France
United Kingdom – North America – Japan
India – Malaysia – China
Emerald Publishing Limited
Howard House, Wagon Lane, Bingley BD16 1WA, UK

First edition 2020

Copyright © Chapter 9. ‘Survival Analysis of Banknote Circulation: Fitness, Network

Structure, and Machine Learning’, © 2020 Bank of Canada. Published under exclusive
licence by Emerald Publishing Limited. All other chapters © 2020 Emerald Publishing
Limited.

Reprints and permissions service

Contact: permissions@emeraldinsight.com

No part of this book may be reproduced, stored in a retrieval system, transmitted in any
form or by any means electronic, mechanical, photocopying, recording or otherwise without
either the prior written permission of the publisher or a licence permitting restricted copying
issued in the UK by The Copyright Licensing Agency and in the USA by The Copyright
Clearance Center. Any opinions expressed in the chapters are those of the authors. Whilst
Emerald makes every effort to ensure the quality and accuracy of its content, Emerald
makes no representation implied or otherwise, as to the chapters’ suitability and application
and disclaims any warranties, express or implied, to their use.

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

ISBN: 978-1-83867-576-9 (Print)

ISBN: 978-1-83867-575-2 (Online)
ISBN: 978-1-83867-577-6 (Epub)

ISSN: 0731-9053 (Series)

CONTENTS

Introduction Econometrics of Networks

Áureo de Paula, Elie Tamer and Marcel Voia

SECTION 1
IDENTIFICATION OF NETWORK MODELS
Chapter 1 Identification and Estimation of
Network Models with Heterogeneous
Interactions
Tiziano Arduini, Eleonora Patacchini and Edoardo
Rainone

Chapter 2 Identification Methods for Social

Interactions Models with Unknown Networks
Hon Ho Kwok

Chapter 3 Snowball Sampling and Sample

Selection in a Social Network
Julian TszKin Chan

SECTION 2
NETWORK FORMATION
Chapter 4 Trade Networks and the Strength
of Strong Ties
Áureo de Paula

Chapter 5 Application and Computation of a

Flexible Class of Network Formation Models
Seth Richards-Shubik

SECTION 3
NETWORKS AND SPATIAL ECONOMETRICS
Chapter 6 Implementing Faustmann–
Marshall–Pressler at Scale: Stochastic
Dynamic Programming in Space
Harry J. Paarsch and John Rust

Chapter 7 A Spatial Panel Model of Bank

Branches in Canada
Heng Chen and Matthew Strathearn

Chapter 8 Full-information Bayesian

Estimation of Cross-sectional Sample Selection
Models
Sophia Ding and Peter H. Egger

Chapter 9 Survival Analysis of Banknote

Circulation: Fitness, Network Structure, and
Machine Learning
Diego Rojas, Juan Estrada, Kim P. Huynh and David
T. Jacho-Chávez

SECTION 4
APPLICATIONS OF FINANCIAL NETWORKS
Chapter 10 Financial Contagion in Cross-
holdings Networks: The Case of Ecuador
Pablo Estrada and Leonardo Sánchez-Aragón

Chapter 11 Estimating Spillover Effects with

Bilateral Outcomes
Edoardo Rainone

Chapter 12 Interconnectedness Through the

Lens of Consumer Credit Markets
Anson T. Y. Ho

Chapter 13 FRM Financial Risk Meter

Andrija Mihoci, Michael Althof, Cathy Yi-Hsuan Chen
and Wolfgang Karl Härdle

Index
INTRODUCTION

ECONOMETRICS OF NETWORKS
Áureo de Paula, Elie Tamer and Marcel Voia

The Econometrics of Networks. Volume 42 of the series Advances in

Econometrics (AiE) aims at providing novel methodological and
empirical research on the econometrics of networks. The volume
includes both theoretical and empirical/policy papers with the
specific purpose of providing an opportunity for a dialogue between
academics and practitioners to better understand this new and
important area of research and its role in policy discussions.
The volume is a good resource for graduate students and
researchers. It includes 13 chapters covering various topics such as
identification of network models, network formation, networks and
spatial econometrics and applications of financial networks. One can
also learn about network models with different types of interactions,
sample selection in social networks, trade networks, stochastic
dynamic programing in space, spatial panels, survival and networks,
financial contagion, spillover effects, interconnectedness on
consumer credit markets and a financial risk meter.
The book can be also a resource for data scientists and
professionals from the industry as it provides a useful resource for
applications, such as, for example, offering theoretical discussions
about within and between groups interactions, the role unknown
networks play in social interactions and the role of strong ties on
trade networks.
A brief description of the chapters of the book which are grouped
in four sections is presented below.

SECTION 1: IDENTIFICATION OF NETWORK

MODELS
This section comprises of three chapters. The chapter by Tiziano
Arduini, Eleonora Patacchini and Edoardo Rainone, “Identification
and Estimation of Network Models with Heterogeneous Interactions,”
generalizes the standard linear-in-means model to allow for multiple
types of between and within-type interactions. It extends the
Bramoulle, Djebbari, and Fortin’s (2009) and Calvo-Armengol,
Patacchini, and Zenou’s (2009) identification conditions and Liu and
Lee’s (2010) estimation approach when network data are available
and peer effects are heterogeneous by peer-type. The proposed
methodology is inspired by Liu and Lee (2010) and Liu (2014) where
the many instruments are derived from many networks (groups)
observed in the sample. Differently, in the model proposed by the
authors the many instruments derive from the multiple subnetwork
framework. A multiple subnetwork framework does not only result in
a larger number of instruments but also yields multiple
approximations of the optimal instruments. The bias arising when
interactions are ignored is analytically derived and evaluated in finite
samples using simulation experiments. The authors show that the
form of the many-instrument bias differs, though the leading order
remains unchanged.
The chapter by Hon Ho Kwok, “Identification Methods for Social
Interactions Models with Unknown Networks,” develops a two-step
identification method for social interaction models with unknown
networks and discusses how the proposed methods are connected to
the identification methods for models with known networks. In the
first step, a linear regression is used to identify the reduced forms,
while the second step decomposes the reduced forms to identify the
primitive parameters. The proposed methods use panel data to
identify networks. Two cases are considered: the sample exogenous
vectors span Rn (long panels), and the sample exogenous vectors
span a proper subspace of Rn (short panels). For the short panel
case, in order to solve the sample covariance matrices’ non-
invertibility problem, the chapter proposes to represent the sample
vectors with respect to a basis of a lower-dimensional space so that
fewer regression coefficients are needed in the first step. This allows
for the identification of some reduced form submatrices, which
provide the equations necessary for identifying the primitive
parameters.
The chapter by TszKin Julian Chan, “Snowball Sampling and
Sample Selection in a Social Network,” studies a snowball sampling
method for social networks with endogenous peer selection.
Snowball sampling is a sampling design which preserves the
dependence structure of the network. It sequentially collects the
information of vertices linked to the vertices collected in the previous
iteration. The snowball samples suffer from a sample selection
problem because of the endogenous peer selection. The chapter
proposes a new estimation method that uses the relationship
between samples in different iterations to correct for selection. In
the application, the snowball samples collected from Facebook is
used to estimate the proportion of users who support the Umbrella
Movement in Hong Kong.

SECTION 2: NETWORK FORMATION

This section has two chapters. The chapter by Áureo de Paula,
“Trade Networks and the Strength of Strong Ties,” surveys the
relevant literature on strategic formation of networks and uses it to
motivate looking at questions related to the behavior of individuals in
the presence of imperfect market institutions. In particular, the
chapter is interested in how individuals devote resources to the
establishment of reliable connections in order to attenuate the
frictions that reduce trading and insurance opportunities by looking
to answer questions such as: When should we expect to see the
appearance of such interpersonal networks as a stable support of
economic transactions? Having been established as a stable
phenomenon, does voluntary networking improve upon the situation
in which no such connections can be established?
To answer these questions de Paula extends a trade network first
suggested by an example in Jackson and Watts (2002) and builds a
model which shows that the investment in strong ties often, though
not always, produces stable configurations that manage to improve
upon the imperfections of market institutions.
The author finds that such voluntary networks of “strong ties” can
usually be sustained as a stable outcome, though examples are not
hard to achieve in which no equilibrium configuration occurs.
Additionally, he finds that whenever such a structure exists, it
improves general well-being over a situation in which only formal
unreliable markets existed. And finally, the analysis suggests that
though voluntary networking efforts are no substitute for an
improvement in the reliability of formal institutions, emergence of
informal insurance networks or extensive investment in connections
should come as no surprise in the presence of “noisy” market
institutions.
The chapter by Seth Richards-Shubik, “Application and
Computation of a Flexible Class of Network Formation Models,”
discusses the empirical application of a class of strategic network
formation models, using the approach to identification introduced by
de Paula, Richards-Shubik and Tamer (2018). The chapter
emphasizes the interplay between model specification and
computational complexity and suggests approaches that help in
improving empirical/computational tractability. Two detailed
examples, one on friendship networks and another on coauthorship
networks, are used to illustrate these issues and to demonstrate the
performance of the approach with both simulation and empirical
evidence.
The analysis shows how the utility specification impacts
dimensionality. Additionally, the author shows how machine learning
techniques can be used for dimension reduction in the coauthorship
model to make the model computationally feasible while including a
rich set of covariates. The chapter presents a more general
estimation method, which expands the potential range of
applications. Also, a statistical inference is provided with minimal
computational burden.

