Int. J.

Human—Computer Studies (1997) 47, 429—452

Attribution accuracy when using anonymity in

group support systems
STEPHEN C. HAYNE
School of Management, Arizona State University West, 4701 W. Thunderbird Rd, Phoenix,
AZ 85069-7100, USA. email: hayne@asu.edu
RONALD E. RICE
School of Communication, Information & Library Studies, Rutgers University,
New Brunswick, NJ 08903-5067, USA. email: rrice@scils.rutgers.edu

This study explores the taken-for-granted assumption that ‘‘anonymous’’ comments

posted on a group support system (GSS) are socially as well as technically anonymous. It
analyses the accuracy of, and influences on, attributions of authors’ identities in seven
field groups with considerable work history after they used the system to enter technically
anonymous comments about salient topics during a brainstorming session. GSS particip-
ants made attributions about authors’ identities, but overall these attributions were
about 12% accurate (ranging from 1 to 29%). The expected predictors of accuracy (an
individual’s total communication with the group, network centrality, and length of
membership in the group) were inconsistent influences across the seven groups.
( 1997 Academic Press Limited

1. Introduction
Organizational work style and pace are changing, with greater reliance on the use of
teams, cross-functional and inter-organizational interactions involving groups of skilled
workers with diverse (Keltner, 1989). To maintain coordination and communication
amongst these teams, companies hold meetings. While such meetings are the most
critical communication forum for many organizations, frequently they are time-
consuming and unproductive experiences (Hackman & Kaplan, 1974; Mosvick &
Nelson, 1987; Rice & Shook, 1990; Pollard & Hayne, 1996).
A fairly new set of technologies designed to assist groups is called group support
systems (GSS) (Jessup & Valacich, 1993). GSS support a range of different group
activities such as brainstorming, categorization, decision-making and voting (Hackman
& Kaplan, 1974; DeSanctis & Gallupe, 1987; Dennis, George, Jessup, Nunamaker &
Vogel, 1988; Zigurs, Poole & DeSanctis, 1988; Nunamaker, Dennis, Valacich, Vogel &
George, 1991; Hiltz & Turoff, 1993).
One difference between a GSS meeting and a face-to-face meeting is that a GSS can
provide the option for group members to contribute comments or messages anonym-
ously, i.e. without revealing the author’s identity. Anonymity in GSS is interesting for
several reasons. First, prior research indicates that anonymity is an important variable
in studying, evaluating and managing GSS. Second, there is increasing interest in
examining anonymity as a determinant of a variety of group outcomes, in particular idea
generation. Third, anonymity is a complex construct. For example, is anonymity simply
429
1071-5819/97/090429#24$25.00/0/hc 970134 ( 1997 Academic Press Limited
430 S. C. HAYNE AND R. E. RICE

a feature of the technology that can be turned ‘‘on’’ or ‘‘off ’’ or are there more social
components?.
This paper describes results from surveys of seven groups’ responses concerning their
GSS use of anonymous brainstorming GSS for real organizational discussions. We have
conducted over 200 meetings using GSS of various types with the anonymity feature
engaged. It is our experience that during coffee breaks, participants repeatedly attempt to
decipher ‘‘who said what’’, and often attribute certain comments to other individuals
(frequently incorrectly), based on social cues in the textual comments. In this context, the
two basic questions considered here are (1) can we accept the assumption that using the
technical ‘‘anonymity’’ feature of a GSS will truly preserve ‘‘anonymity’’ and (2) what might
influence attributions of identity even under conditions of ‘‘anonymity’’?

2. Theoretical foundations: attribution, social cues, anonymity,

mediated communication and prior communication

2.1. THE GENERAL ARGUMENT

The general argument is fairly straight-forward. First, anonymity instantiated through
technical features may not guarantee social anonymity. Even in conditions of technical
anonymity, participants may attribute authorship to comments entered into a GSS.
Attribution, in general, is a common and apparently inherent human process. Attribu-
tions of authorship identity may arise from a variety of stimuli, and these attributions
may be more or less accurate. Influences on this attribution accuracy may include
sources of context such as group history and prior communication and individual
concern for anonymity. To the extent that participants do attribute authorship to
comments, they may then also use these attributions to evaluate the content of comments
that appear to be otherwise ‘‘anonymous’’. In turn then, GSS decision processes and
outcomes may not achieve the theorized benefits of anonymity, and may even suffer from
unidentified misattributions and social evaluations. Thus, this paper explores the con-
ceptual and empirical nature of anonymity in GSS by first reviewing the issues
of anonymity, attribution, social cues and mediated communication and prior
communication.