SECTION 3: NETWORKS AND SPATIAL

ECONOMETRICS
This section comprises of four chapters.
The chapter by Harry J. Paarsch and John Rust, “Implementing
Faustmann–Marshall–Pressler at Scale: Stochastic Dynamic
Programming in Space,” constructs an intertemporal model of rent-
maximizing behavior on the part of a timber harvester under
potentially multi-dimensional risk as well as geographical
heterogeneity. Subsequently, the authors use recursive methods (the
method of stochastic dynamic programing) to characterize the
optimal policy function which is the rent-maximizing timber-
harvesting profile.
One feature of their application to forestry in the province of
British Columbia is the unique and detailed information that is
organized in the form of a dynamic geographical information system
which helps to account for site-specific cost heterogeneity in
harvesting and transportation, as well as an uneven-aged stand
dynamics in timber growth and yield across space and time in the
presence of stochastic lumber prices.
Their framework is a powerful tool, and by using it one can
conduct a variety of different policy experiments. First, the authors
use geography both in the planar sense and in the three-dimensional
sense. Second, they consider site-specific heterogeneity both on the
cost side in terms of harvesting and transportation and on the
growth and yield side in terms of heterogeneous stands of timber.
Third, they use best-practice biological methods to model the
dynamics of uneven-aged forest growth and yield. Fourth, in the
past, economists have typically demonstrated their methods by
solving simple, prototypical examples in closed-form or they have
imposed conditions sufficient to sign comparative static predictions.
Alternatively, in this chapter the authors use recent developments in
computational methods to solve numerically for the optimal policy
function. Finally, they compared their optimal harvesting policy with
the harvests that have occurred during the past eight years, or so,
finding striking and significant differences.
The chapter by Heng Chen and Matthew Strathearn, “A Spatial
Panel Model of Bank Branches in Canada,” empirically analyzes the
spatial bank branch network in Canada. The authors study the
market structure (both industrial and geographic concentrations) via
its own or adjacent postal areas. The empirical framework considers
branch density (the ratio of the total number of branches to area
size) by employing a spatial two-way fixed effects model.
The addition of geographic concentration, as measured by the
average distance to the closest bank branch, allows the authors to
reflect on the degree of spatial clustering among bank branches in a
given postal area. By controlling for both industrial and geographic
concentrations the authors can capture not only the degree of
competitiveness in a given area, but also how well the area is
serviced in terms of travel distance to the nearest branch.
The chapter finds that there are no effects associated with market
structure but that there are strong spatial within and nearby effects
associated with the socioeconomic variables. In addition, the chapter
studies the effect of spatial competition from rival banks and finds
that large and small banks tend to avoid markets dominated by their
competitors.
The chapter by Sophia Ding and Peter H. Egger, “Full-information
Bayesian Estimation of Cross-sectional Sample Selection Models,”
proposes an approach to estimate cross-sectional sample selection
models, where the shocks on the units of observation feature some
interdependence through spatial or network autocorrelation.
There is research that aims at addressing these two issues in
conjunction. The authors are showing that previous Bayesian
algorithms such as the ones developed by Dogan and Taspinar
(2018) do not allow for the latent variables (and the associated
random shocks) to affect the latent outcome of the units whose
outcome is observed. This may lead to biased parameters, in
particular, of the covariances of the disturbances between the
equations. This chapter improves on the prior Bayesian work on this
subject by proposing a modified approach toward sampling using the
multivariate-truncated, cross-sectionally dependent latent variable of
the selection equation. The chapter outlines the model and
implementation approach and provides simulation results
documenting improved performance.
The chapter by Diego Rojas, Juan Estrada, Kim P. Huynh and
David T. Jacho-Chavez, “Survival Analysis of Banknote Circulation:
Fitness, Network Structure, and Machine Learning,” utilizes machine
learning techniques to study the distribution network patterns of
over 900 million banknotes using an administrative data set from the
Bank of Canada’s Currency Information Management System. The
data contain information regarding the printing date, physical fitness
and where and when these banknotes return to the Bank of
Canada’s distribution centers. Having constructed networks at the
region and at the financial institution (those requesting as well as
depositing banknotes) level, the authors use a K-prototypes
clustering unsupervised machine learning algorithm to classify notes
into different types. This information is then used when fitting a
hazard model to explain how long a banknote stays in circulation.
The results show that their denominations, and not fitness
measures, are the main determinants of a banknote duration in
circulation after controlling for the network structure.

SECTION 4: APPLICATIONS OF FINANCIAL

NETWORKS
The last section comprises of four chapters and provides applications
to financial networks.
The chapter by Pablo Estrada and Leonardo Sánchez-Aragón,
“Financial Contagion in Cross-holdings Networks: the case of
Ecuador,” applies a financial contagion model proposed by Elliott,
Golub, and Jackson (2014) to a cross-shareholding network of firms
in Ecuador where the nodes are the firms and the links are the
cross-shareholdings among firms. A novel data set provided by
SUPERCIAS is used in the analysis.
The financial contagion model uses a network of financial
interdependencies among firms in a dependency matrix where each
element represents the cross-shareholding. The authors study how a
negative shock that affects one firm propagates through the network
and generates a cascade of failures. The results show that the
Ecuadorian market exhibits low levels of diversification and
integration, which means that the effects of cascades cannot be
amplified throughout the network. Low integration implies the
presence of weak links in the network. Results also show the
presence of a giant weakly connected component (40% of the total
firms) because diversification is moderate suggesting that cascade
effects are still weak.
Furthermore, a sensitivity analysis is conducted to determine
which parameters contribute to firm’s failure. When allowed the
threshold, the failure cost, and the drop market value to vary, only
two waves of contagion are noticeable. It was also found that two
firms coming from the finance and trade industry cause the highest
contagion and when a shock affects an entire industry there are
more firm failures from trade and manufacturing industries than
other industries. The results can be relevant for policymakers as they
are better able to monitor the market and anticipate future losses.
The chapter by Edoardo Rainone, “Estimating Spillover Effects
with Bilateral Outcomes,” is concerned with the estimation of
spillover effects when outcomes arise as a consequence of bilateral
interactions instead from individual actions, in other words the
analysis refers to network effects when outcomes are generated on
links and not on nodes.
With the diffusion of over-the-counter (OTC) platforms and the
advances in the economic theory related to networks, the chapter
emphasizes the importance of assessing network effects with link-
based outcomes. A link-based spatial autoregressive (SAR) model is
proposed together with identification conditions and a two step least
square (2SLS) estimation procedure. The author shows analytically
and with Monte Carlo simulations that using a standard node-based
SAR, which models nodes’ instead of links’ outcomes, produces
misleading results when the data generating process (DGP) is link-
based. The methodology is illustrated using real data from an
interbank network. The results highlight that under conditions that
are often met in OTC markets, modeling nodes’ outcomes can lead
to biased results and misleading policy implications.
The chapter by Anson T. Y. Ho, “Interconnectedness through the
Lens of Consumer Credit Markets,” looks at the interconnectedness
between financial institutions (FIs) through the lens of consumer
credits. Financial systemic risk is often assessed by the
interconnectedness of FIs in terms of cross ownership, overlapping
investment portfolios, interbank credit exposures and other factors.
Using detailed consumer credit data in Canada, this chapter
constructs a novel banking network to measure FIs’
interconnectedness in consumer credit markets. Results show that
FIs on average are more connected to each other over the sample
period, when the interconnectedness measure increases by 21%
from 2014 to 2019.
The FIs with more diversified portfolios are also more connected
in the network. Participation in mortgage markets has strong positive
influence on FIs’ connectedness, because FIs with mortgage
operations have more similar portfolios to the large FIs. Findings in
this chapter highlight the importance of FIs exposure to the
household sector, which may have important implications on
systemic risk and the risk of multiple stress incidences across FIs.
Measuring the connectedness in consumer lending networks is the
first step in quantifying the potential systemic risk generated by FIs’
consumer lending operations. Deeper understanding on how FIs
finance their consumer lending is required to further translate the
connectedness in consumer lending network into systemic risk
measures.
Finally, the chapter by Andrija Mihoci, Michael Althof, Cathy Yi-
Hsuan Chen and Wolfgang Karl Härdle, “FRM Financial Risk Meter,”
proposes a systemic risk measure that accounts for links and mutual
dependencies between financial institutions utilizing tail event
information. The proposed Financial Risk Meter (FRM) is based on a
Lasso quantile regression and it is designed to capture tail event co-
movements. The FRM focus lies on understanding active data sets
characteristics and the interdependencies in a network topology.
The focus of the chapter is on two selected FRM indices, namely
FRM@Americas and FRM@Europe for the equity markets, and
SRM@EuroArea as an application to the asset class of government
bonds. Augmenting them, for example, by simultaneously checking
varieties of quantiles of FRM components, one can monitor economic
activity and network dynamics, and suggest further improvements in
portfolio risk management.
The chapter’s findings are: first, FRM correlates positively with
other measures of systemic risk and peaks around crises; second, a
detailed inspection of the active set across time allows to detect the
network’s nodes presenting the highest risk of spillover and third,
FRM is shown to predict upcoming recession periods and serves as a
leading indicator for systemic risk in a variety of world regions, the
US and the EU market.
Therefore, the FRM can be viewed as an early recession indicator
that can help to detect distressed areas in the financial system
network consisting of banks and non-banks, and thereby can help
prevent spillovers into the wider financial industry. Finally, FRM can
measure tail event risk, accounts for network dynamics
characteristics and offers a flexible risk measuring platform.
In practice, FRM can be applied to the return time series of
selected financial institutions and macroeconomic risk factors.