2.2. ANONYMITY
A major feature of some GSS is an ‘‘anonymity’’ condition: individuals can create
messages and vote or express preferences without identifying themselves and thus
without being able to identify the authors of contributions to the GSS discussion
(Nunamaker et al., 1991). Different ‘‘levels’’ of anonymity could be achieved in several
ways, i.e. by not introducing participants who have not previously met, not requiring or
posting author identity, or allowing participants to use ‘‘pen names’’ or aliases (Hiltz,
Turoff & Johnson, 1989; Jessup, Connolly & Galegher, 1990; Connolly, Jessup & Val-
acich, 1991; Lim & Benbasat, 1991). Much of the prior GSS research is equivocal on the
effects of anonymity. This ambiguity may be due to the wide variety of research
paradigms and settings employed, the variability of the technology and its use, and
different conceptualizations of anonymity.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 431

Efforts by many researchers (Hackman & Kaplan, 1974; Kerr & Bruun, 1981; Siegal,
Dubrovsky, Kiesler & McGuire, 1986; George, Easton, Nunamaker & Northcraft, 1990;
Jessup et al., 1990a; Pinsonneault & Kraemer, 1990; Connolly et al., 1991; Jessup &
Tansik, 1991; McLeod, 1992; Jessup & Valacich, 1993) have generally found an increase
in production and satisfaction when anonymous group brainstorming is used. Other
advantages of anonymous participation include decreased evaluation apprehension,
decreased member domination, decreased conformance pressure and decreased status
competition, which can lead to increased exploration of alternatives and surfacing of
assumptions.
However, potential disadvantages include social loafing, failure to listen, disinhibition,
deindividuation and poor socialization, which can lead to decreased effectiveness and
group dissatisfaction. Anonymity may not affect the ability to contribute, but it does
seem to affect what a participant is willing to say, and how it is said.
Valacich, Jessup, Dennis and Nunamaker (1992) provide a major review and concep-
tual framework which discusses the role of anonymity in GSS. They propose two types of
anonymity: process and content. Process anonymity promises a participant that others
will not know whether or not the particular participant is in fact participating. (Process
anonymity is not often available during same time, same place GSS meetings, because
participants can readily see who is there.) Content anonymity prevents others from
knowing who contributes particular messages. Content anonymity may be either local
(during the process) or global (extended even to later analysis of transcripts or logs). An
individual can take advantage of process anonymity, but inadvertently or intentionally
reveal themselves with their message content. Both of these forms of anonymity focus on
the technical aspect of anonymity.
Such emphasis on technical anonymity tends to ignore group members’ cognitive and
social processes, seeing individuals instead as minimediators of larger group processes.
This is inappropriate, because individuals bring with them a myriad of information that
can be brought to bear upon anonymously entered text to uncover cues and then
associate text with other group members.
Further, Licker (1992a, b) proposed that anonymity is a state of the individual,
experienced either as Identitylessness or Source Dissociation. Identitylessness is the
perception that one does not have an identifiable locus in a setting—i.e. others do not
know I am a participant or what my role is—whereas Source Dissociation is a feeling
that others cannot identify one as the source of specific messages—i.e. others do not
know I entered this comment. It is the source dissociation aspect of anonymity, a social
and not a technical condition, that we find interesting.
However, most GSS researchers adopt a definition of anonymity that treats anonym-
ity as a purely technical feature. This techno-centric definition hides the more social
nature of anonymity in communication settings; we feel anonymity is not just a switch
that can be turned on and off. First, people are seldom unidentified in every way. Second,
it is generally artificial to be an anonymous participant in most kinds of non-GSS
meetings (George et al., 1990), although we acknowledge that short-lived anonymous
situations such as expressing preferences through blind balloting do exist. Moreover,
Mennecke, Hoffer and Wynne (1992) suggest that even during GSS sessions, ‘‘extra-
meeting’’ influences, such as those discussions during coffee breaks, may serve as inputs
into future interactions and overcome anonymity.
432 S. C. HAYNE AND R. E. RICE

Consequently, we argue that there are two types of anonymity, technical and social.
¹echnical anonymity occurs when any meaningful identifying information about others
(or yourself ) is removed from any material exchanged. Thus, even content can be
technically anonymous if it cannot be linked (and therefore associated with prior
content) as coming from the same author, especially if that author is not specifically
identifiable. This may be enacted through the use of technical GSS features such as using
‘‘pen names’’, unsigned messages or removal of all account and participant identity in
any communication exchanged through the system (one can imagine voice signatures
and tones being removed from audio conferencing, etc.). Even asynchronous interaction
may reduce, to some extent, the identity of participants, by delaying responses and
ongoing contexts for identifying authors. Technical anonymity supposedly removes all
evaluations and responses based on social cues or personal identity and allows the
content alone to influence subsequent comments and decisions.
Social anonymity occurs when users actually perceive others (and perhaps even
oneself ) to be deindividuated or unidentifiable, perhaps because they do not perceive any
cues or contexts available to use in attributing the identities of others in relating multiple
messages. In GSS, technical anonymity seems to be a necessary, but not sufficient,
condition for social anonymity. As noted above, users may perceive others to be
identifiable, even under conditions of technical anonymity. Hence, technical anonymity
may be high, but social anonymity may be low.
Given the possibility of low social anonymity in technically anonymous GSS, we find
no prior research that empirically considers either the level or accuracy of attributions.
Group communication, membership, task, proximity, synchroneity, process, identity and
concern for anonymity are all issues that may influence levels of social anonymity, as
discussed below. (By the way, this is a general concern, not necessarily limited to
computer-mediated anonymity. For example, pieces of paper used by group members to
record their vote for later tabulation might be considered as the ‘‘technical’’ aspect of
voting anonymity. There are many situations where this ‘‘technically’’ anonymous voting
is highly charged with attributions and, thus, not ‘‘socially’’ anonymous.)

2.3. ATTRIBUTION THEORY

Attribution theory provides some conceptual basis for proposing that GSS users in
technically anonymous situations may not be completely socially anonymous, and leads
to some implications for possible effects of low social anonymity in ‘‘anonymous’’ GSS
decisions.
The central assumption of attribution theory is that humans are constantly engaged in
a search for the causes of events and that a understanding of such causes will permit the
person to function more adaptively (Heider, 1958; Kelley, 1967; Weiner, 1985). People
behave as ‘‘naive’’ scientists (Kelley, 1955) and try to reason in a more or less rational
way. Attribution theory examines the conclusions people reach about their own and
other people’s characteristics and the effects of these conclusions on future judgments
and choices. For example, one finding of this theory is the ‘‘fundamental attribution
error’’ by which we attribute others’ behavior largely to their personality factors, while
we attribute our own behavior largely to situational factors to which we respond (Jones
& Nisbett, 1972).
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 433

Researchers have attempted to go beyond the basic premise of attribution theory to