REFERENCES
Bramoulle, Y., Djebbari, H., & Fortin, B. (2009). Identification of peer effects through social
networks. Journal of Econometrics, 150, 41–55.
Calvo-Armengol, A., Patacchini, E., & Zenou, Y. (2009). Peer effects and social networks in
education. The Review of Economic Studies, 76(4), 1239–1267.
Chan, T. J. (2020). Snowball sampling and sample selection in a social network. In
Advances in Econometrics (Vol. 42).
Chen, H., & Strathearn, M. (2020). A spatial panel model of bank branches in Canada. In
Advances in Econometrics (Vol. 42).
de Paula, A. (2020). Trade networks and the strength of strong ties. In Advances in
Econometrics (Vol. 42).
de Paula, A., Richards-Shubik, S., & Tamer. E. (2018). Identifying preferences in networks
with bounded degree. Econometrica, 86(1), 263–288.
Ding, S., & Egger, P. (2020). Full-information Bayesian estimation of cross-sectional sample
selection models. In Advances in Econometrics (Vol. 42).
Dogan, O., & Taspinar, S. (2018). Bayesian inference in spatial sample selection models.
Oxford Bulletin of Economics and Statistics, 80(1), 90–121.
Elliott, M., Golub, B., & Jackson, M. O. (2014). Financial networks and contagion. American
Economic Review, 104(10), 3115–3153.
Estrada, P., & Sanchez-Aragon, L. P. (2020). Financial contagion in cross-holdings networks:
The case of Ecuador. In Advances in Econometrics (Vol. 42).
Ho, A. T. Y. (2020). Interconnectedness through the lens of consumer credit markets.
Advances in Econometrics (Vol. 42).
Jackson, M., & Watts, A. (2002). The evolution of social and economic networks. Journal of
Economic Theory, 106(2), 265–295.
Kwok, H. H. (2020). Identification methods for social interactions models with unknown
networks. In Advances in Econometrics (Vol. 42).
Liu, X., & Lee, L. F. (2010). GMM estimation of social interaction models with centrality.
Journal of Econometrics, 159, 99–115.
Liu, X. (2014). Identification and efficient estimation of simultaneous equations network
models. Journal of Business & Economic Statistics, 32(4), 516–536.
Mihoci, A., Althof, M., Chen, C. Y.-H., & Hardle, W. K. (2020). FRM financial risk meter. In
Advances in Econometrics (Vol. 42).
Paarsch, H. J., & Rust, J. (2020). Implementing Faustmann–Marshall–Pressler at scale:
Stochastic dynamic programming in space. In Advances in Econometrics (Vol. 42).
Rainone, E. (2020). Estimating spillover effects in OTC networks. In Advances in
Econometrics (Vol. 42).
Richards-Shubik, S. (2020). Application and computation of a flexible class of network
formation models. In Advances in Econometrics (Vol. 42).
Rojas, D., Estrada, J., Huynh, K. P., & Jacho-Chavez, D. T. P. (2020). Survival analysis of
banknote circulation: Fitness, network structure, and machine learning. In Advances
in Econometrics (Vol. 42).
SECTION 1
IDENTIFICATION OF NETWORK MODELS
CHAPTER 1

IDENTIFICATION AND ESTIMATION OF

NETWORK MODELS WITH
HETEROGENEOUS INTERACTIONS
Tiziano Arduini, Eleonora Patacchini and Edoardo
Rainone

ABSTRACT
The authors generalize the standard linear-in-means model to
allow for multiple types with between and within-type
interactions. The authors provide a set of identification
conditions of peer effects and consider a two-stage least
squares estimation approach. Large sample properties of the
proposed estimators are derived. Their performance in finite
samples is investigated using Monte Carlo simulations.
Keywords: Networks; heterogeneous peer effects; spatial
autoregressive model; two-stage least squares; efficiency
JEL classifications: C13; C21; D62

1. INTRODUCTION
Interaction among agents can be modeled in several ways. When
the exact topology of connections is known, one possibility is to
consider the peer effects that stem from the given network
structure. There is a large and growing literature on peer effects in
economics using network data.1 A popular model employed in
empirical work is the linear-in-means model (Manski, 1993). Three
assumptions underlie this statistical model: (i) the network is
exogenous; (ii) the effects of all peers are equal; and (iii) peer status
is measured without error.
This chapter is concerned with the specification and estimation of
a peer effects model when Assumption (ii) is removed. Assumptions
(i) and (iii) are maintained.2 Specifically, we consider a model of
peer effects where different types of peers are allowed to exert a
different influence and where social interactions are different
between and within types.
We extend the conventional identification conditions of network
models (Bramoullé, Djebbari, & Fortin, 2009; Calvó-Armengol,
Patacchini, & Zenou, 2009) to multiple endogenous variables and
multiple adjacency matrices. We propose efficient two-stage least
squares (2SLS) estimators using instruments based on the multiple
reduced form approximations. We show that the standard IV
approximation (Kelejian & Prucha, 1998, 1999; Liu & Lee, 2010)
involves a very large number of IVs, even if we use a low degree
approximation of the optimal instruments.3 For this reason, we
consider many-instrument asymptotics (Bekker, 1994) allowing the
number of IVs to increase with the sample size. We also propose a
many-instrument bias-correction procedure. Simulation experiments
show that the bias-corrected estimator performs well in finite
samples. Finally, we investigate the bias occurring when the
interaction structure is misspecified. We derive analytically the bias
that occurs when only within-type peer effects are considered, that
is, when interactions between types are at work but ignored by the
econometrician. We then use a simulation experiment to evaluate
this bias in finite samples.
 Our framework is a generalization of the model proposed by
Arduini, Patacchini, and Rainone (2019) to study treatment effects
with heterogeneous externalities. In their model connections are
defined by membership in a given group and all agents within the
group are connected. The availability of information on the precise
structure of interactions between agents of different types offers
alternative identification conditions based on the sparsity of
interactions within or between agent-types.4
There is a long tradition in spatial econometrics looking at spatial
autoregressive models with multiple endogenous variables (see
Kelejian & Prucha, 2004). In the spatial econometrics context,
however, the adjacency matrix is the same for all endogenous
variables. The presence of different adjacency matrices provides
alternative possibilities to identify the model that have not been
explored so far.
Our methodology is inspired by Liu and Lee (2010) and Liu
(2014). Differently from their approach where the many instruments
derive from the many networks (groups) observed in the sample, in
our model the many instruments derive from the multiple
subnetwork framework. A multiple subnetwork framework does not
only result in an increasing number of instruments but also yields
multiple approximations of the optimal instruments. We show that
the form of the many-instrument bias differs, though the leading
order remains unchanged.5
This chapter is organized as follows. The next section introduces
the econometric model. Identification conditions are given in Section
3, and in Section 4 we consider 2SLS estimation for the model.
Section 5 investigates the bias occurring when the interaction
structure is misspecified. Section 6 concludes.