specify the information (or social cues) that humans use in inferring the causes of an
observed event (or statement), and the consequences of such inferences for the individual.
Kelley (1973) postulates that individuals employ certain informational cues—‘‘consen-
sus’’ (conforming behavior implies situational influences; unique behavior implies indi-
vidual characteristics as the influence), ‘‘distinctiveness’’ (highly distinctive behavior is
attributed to environmental conditions, while consistent behavior is attributed to indi-
vidual personality) and ‘‘consistency’’ (consistent individual behavior implies the causes
of that behavior are also stable)—to determine the causes to which behavior is at-
tributed. These three sources of information are obtained through observation of
behavior across people, situations and time, respectively (McArthur, 1972; Hewstone
& Jaspars, 1987; Hilton & Jaspars, 1987; McGill, 1989).
Iacobucci and McGill (1990) have meta-analysed published data to demonstrate that
several effects are particularly robust. Even over varying paradigms and procedures,
consensus and distinctiveness contribute to person attributions, distinctiveness contrib-
utes to stimulus attributions, and consistency contributes to circumstance attributions.
Thus, the above three major effects involve making attributions about people.
Hayes and Hesketh (1989) argue that attribution theory can be further enhanced
by considering the evidence that a wide range of cognitive biases and errors exist
(Kahneman & Tversky, 1973; Tversky & Kahneman, 1974; Hamil, Wilson & Nisbett,
1979; Henrion, 1980; Nisbett & Ross, 1980). That is, people make attributions based on
a variety of cognitive processes that may involve considerable bias and inaccurate
inference. Worse, Arkes (1981) has stated that merely explaining a bias to people and
then instructing them not to be influenced by it is not an effective strategy for removing
the bias. Therefore, more ‘‘active’’ strategies have been suggested (Koriat, Lichtenstein
& Fischoff, 1980; Nisbett & Ross, 1980; Arkes, 1981).
Why is this important to our current argument? As summarized in the prior section,
use of an anonymous GSS alters the attribution process. Anonymity in GSS comments
supposedly removes personal biases by removing the identity of comments’ authors from the
content of those comments during the process of decision making. However, (1) people may
continue to attribute statements made in the GSS even in anonymous settings (as people
seem disposed to constantly make attributions, frequently, about others), (2) these
attributions may not be accurate and may be highly biased and so (3) individual
decisions, contributions, discussions and judgments may be affected by mis-attributions
in conditions of anonymous input. Note we are concerned with individual-level issues,
but in the context of group work.

2.4. SOCIAL CUES AND MEDIATED COMMUNICATION

All cues that individuals provide, or other individuals perceive, during an interaction
may both stimulate, and be used as evidence for, one’s attributions. Obviously, we pay
close attention to the content of what a person is saying or doing. However, other specific
cues, such as attractiveness, dress or speech may be especially salient due to context,
expectations or attention. These salient cues are then the basis for forming impressions
and evaluating others (Taylor & Fiske, 1978). Salient stimuli encourage the use of simple
inferential rules, based on limited information about others (Blank, 1991). The more cues
434 S. C. HAYNE AND R. E. RICE

we successfully decode, the more accurate our attribution. Thus, available or perceived
cues can affect the social anonymity of participants in technically anonymous GSS,
evaluations of specific participants and their contributions and perhaps general decision
outcomes. But on what could these attributions be based, if people are using not only
a technically anonymous, but also a text-based, GSS?
Two somewhat similar theories focus on the interaction between social cues and use of
new organizational media: social presence (Short, Williams & Christie, 1976; Rice, 1987)
and media richness (Daft & Lengel, 1984, 1986). Both emphasize how communication
media differ in the extent to which (1) they can overcome various communication
constraints of time, location, permanence, distribution and distance, (2) transmit the
social, symbolic and nonverbal cues of human communication and (3) allow interpreta-
tion of equivocal information. Rice (1987) also discusses these and other attributes of
media.
According to these two theories, computer-mediated communication generally trans-
mits fewer social and nonverbal cues. This is because text-based media such as electronic
mail and group support systems have limited technical bandwidth compared to face-to-
face communication. In many cases, especially for tasks that require high social presence
or information richness, this cue reduction is associated with negative effects such as
disinhibition (flaming), misunderstandings, lack of consensus and feelings of imperson-
ality (Kiesler, Siegel & McGuire, 1984; Rice, 1984). However, the reduction of such cues
may also lessen status differences and domination based on appearance, oral skills, etc.,
that are common in face-to-face group discussions (Krauss, Apple, Morency, Wenzel
& Winton, 1981). In computer-mediated communication, this reduction of cues is often
reflected in more equal participation, more reciprocal communication and less sequential
or linear comments (Rice, 1984; Dennis et al., 1988; Hiltz & Turoff, 1993).
This proposed ability of new media to reduce the communication of social cues is one
reason for the development and testing of anonymity features in group support systems.
For certain group activities, such as brainstorming or airing disagreements, it is the-
oretically advantageous to allow participants to offer comments simultaneously and
anonymously, to further decrease possible negative effects of social, status and authority
cues (Jessup et al., 1990b). Thus, a participant’s internal and public responses to the
comments are disassociated from responses to specific others, as well as from attributions
of others’ motivations behind the comments. As a result, in medium to large groups,
electronic brainstorming and similar techniques (such as the nominal group—Moore,
1987) seem to stimulate more diverse, more numerous and more honest comments.
Especially for certain politically charged or difficult group issues, this anonymity of
authorship can be a crucial component of group success (that the difficulty of making
accurate attributions under these conditions has been presumed to imply that people do
not make attributions under these conditions).
In spite of the research summarized earlier that presume CMC and GSS filter out
most such social cues, some researchers argue that text-based media may still convey
substantial emotional and personal content (see reviews in Kiesler et al., 1984; Culnan
& Markus, 1987; Rice & Love, 1987). Further, CMC users may develop richer relations
and ways of increasing social cues in their electronic mail messages over time (Sproull
& Kiesler, 1986; Walther, 1992). In attempts to improve one’s relational understanding of
other communicators, users may ‘‘develop distinctive impressions of other interactants
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 435

by decoding text-based cues and derive psychological-level knowledge about other

actors from CMC interaction’’ (Walther, 1992, p. 67). Walther argues (1992), and finds
(1993) that impression formation through CMC may increase over time to approach the
initially higher levels of face-to-face impression formation, and that this will be more
likely for ongoing CMC sessions and for smaller groups. With groups that already have
considerable experience communicating through a variety of channels with other group
members, users may be able even more easily to interpret and decode textual entries, and
associate relational impressions to stylistic characteristics and stated positions available
in others’ comments, even though the comments are technically anonymous. Lea and
Spears (1991) note that paralanguage (the use of typographical and punctuation symbols
to convey tone and emotion) is one means by which social information is communicated
in CMC. The meaning of such paralinguistic marks is dependent on the group or
individual communication context. They also found that deindividuated subjects for
whom group identity had been made salient would evaluate users of paralanguage more
positively than when group salience was low, supporting their argument that social
context could be highly influential in mediated communication.
Even if nonverbal cues (visuals, tone, inflection, eye-gaze, appearance, etc.) are ‘‘filtered
out’’, it may not be the case that the majority of evaluative content and social cues is, in
fact, carried in the nonverbal portion. Krauss et al. (1981) argue that the high percentage
of communication of emotion carried in nonverbal modes commonly cited are derived
from atypical experiments. Their own study compared the extent to which people
evaluated, and correctly evaluated, the emotional affect of a presenter, through several
different channels in two studies: full-channel (audiovisual), verbal written (transcripts),
vocal (content-filtered speech) and visible (silent video). In both studies, verbal content
was the best predictor of evaluative judgments. ‘‘Overall, no support was found for the
widespread assumption that nonverbal channels (vocal or visible information) form the
primary basis for the communication of affect’’ (p. 312). They also noted several other
studies that found similar results, but the authors did agree with the proposition that
nonverbal cues provide the context for accurate interpretation of verbal content.
What all these propositions and results mean is that readers of text-only com-
ments in a technically anonymous GSS might perceive enough cues to make attri-
butions, and that those attributions might be at least somewhat accurate. For example,
an author might always use an accepted abbreviation or jargon (consensus), might use
a unique grammatical style (distinctiveness), and might repeatedly state a position that
a reader knows the author holds (consistency). In this case, accurate attribution may be
possible.

2.5. PRIOR COMMUNICATION

We suspect that participants might, in fact, be able to identify the authors of such
technically anonymous comments, especially in groups that have a considerable history
of interaction. Knowledge of the other participants’ identities brings to bear prior
knowledge of relevant cues such as status, expertise, authority, conflict or social role,
thereby reducing the benefits of anonymity. This could give some participants an
advantage, biasing the evaluation of comments or even allowing some individuals to
consider later sanctions for others’ critical contributions.
436 S. C. HAYNE AND R. E. RICE