2. THE NETWORK MODEL WITH

HETEROGENEOUS PEER EFFECTS
Suppose the n observations in the data are partitioned into
networks, with nr agents in the rth network. For the rth network, let

where Yr = (y1r, …, ynr)′ is an nr-dimensional vector of outcomes,

Gr = [gij,r] is an nr × nr adjacency matrix, gij,r is equal to 1 if i and
j are connected, 0 otherwise. Xr is a nr × p matrix of exogenous
variables capturing individual characteristics. For ease of
presentation, we assume Gr and Xr are non-stochastic. εr = (ε1r, …,
εnr)′ is a vector of errors whose elements are i.i.d. with 0 mean and
variance for all i.6
For model (1), ϕ represents the endogenous effect, where an
agent’s choice/outcome may depend on those of his/her peers on
the same activity, and γ represents the contextual effect, where an
agent’s choice/outcome may depend on the exogenous
characteristics of his/her peers.
Let us suppose there is a finite number of types of agents in the
population. For simplicity, let us consider two types a and b. A and B
are thus sets that partition {1,…, n}. The na and nb denote the
cardinalities of A and B. Let and β = (β ′, γ ′)′.
Let us now define

where Ga,r is

formed only among nodes of type A and Gab,r keeps trace of links
from b to a.7 Regularity conditions are listed in Appendix 1. Model
(1) can be generalized in the following way:
where

and εa,r and εb,r are i.i.d. errors with variance and
respectively. Let us suppose for simplicity that If we
stack up equations (2) to (3) and restrict the endogenous effect
parameters of the two equations to be the same (i.e., ϕa = ϕb = ϕab
= ϕba), then we obtain model (1). Let us define the following
matrices

where Am = (Xa,r, Gab,rXb,r, εa,r), Bm = (Xb,r,

Gba,r Xa,r, εb,r) and By plugging Yb,r in equation (2)
we have

where provided that || ϕbGb,r ||∞

< 1, where || ⋅ ||∞ is the row-sum matrix norm. The ijth element of
Jb,r sums all k-distance paths from j to i when i, j ∈ B scaling them
by 8 Therefore the reduced form of model
(2) is

where Ma,r = (I − ϕaGa,r − ϕabϕbaCa,r)−1.9 A sufficient condition

for the non-singularity of (I − ϕaGa,r − ϕabϕbaCa,r) is || ϕaGa,r ||∞
+ || ϕabϕbaCa,r ||∞ ≤ 1. This condition also implies that Ma,r is
uniformly bounded in absolute value.10
We note that: (i) we present an aggregate model specification
(i.e., G which multiplies y in model (1) is not row-normalized), but
the approach also applies to an average model (i.e., when G which
multiplies y in model (1) is row-normalized)11; (ii) our model
specification has two types, but all the assumptions, propositions
and proofs can be naturally extended to a finite number of types;
(iii) we consider a single network, but the approach can be easily
extended to the case of multiple networks (i.e., a network with
several components) with the addition of network fixed effects in the
model specification12; and (iv) we can also add a heterogeneous
spatial lag in the error term εa = ρaWaεa + ρabWabεb.13

3. IDENTIFICATION
For notational convenience, from now on we omit the network
observation index. Let us define
Endogenous effects in equation (2) are identified if E(Za) has full
column rank for large n.14 In this section, we find sufficient
conditions for E(Za) to have full column rank.15 The detailed proof is
given in Appendix 3. Here, identification means that a consistent
estimator of the parameters of model (2) exists.
Proposition 1. Let Xa and Xb have full column rank. If Ma, Mb,
Ja and Jb are invertible,16 then E(Za) has full column rank in the
following cases

1. (a) i. βaϕa + γa ≠ 0,
ii. Ia, Ga and are linearly independent.
[and]
(b) i. βbϕb + γb ≠ 0,
 ii. Gab and GabGb are linearly independent.
[or]
2. (a) i. γab ≠ 0,
ii. Gab and GaGab are linearly independent.
[and]
(b) i. γba ≠ 0,
ii. Ia, Ga and GabGba are linearly independent.

Fig. 1. Identification with Heterogeneous Nodes.

Proposition 1 generalizes the identification conditions in Bramoullé et

al. (2009). Note that conditions (1a) are exactly the same
identification conditions found by Bramoullé et al. (2009) in the case
of homogeneous effects (i.e., only one type). Proposition 1 here is
more general as it provides alternative possibilities. When more than
one type is considered we do not need linear independence of a
particular set of matrices – we have multiple sufficient conditions.
Even if Ia, Ga and are linearly dependent we can still identify ϕa,
and the other parameters, relying on linear independence of network
paths passing through type B nodes.17 The set of adjacency
matrices’ combinations can be represented as a Tree-indexed Markov
chain – the parameters can be identified because of the multiple
branches of the tree (see Appendix 3). Obviously, if Ga, Gba, Gab
and Gb are complete and consequently all products among them are
linearly dependent, then the parameters of the model remain not
identified. However, if type A nodes are in a complete network, but
the matrices representing between-type interactions are sparse (i.e.,
Gab and Gba are not complete), then identification can be achieved
and ϕa can be estimated even if Ga is complete. Systems in panels
(b) and (c) in Fig. 1 can be identified because the adjacency matrix
of type B nodes (blue nodes in Fig. 1) is sparse, whereas systems in
panels (a) and (d) cannot. The additional parameters’ restrictions
(condition (1b, 2a or 2b)) are due to an additional vector in the full
rank condition (i.e., E(Gabyb)).
Proposition 1 has a natural interpretation in terms of instrumental
variables. A multiple type framework adds an extra layer of exclusion
restrictions. In fact, multiple sets of matrices provide additional
instruments. The intuition is that when we distinguish nodes in
different types, a higher number of possible network intransitivities
are formed. Appendix 2 provides technical details on the connection
between identification in a single type model and a multiple type
one.

4. THE 2SLS ESTIMATOR

We consider 2SLS estimation for the model in the spirit of Liu and
Lee (2010). Following the standard technique used in spatial
econometrics literature, we have the following optimal instruments
from the two (symmetric) reduced forms
Recalling that Za = [Gaya, Gabyb, E(Am)] is a n × (k + 2) matrix,
we have fa = E(Za) = [E(Gaya), E(Gabyb), E(Am)]. Therefore, from
equations (6) and (7) we have

where e1 is a first unit vector of dimension (k + 2), Sa = GaMa and

Sab = GabMb. These instruments are infeasible given the embedded
unknown parameters. fa can be considered a linear combination of
IVs in E(Am)), Sab (GbaJaE(Am), E(Bm)),
E(Am)). Furthermore, since Sa = GaMa and Sab = GabMb provided
that || ϕaGa ||∞ + || ϕabϕbaCa ||∞ ≤ 1 and || ϕbGb ||∞ < 1, we
have

The same approximation holds for Sab. It follows that

This implies
Hence, the approximation error by series expansion diminishes very
quickly in a geometric rate, as long as the degree of approximation
(p) increases as n increases. We can also replace Sa and Sab by a
linear combination. The instruments become

with an approximation error diminishing very quickly when K (or p)

goes to infinity, where K denotes the number of instruments. Let us
define as the matrix of instruments and select
a na × K submatrix HK based on a p-order approximation of H∞.18
For instance, if we use the second-order approximation of the
infinite sums, will be the first step best
projector. The feasible 2SLS estimator for model (2) is

where

4.1. Asymptotic Properties

This section derives the asymptotic properties of the many
instrument 2SLS estimator for heterogeneous network models.
The following propositions establish the consistency and
asymptotic normality of the many-instrument 2SLS estimator in
equation (8). We assume an i.i.d. sample of size from a population
of networks with a fixed and known structure.19 Regularity
conditions together with some discussion can be found in Appendix
1. Some useful Lemmas are provided in Appendix 2. All the proofs
are listed in Appendix 3.
 Let

20

PKSa = ψa and PKTba = Ξba, where Tab = SabGbaJb.21

Proposition 2. Under Assumptions 1–5, if

where

From Proposition 2, when the number of instruments K grows at a

slower rate than the sample size n, the 2SLS estimator is consistent
and asymptotically normal. However, the asymptotic distribution of
the 2SLS estimator may not be centered around the true parameter
value due to the presence of many-instrument bias of order Op(K/n)
(see e.g., Liu & Lee, 2010). We note that the leading order of the
bias is the same as in Liu and Lee (2010) and Liu (2014). However,
the structure of the bias differs. Here, it depends on multiple
approximations of the optimal instruments (see the beginning of
Section 4). The condition that K/n → 0 is crucial for the 2SLS
estimator to be consistent. This is evident if we look at the normal
equation of our estimator: When we have

that by
Lemma B.2 in the Appendix. This converges to 0 only if the number
of instruments grows more slowly than the sample size.22 Similarly
to Liu and Lee (2010), the following corollary characterizes different
scenarios for different rates in which K diverges from n.
Corollary 1. Under Assumptions 1 –5 , (i) if
(ii) if
 where

The many-instrument bias of the 2SLS estimator can be corrected

by the estimated leading-order bias (b) given in Proposition 2. Given
consistent estimates of the bias-corrected
2SLS estimator is

The following proposition shows that the bias-corrected estimator

is properly centered around the normal distribution.
Proposition 3. Under i, if
−consistent initial
estimators, then

4.2. Finite Sample Performance

In this section, we use simulation experiments to investigate the
performance of the proposed estimator in small samples.
We conduct a Monte Carlo simulation study based on the
following model

where Xa, Xb and ε ∼ N(0, 1). Borrowing from Liu and Lee (2010),
we generate the G matrix as follows. First, for the ith row of G, we
generate an integer di ε [0, 1, …, m] with a uniform probability
function, where m = 10, 20, 30. Then we set the (i + 1)th, ···, (i +
di)th elements of the ith row of G to be ones. If (i + di)th < na, the
other elements in that row are zeros; otherwise, the entries of ones
will be wrapped around such that the number of (di − na) entries of
the ith row will be ones. We partition the matrix into four
submatrices Ga, Gb, Gab and Gba with a random selection of rows
and correspondent columns. The identifier variable used to select
the two types is generated by a Bernoulli distribution with p = 0.5.
The number of replications is 1,000 and na = nb = 500. We perform
two experiments that are summarized in Tables 1 and 2. Each
column reports mean and standard error (in parenthesis) of the
empirical distributions of different estimators. The first column
shows 2SLS few IVs. It is based on equation (8) with the IV matrix
HK derived by the first-order approximation of the best instruments
(K = 24). The second column reports the 2SLS many IVs, it is
derived by the second-order approximation of the best instruments
(K = 84). Finally, the third column shows the 2SLS bias-corrected. It
is based on equation (9) with consistent estimates derived from the
2SLS few IVs.
Table 1 reports the performance of the estimators when changing
the density of the network, that is, the number of connections. Each
panel represents a different value of m, which indicates the
maximum number of connections. The data are generated with βa =
βb = γa = γb = γab = γba = 0.5. The peer effects parameters are
set to: ϕa = ϕb = 0.1 and ϕab = ϕba = 0.2. The results show that all
estimators perform well, with different nuances. In particular, one
can observe the trade-off between bias and efficiency for the 2SLS
many IVs when network density increases the higher the density, the
higher the gain in terms of efficiency with respect to the 2SLS few
IVs. However, the bias (due to the many instruments) increases as
well. The bias correction that we propose is thus particularly
beneficial when the network is dense. Table 2 reports the
performance of the estimators when changing the heterogeneity of
within and between-type parameters. The simulation setup remains
unchanged, but we now set the maximum number of connections to
20 and let the ϕ parameters vary. In the first panel, we consider ϕa
= ϕab = ϕb = ϕba = 0.1. This is the benchmark framework in which
peer effects are homogeneous. In the second panel, we introduce
some heterogeneity in the within-type interaction effects. We set ϕa
= ϕb = 0.1 and ϕab = ϕba = 0.3. In the third panel, peer effects are
different both within and between types. We set ϕa = 0.1, ϕb = 0.2
ϕab = 0.4 and ϕba = 0.05. Table 2 shows that the performance of
the estimators does not depend on the values of the parameters –
the ranking of the estimators in terms of efficiency and bias remains
unchanged.

5. MODEL MISSPECIFICATION BIAS

In this section, we investigate the bias occurring when the
interaction structure is misspecified.
First, we analytically derive the bias that occurs when only within-
type peer effects are considered, that is, when interactions between
types are at work but ignored by the econometrician. We then use a
simulation experiment to evaluate this bias in finite samples.23 Let
us suppose the econometrician estimates the following model

whereas the real DGP is

This model misspecification results in an estimator of the

endogenous effect ϕa that is inconsistent. First, we are omitting the
influence of the outcome of type B agents. Second, we do not
consider the indirect connections among type A nodes passing
through type B nodes. As a result, with κ ≥ 2, is misspecified.
Therefore, the commonly used instrument might not be valid as
the exclusion restrictions might be violated. Third, we misspecify the
contextual effects (GaXa) by ignoring the characteristics of other-
type peers. Analytically, the bias is

where Za = [Gaya, Xa, GaXa] and The bias is

positively correlated with the direct influence of type B on type A, as
captured by the peer effects from B to A and the influence of the
characteristics of type B on type A.
Table 3 shows the extent of this bias in finite samples through a
Monte Carlo simulation and represents the performance of the 2SLS
few IVs following the same experiment design as in the previous
section.24 We report on the case in which the maximum number of
connections is 10 for each node (as in panel 2 of Table 1).25 The
first column reports the real value of the parameters. The second
column shows the performance of the 2SLS estimator in the
misspecified model. When interactions between types are at work
but ignored by the econometrician it results in the size of the bias
derived above. The third column shows the results of the estimator
when the econometrician considers the correct DGP (equations 11
and 12), but does not use the approximation of optimal instruments
(equation 8). In other words, we consider the case where the
traditional network IV approach is applied mechanically, thus
and GabGbXb are used as instruments, respectively, for GaYa and
GabYb. In short, only within-type instruments are considered. The
resulting 2SLS estimator is consistent but not efficient. The fourth
column reports the performance of our 2SLS few IVs (equation 8),
which considers the Hk matrix derived in Section 4 (i.e., which also
includes between type instruments).26 Mean values for each
coefficient’s empirical distribution and standard errors (in
parenthesis) are reported.
Table 3 shows that the bias is large in the second column,
especially for the β coefficients. In the second column the bias is not
large, but the problem is efficiency. Our approach (third column)
reveals no bias and improved efficiency.

6. CONCLUSION
This chapter extends the linear-in-means model with endogenous
and contextual peer effects by allowing them to differ within and
between types. We extend the Bramoullé et al. (2009) and Calvó-
Armengol et al. (2009) identification conditions and Liu and Lee
(2010) estimation approach when network data are available and
peer effects are heterogeneous by peer type. The bias arising when
interactions are ignored is analytically derived and evaluated in finite
samples using simulation experiments.

Table 1. Monte Carlo Simulations – Changing Density.

Note: 1,000 observations, 1,000 replications. yb is generated with ϕb = 0.1, ϕba = 0.2, βb
= 0.5, γb = 0.5, γba = 0.5.

Table 2. Monte Carlo Simulations – Changing Parameters.

Note: 1,000 observations, 1,000 replications.

Table 3. Monte Carlo Simulations – Misspecification Bias.

Note: 1,000 observations, 1,000 replications, 10 max connections, ϕb = ϕab = ϕba = 0.3,
βb = 0.5, γb = γab = γba = 0.5.

NOTES
1. See Jackson and Zenou (2014) and Jackson, Rogers, and Zenou (2017) for a collection
of recent studies.
2. For recent reviews of models and methods aiming at removing (i) and (iii) see Hsieh,
Lin, and Patacchini (2019) and Advani and Malde (2018), respectively. The approach
proposed here can also be used in conjunction with methods dealing with both (iii) and (i).
 3. See Prucha (2012) for a review of Generalized Method of Moments estimators in a
spatial framework.
4. We also show that identification in this model is related to the concepts of chains and
Tree-indexed Markov chains (see Benjamini & Peres, 1994).
5. Linear-in-means models with heterogeneous peer effects may also be estimated using
Bayesian methods (see e.g., Goldsmith-Pinkham & Imbens, 2013; Hsieh & Lin, 2017). While
Bayesian methods are computationally challenging, our IV approach is of immediate
applicability for the practitioner.
6. Observe that εr generally depends upon n. Thus, the vector is a triangular array. We
suppress the subscript n for notational convenience. See Kelejian and Prucha (1998) for
details.
7. More formally,
where Ra = (Ina,r,, Ona,r,nb,r) and Rb = (Onb,r,na,r, Inb,r) are matrices that select the
nodes in types a and b, respectively. Ok,l is a k × l matrix of 0’s.
8. Ca,r is a matrix which captures all the indirect connections among nodes of type A
passing through one or more nodes of type B. Note that the ijth generic element of
Gab,rGba,r is equal to the number of length-2 paths directed from j ∈ A to i ∈ A passing
through a node l ∈ B. This matrix accounts only for distance-2 indirect connections while
Ca,r = Gab,rJb,rGba,r captures all the paths starting from j ∈ A and ending to a generic
node in B, eventually passing through other nodes of type B and finally arriving in i ∈ A
scaling them by ϕb.
9. This matrix captures all direct and indirect paths among type A nodes passing through
others type A nodes and type B nodes.
10. The assumption is crucial for identification of the model and asymptotic normality of
the estimator (see Appendix 1 for a detailed definition of a matrix uniformly bounded in
absolute value).
11. Aggregate and average models are different in terms of behavioral foundations,
contextual effects are supposed to be averages over peers in both cases w.l.o.g. see (Liu,
Patacchini, & Zenou, 2014).
12. In this case, we only have to transform equations (2)–(3), using a deviation from
group mean projector (see Liu, 2014). All the statistical results of the proposed estimator
still hold.
13. The resulting model is a SARARMAG(p; q; g) with p = 1, q = 1 and g = 2, where p
and q are the number of spatial lags for outcome and error, respectively, and g is the
number of types (see Kelejian & Prucha, 2007).
14. This implies that Assumption 4 in Appendix 1 holds.
15. Symmetric conditions and results hold for equation (3).
16. In practice we need a series expansion to approximate the inversion of the matrices.
17. For example, we can take advantage of linear independence of Ia, Ga and
(instead of Ia, Ga and ); and Gab and GaGba.
18. Note that K is a function of the degree of approximations p.
19. Our results also hold true if we observe an i.i.d. sample from a population of
networks with a stochastic but strictly exogenous structure.
20. This is a crucial assumption. See the discussion in Appendix 1 after Assumption 4.
21. To simplify the notation, we assume that n → ∞ implies na → ∞ and nb → ∞.
 22. Indeed, if we use a fixed number of instruments given by the asymptotic

distribution will be Note that

which is positive semi-definite in

general. The 2SLS estimator with fixed number of instrument is generally not efficient. In
order to have efficiency, we need to index our matrix of instruments with K and let K grow
more slowly than the sample size.
23. Observe that the literature looking at spillovers on networks with missing data (point
(iii)) aims at obtaining a consistent estimator for the spillover parameter once missing data
are taken into account. Our context is different, since it allows to estimate multiple different
values for the spillover effects.
24. We use the 2SLS few IVs to ease the comparison of 2SLS estimators with the
misspecified set of instruments. Observe that the bias considered here is due to the
misspecification of the model rather than the many instrument issue.
25. The simulation results for the other cases, that is, when the maximum number of
simulations is 20 or 30, are very similar.
26. First-order approximation of optimal instruments is considered.
27. See Staiger and Stock (1997) or Baltagi, Kao, and Liu (2012) for a panel data version
of weak instrument asymptotics. Another interesting extension could be to derive the
estimator’s asymptotic properties under many weak instruments. In doing so, we are
allowing the rate of concentration parameter to be different than the rate of the sample
size. Consequently, we can compare it with the rate in which K increases. See for example,
Chao and Swanson (2005).
28. The matrices sequence is multiplied by Xa or Xb depending on the last interaction
matrix. For instance is multiplied by Xb while GbGba is multiplied by Xa.
29. In this notation a chain includes all possible paths that have common features. For
instance, all of paths are starting from A and arriving to B in the same chain.
30. Note that H1 is the IV matrix considered in Section 4, which is approximated by HK in
the feasible 2SLS estimation.
31. Note that here we need at least three chains from Ca and two from Cab because we
are considering the outcome equation for type A nodes, that is, the starting point of chains
is always type A node.
32. Given that here we are not interested in determining the transition probability law of
a chain, even if it is simple to estimate and is basically the link formation probability
considered for all of the possible combinations of nodes’ type. Benjamini and Peres (1994)
give a detailed discussion on Tree-indexed Markov chain.
33. It is equivalent to say that the probability 0 < P (gij = 1) < 1, i ∈ A, B and j ∈ A, B.
Note that transition probability can be derived from Ga, Gb, Gab and Gba. Here we are
simply excluding the classical linear in mean framework (when the matrices are complete)
and the case in which there are no connections (when the matrices are empty).
34. From a Markov Chain perspective again, a more restrictive condition consists in
assuming that the underlying Markov Chain is irreducible and aperiodic. This means that
type A is connected with type B or type A with the same probability (and the same holds for
type B). Thus, in this case tree branches with the same length have the same probability of
being observed. The aperiodicity and irreducibility are not necessary for the identification
condition to hold, but of course are sufficient.
35. Holding condition (2) instead of (1). We basically take advantage of linear
independence of Ia, Ga and GabGba instead of
36. The additional parameter restrictions (conditions (1b, 2a, or 2b) in Proposition 1) are
basically due to an additional vector in the full rank condition (i.e., E(Gabyb)).
37. Borrowing from Markov chains vocabulary again, this is because the state that
characterizes the chain is only one (A).

REFERENCES
Advani, A., & Malde, B. (2018). Credibly identifying social effects: Accounting for network
formation and measurement error. Journal of Economic Surveys, 32(4), 1016–1044.
Arduini, T., Patacchini, E., & Rainone, E. (2019). Treatment effects with heterogeneous
externalities, Journal of Business & Economic Statistics, 1–13.
https://doi.org/10.1080/07350015.019.159755
Baltagi, B. H., Kao, C., & Liu, L. (2012). On the estimation and testing of fixed effects panel
data models with weak instruments. Advances in Econometrics, 30, 199–235.
Bekker, P. A. (1994). Alternative approximations to the distributions of instrumental variable
estimators. Econometrica, 657–681.
Benjamini, I., & Peres, Y. (1994). Markov chains indexed by trees. The Annals of Probability,
22, 219–243.
Bramoullé, Y., Djebbari, H., & Fortin, B. (2009). Identification of peer effects through social
networks. Journal of Econometrics, 150, 41–55.
Calvó-Armengol, A., Patacchini, E., & Zenou, Y. (2009). Peer effects and social networks in
education. The Review of Economic Studies, 76(4), 1239–1267.
Donald, S., & Newey, W. (2001). Choosing the number of instruments. Econometrica, 69(5),
1161–1191.
Goldsmith-Pinkham, P., & Imbens, G. (2013). Social networks and the identification of peer
effects. Journal of Business and Economic Statistics, 31, 253–264.
Hsieh, C.-S., & Lin, X. (2017). Gender and racial peer effects with endogenous network
formation. Regional Science and Urban Economics, 67, 135–147.
Hsieh, C.-S., Lin, X., & Patacchini, E. (2019). Social interaction methods. In Handbook of
labor, human resources and population economics.
Jackson, M. O., Rogers, B. W., & Zenou, Y. (2017). The economic consequences of social-
network structure. Journal of Economic Literature, 55(1), 49–95.
Jackson, M. O., & Zenou, Y. (2014). Games on networks. In Handbook of game theory (Vol.
4).
Kelejian, H., & Prucha, I. R. (1998). A generalized spatial two-stage least squares procedure
for estimating a spatial autoregressive model with autoregressive disturbances. The
Journal of Real Estate Finance and Economics, 17(1), 99–121.
Kelejian, H., & Prucha, I. R. (1999). A generalized moments estimator for the
autoregressive parameter in a spatial model. International Economic Review, 40(2),
509–533.
Another random document with
no related content on Scribd:
You may laugh, children, but it is true. The dolphin had a servant,
who was also a dolphin, but of the family of the Globiceps. These are
so called because of their round heads, which look like the globes
used in the electric lighting of streets.
The young dolphin was playing in the water. He tried to attract
Pinocchio’s attention in many ways. He spouted water through the
hole which every dolphin has at the top of his head. He called to the
marionette. He smiled at the youngster. It was of no use. Pinocchio,
with his wooden nose in the air and his dough cap on one ear, would
not even turn his head.
“I wonder if he is deaf or blind?” the dolphin finally said, loudly
enough to be heard.
Pinocchio turned with a start.
“For your own benefit, I just wish to say that I am not now and never
have been deaf,” he said as haughtily as he could.
“Then why do you look at me in that fashion? And why don’t you
answer me?” was the reply.
“I am acting just as a gentleman should toward those who are
beneath him,” said Pinocchio.
“I don’t know which of us is the better of the two. All I do know is, that
my father was the richest inhabitant of the sea and that the other
dolphins considered him their king.”
“King?” mumbled Pinocchio, who knew himself to be the son of a
poor carpenter, earning so little that he never had a penny in his
pocket.
“But king or not, what does it matter? In this world we are all equal,
for we have all been created by God. Listen, my dear marionette.
Come here. As we are to travel such a long distance together, we
should be friends. Are you willing to be my friend?”
These pleasing words made Pinocchio see how stupid and how rude
he had been.
“Think of it! A fish (oh, no, I mean a sea animal) giving me lessons in
politeness!” Then turning to the dolphin, he said, “Yes, we shall be
friends. What is your name?”
“Marsovino. And yours?”
“Pinocchio.”
“A beautiful name. Come, shake hands.”
“Very willingly,” replied Pinocchio.
The good little animal stuck one of his fore fins out of the water for
Pinocchio to shake.
“And what is the tutor’s name?” said the boy of wood to the boy of
the sea.
“The tutor is a dolphin of the Tursio family, but I call him father. Is it
true that you are coming with us on our travels?”
“Yes,” said the marionette, proudly. “And I am able to teach you.”
“Teach me! That’s strange. How do you expect to teach me?”
“You will soon find out. You talk rather disrespectfully to me. I have
been in all the schools of the kingdom. And you? You probably have
never been on land for twenty-four hours.”
Marsovino looked at the marionette smilingly, but made no reply.
Pinocchio walked up and down with his hands in his pockets and his
hat at an angle of forty-five degrees, ruffling his feathers at the
brilliant remark he had made.
As soon as Tursio came near, Marsovino asked him if he were ready.
“Yes. Everything is finished,” was the reply. “Are you ready,
Pinocchio?”
“Yes. I am ready. Let us start.”
“Start? How? Do you mean to say that you are coming under the sea
with that suit?”
“Of course. It’s the only one I have.”
“A suit of paper! The very idea! Luckily I have prepared for this.
Here, Globicephalous,” he said to his servant, “give me that little suit
of ray leather,—the one I had you make this morning.”
“Splendid,” cried Pinocchio, clapping his hands. “Now I have a new
suit.”
Putting it on, he looked at himself in the water. Seeing how dark and
unbecoming it appeared, he turned to Tursio and said excitedly:
“I don’t want this. It is too ugly. I like my pretty flowered-paper one
better.”
“Your paper one Globicephalous will carry in his satchel for you.
Should you wear it in the water, it would be spoiled.”
“I want my pretty suit,” insisted Pinocchio. “If any one saw me in this
thing, he would ask me if I had been through the coal-hole.”
“But yours will be ruined if you wear it in the water, I tell you.”
“I want mine. I want mine,” wailed Pinocchio.
“Very well. Globicephalous, take the paper suit out of the traveling
bag and give it to the boy.”
The marionette turned, expecting to see an ordinary traveling bag.
Instead, he saw Globicephalous take an enormous oyster out of the
water.
“Isn’t that strange! Oyster shells for a traveling bag!”
“Strange? Why, what is strange about that?” asked Tursio.
“What is its name?” asked Pinocchio.
“That is the giant Tridacna. They are the largest oyster shells
known.”
“How large the animal inside must be,” observed Pinocchio, with a
yawn.
“Yes. It is very large, and also very beautiful. The center of the body
is a violet color dotted with black. Around this is a green border. At
the extreme edge the colors change from deepest to lightest blue.
Yes, indeed. It is very beautiful.”
“What a good meal it would make,” thought Pinocchio. His only wish
was for a good dinner, but in order to be polite he said, “Who would
ever think that there are such things under the sea!”
“Why, you have been in every school in the kingdom and don’t know
that?”
“Books on the subject you can find everywhere.”
Pinocchio bit his lips, but did not say a word. Quickly he dressed
himself again in his paper suit and declared himself ready to start.
“All right! Come along!” said the dolphin, stretching a fin out to help
Pinocchio along.
The marionette started to walk into the water. He had not gone far,
however, before his paper suit began to leave him. Hastening back
to the shore, he very meekly put on the ray-leather suit which
Globicephalous handed to him.
“Remember, my boy,” said Tursio, “that in this world of ours we must
think not only of the beauty but also of the usefulness of things. Also,
do not forget that a boy who never learns anything will never be
anything.”
“But I have learned much,” answered Pinocchio. “To prove this to
you, I can now tell you of what material this suit is made.”
“I have told you already. It is of ray leather. Do you know what a ray
is?”
“Surely I know. You may give it another name. Still, it must be that
white animal on four legs. You know. The one the shepherds shear
during some month or other.”
“Mercy!” cried Tursio. “You are talking about sheep. They give wool
to man.”
Pinocchio, without moving an eyelid, went on:
“Yes, that’s true. I have made a mistake. I should have said it is that
plant that bears round fruit, that when it opens....”
“Worse and worse,” interrupted the old dolphin. “What are you
talking of, anyway? That is the cotton plant. Marsovino, please
explain to this boy, who has read all the books in the world, what a
ray is.”
So Marsovino went on: “A ray is a fish, in shape like a large fan. It
has a very long tail, which it uses as a weapon.”
“To what class of fishes does it belong?” asked Pinocchio.
“It belongs to the same class as the lampreys, which look like
snakes, the torpedo,—”
“Be careful never to touch that fellow,” here interrupted Tursio.
“—the sawfish and the squaloids,—that is, the common shark and
the hammerhead.”
“The saw? The hammer?” observed Pinocchio. “If I find them, I must
keep them for my father. He is a carpenter, but so poor that he
seldom has money with which to buy tools.”
“Let us hope that you will never meet the saw, the terrible
hammerhead, or even the common shark,” said Tursio.
Pinocchio made no answer, but in his heart he kept thinking, “I am
very much afraid that the dolphins are teaching me, not I the
dolphins.”
Tursio then handed Pinocchio a small shell of very strange shape. It
looked like a helmet.
“Wear this, Pinocchio,” he said. “It will make a pretty cap for you.”
“It is very pretty. What is it?”
“It is a very rare shell.”
“But it is only one shell. Where is its mate?”
“It has none. It is a univalve. That means it has only one shell. The
tellines have two shells, and are therefore called bivalve. Another
kind looks like a box with a cover.”
“But does an animal live in there?”
“Of course. Every shell has its mollusk.”
“Mollusk?” repeated Pinocchio.
“Yes. The small animals that live in shells are called by that name.”
“They have a very soft body. By means of a member, called a foot,
they get such a strong hold on rocks that it is very hard to tear them
off.”
“Some mollusks have a strong golden-colored thread by which they
also hang to rocks. Why, people have even made cloth out of these
threads.”
Pinocchio cared little for all this explanation. He looked at himself in
the water, and was, after all, very much pleased with himself.
“This cap seems made for me,” he said. “Too bad I have no feather
for it.”
“Perhaps we shall find one on our journey,” laughed Tursio.
“Where will you get it? In the sea?”
“Yes, in the sea,” answered Tursio, in a tone which made the
impudent marionette almost believe him.
 CHAPTER IV
“Well, children, let us hasten. If we talk so much,
the sun will rise and find us here. Come, Pinocchio!
Jump on my back and let us start.”
There was no need for Tursio to repeat his
command. In the twinkling of an eye, Pinocchio was
riding on the dolphin’s back, holding on tightly to the
dorsal fin.

“Gallop and gallop, my pretty horse,

Swiftly over the boundless sea.
Straight through the water take thy course,
Till my dear father again I see.”

“Gallop and gallop, my pretty horse,

Gallop away under the sea.
Swim to the south, and swim to the north,
Till my dear father again I see.”

So sang Pinocchio gleefully.

Tursio and his swimming companions, with a few shakes of their
strong tails, were soon far away from shore. This is not to be
wondered at, for dolphins are known to be very swift. Very soon
Pinocchio saw nothing but sea and sky. Always holding on tightly to
Tursio’s fin, he looked to the right and to the left; but nothing could
he see of his dear father.
“Hold fast, Pinocchio,” suddenly cried Tursio.
“All right, Mr. Tursio,” replied Pinocchio, but he could say no more.
For suddenly, with a great jump, the dolphin was under water.
What a moment for our poor wooden hero!
“Now I understand it all,” he thought. “This dolphin wants to get me
into the sea that he may eat me at his leisure. Oh, poor me! I shall
never again see the light of day.”
But marvel of marvels! He suddenly awoke to the fact that, instead of
drowning, he was breathing easily. Not only that, but he could
actually talk!
“This is strange,” said he. “I have always thought that people would
drown in the water.”
“And it is true,” answered the dolphin, “that men usually drown in the
sea. But I have given you the power to live under water. You see,
then, you have become a real amphibian.”
“A real what? What am I now?”
“An amphibian. That is, you have the power to live both in the air and
in the water.”
“But are there such animals?”
“Why, of course, child. Frogs, for example, which belong to the
Batrachia family. In the water they breathe with branchiæ, or gills,
and in the air with lungs. Usually, however, the name is given only to
those mammals that live in the water and move only with great
difficulty on the earth. To this class belong the seals and the sea
lions.”
“Well, then, I shall never drown.”
“No; and you will have a wonderful journey under the sea. Just hold
on to me, and I will carry you. Do not be afraid.”
“Afraid? Of course not. But I don’t like the darkness very much.”
“That is too bad. But the darkness will not last very long. You know, I
promised that we should make our journey by the light of the sun.
Wait awhile.”
Through the water Tursio went like an arrow, followed by Marsovino
and the servant.
Pinocchio, to gain courage, shut his eyes. When he opened them
again, wonder of wonders! Very near to him a large sun was moving
back and forth. It looked as if it were alive.
“The sun at the bottom of the sea!” yelled Pinocchio, frightened
almost to death. “Do you want me to believe that? You must be a
wizard playing tricks on me.”
“I am not a wizard, Pinocchio, and the sun is not a trick. It is nothing
more nor less than a fish.”
“I never heard of such a thing.”
“And you have been in all the schools of the kingdom! Marsovino,
please explain to this boy what a sunfish is.”
“The sunfish is so called because of the bright light that comes from
its body. When several of these fish are together, the sea looks as if
it were full of little, shining suns.”
As usual, Pinocchio was silent. He was beginning to think that even
dolphins knew more than he did.
Stretching out his hand, he touched a small fish that was passing by.
Another surprise! As soon as he touched it, it began to swell and
swell, until it was as round as a ball. And from this ball, countless
points began to stick out.
“Suddenly, with a Great Jump, the Dolphin was under the
Water.”
“Oh!” yelled Pinocchio again. “What is it this time?”
“It is only a globefish, my marionette. It is harmless, if you don’t
touch it.”
“But why should it turn into a balloon?”
“It does that to protect itself,” answered Tursio. “It is possible for the
globefish to do that, because it can take in a large quantity of air.
With bristles ready, it can then meet the attacks of other fish, as each
point is as sharp as a needle.”
“I never knew that before,” exclaimed Pinocchio, forgetting his
previous boast.
Tursio and Marsovino looked at each other and laughed.
 CHAPTER V
The night passed without further adventure.
As soon as morning dawned, the four friends rose
to the surface. Our marionette was delighted to see
the sun again. The pure morning air, though,
reminded him that he was hungry. The day before,
if you remember, he had eaten very little.
“I should like something to eat,” he said in a weak
voice.
“Let us go to breakfast,” answered the dolphin. Gayly he dove into
the water, and led the party deep into the sea. After a short swim, he
stopped. But, unfortunately, the four friends found themselves in a
place where there were very few herring and salmon. These, you
know, are the dolphin’s favorite food.
The salmon is a fish that lives both in rivers and in seas. Like the
swallow, he looks for warm places in which to pass the winter. So, in
large numbers he migrates to the sea at that time of the year, and in
the spring he returns to the rivers.
“This morning our breakfast will be light,” observed Tursio,
swallowing three herring at once.
“I shall not eat anything. I don’t feel very well. Besides, salmon is the
only thing I can eat,” said Marsovino.
Tursio, wishing to please his pupil, started to swim toward two very
high rocks. They were so high that their tops stuck out of the water.
Very probably they were the base of an island in the middle of the
sea. But although he looked here, there, and everywhere, he could
find no salmon.
Globicephalous satisfied his hunger with three dozen herring and
half a bushel of smelts.
And Pinocchio? Pinocchio this time certainly did not suffer from lack
of food.
Tursio had shown him a large rock, attached to which were hundreds
of oysters. Some were of the size of a pinhead. Others were as large
as a boy’s cap, and these were two years old.
“Go and have your breakfast,” said Tursio.
“Must I eat those horrible-looking things?” asked Pinocchio.
“Open them and see what is inside,” was the reply.
“Pinocchio this Time certainly did not suffer from Lack of
Food.”
After Pinocchio had opened and eaten one, he no longer thought of
the looks of the oyster shells. He opened and ate so many, that it
was a wonder to Marsovino that so small a person could hold so
much.
Suddenly Pinocchio noticed numberless tiny, tiny white specks
coming out of some oysters. To him they looked like grains of sand.
But when he saw the specks moving and trying hard to attach
themselves to rocks, he could not help crying out, “O look at the live
sand, Tursio.”
“Who told you it is live sand?” asked Tursio. “Those are the newborn
oysters, looking for a place on which to spend their lives. Where
those small grains hang, there the oysters will live, grow, and die.”
“If no one gets them before that,” added Globicephalous.
“And are all those little dots oysters?”
“Yes. All of them. And many of them come from a single oyster, for
an oyster gives forth almost two millions of eggs at a time. These
little things have so many enemies, however, that very seldom do
more than ten of the millions grow old.”
“Two millions! Then I may eat all I want to,” continued Pinocchio,
unmercifully tearing away the poor oysters, young and old.
“Look, Pinocchio,” here called Tursio, pointing to a small fish, colored
with brilliant blues and reds. “That is the stickleback. You may have
heard that this fish makes a nest, as do birds. Also that the male, not
the female, takes care of the eggs.”
“Surely I have,” answered Pinocchio, seriously.
The stickleback seemed to be very much excited. He moved around
the nest he had made and watched it anxiously. The cause for this
was soon evident. A second stickleback made its appearance from
behind the rocks. At once the two engaged in a terrific struggle. They
bit each other, used their tails as weapons, and charged each other
viciously. During the battle they changed color—to a beautiful blue
mottled with silver.
Pinocchio was struck with wonder. “Look! Look! One is wounded....
He falls.... He dies!” he cried. “And look at the other. How quickly he
returns to the nest to guard the eggs!”
“But how is it,” here asked Marsovino, “that once I saw a stickleback
swallow one of his little ones?”
“If you had followed him, you would later have seen the small fish
come safely out of the large one’s mouth,” answered Tursio.
 “‘Look! Look! One is Wounded.’”

“But why did the large one swallow the small one?” asked Pinocchio.
“Because the little one probably wanted to run away from the nest. It
was too soon, the little one was too young to take care of himself; so
the father took the only means he had to save the youngster from an
enemy,” patiently explained Tursio.
Just then a small fish attracted the dolphin’s attention.
“Boys,” he said, “do you see that tiny fish? It is called the pilot fish. It
is the shark’s most faithful friend. Wherever goes the shark, there
goes the pilot fish.”
“Now, Pinocchio,” he continued after a pause, “I shall leave you with
Globicephalous. Marsovino and I are going to pay a visit to the
dolphin Beluga, who is a great friend of mine. He usually lives in the
polar seas, but on account of his health, he has come to warmer
waters. We shall return this evening, if all be well. Meet us near
those two mountains which are so close together that they form a
gorge. You may take a walk with Globicephalous, but be sure to be
at that spot to-night.”
“I am ashamed to be seen with a servant,” began Pinocchio.
“You are a fine fellow,” answered Tursio, with sarcasm. “Do you
know what you should do? Buy a cloak of ignorance and a throne of
stupidity, and proclaim yourself King of False Pride of the Old and
the New World!”
With this remark Tursio turned to his pupil, and the two swam away.