In line with Walther’s (1992) propositions about impression development in computer-

mediated interactions, and Rice’s (1987, 1993) discussion of the importance of network
structure for understanding the use of computer-mediated systems, individuals with
more prior communication with other members of the group should be better able to
identify otherwise technically anonymous comments from other members. This is be-
cause they would be more exposed to others’ communication styles and cues over time—
in Walther’s terms, they would have formed more detailed impressions of the others—
thereby increasing the likelihood and accuracy of attributions. The simplest indicator of
overall prior exposure to group members’ communication might be the length of time
since joining the group (membership history).
However, communication structure may be conceptualized in several more rigorous
ways. Three possible approaches are justified through network analysis theory, each
increasingly more focused on specific network relationships: amount of communication
with one’s group, centrality in that group’s prior communication network, and the extent
of prior communication with specific others. Rice (1993) discusses the theoretical motiva-
tions for different network mechanisms of social influence in groups.
The first is simply total amount of prior communication with one’s group. Greater the
overall frequency of communication with the group, the more likely one has been
exposed to other members’ communication styles, cues and positions on salient topics.
Overall communication would be important both to be able to identify comments as
belonging to a particular author, as well as to know which other group members do not
share those communication styles, positions or cues.
The second has to do with one’s location within the group’s communication struc-
ture. Freeman (1979) argues that the extent to which each group member lies on
the communication paths of all other pairs in the group—that is, one’s betweenness
centrality in the group—affects that individual’s control of the flow of information from
and about the other group members. This concept is more specific than the total
communication frequency, as it takes into account the patterns of communication
relations among all group members. The more central one is, the more familiar one
should be with others’ opinions and communication cues and, therefore, more likely to
make attributions and be more accurate in those attributions. However, this measure
also takes into account the influence of relations with others whom one might never
directly communicate, so it may not be a strong influence on accuracy of attributions of
others’ comments.
Finally, relationships between communication and attributions may be most clearly
articulated between specific pairs of group members. That is, to the extent that specific
pairs of group members communicate more frequently with each other, they should
be more familiar with each others’ communication styles and issue positions. This
approach takes into account the overall effect of each pairwise relationship sim-
ultaneously, but it is only based on direct relationships between specific others. This
approach also takes into account the notion that some attributions may be made only to
those others who are (perceived to be) well known. Thus, length of group membership,
total group communication frequency, group centrality and pairwise communication
strength all should be related to identification of the otherwise technically anonymous
GSS comments, with each implying a slightly different mechanism of impression devel-
opment and attribution.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 437

3. Hypotheses
Thus, the two general research questions for this study are: (1) can we accept the
assumption that using the technical ‘‘anonymity’’ feature of a GSS will truly preserve
‘‘anonymity’’ and (2) what might influence attributions of identity even under conditions of
‘‘anonymity’’? Three sets of directional hypotheses follow from the prior discussions:
H1. Group members can accurately identify the source of technically anonymous
comments.
H2a. Longer membership in a group is positively associated with one’s mean percent
correct identification of authorship of technically anonymous comments.
H2b. The total amount of one’s prior communication within a group is positively
associated with one’s mean percent correct identification of authorship of technically
anonymous comments.
H2c. One’s position in the prior communication network, as measured by be-
tweenness centrality, is positively associated with one’s mean percent correct identifica-
tion of authorship of technically anonymous comments.
H3. The prior pairwise communication among group members is positively asso-
ciated with the percentage correct identification of authorship of technically anonymous
comments between those pairs of group members.
Hypothesis H1 is conceptually independent from the other hypotheses. That is, if
people cannot accurately make attributions, the attribution process itself might still be
predictable.

4. Methods
We have engaged in an exploratory study involving a ‘‘real-time’’ survey of GSS use. The
groups studied herein are not highly controlled laboratory, student or simulated groups
(as critiqued by McGrath & Hollingshead, 1993), but are field groups attempting to
resolve actual business problems. Natural groups are often part of a larger organization,
are likely to have worked together previously and may be working on several projects
simultaneously. Groups do not, of course, have to represent projects or teams—they may
be any set of people with regular interaction and a common goal.

4.1. PARTICIPANTS
The seven organizational units involved here are all operations divisions of mid-
sized firms—i.e. subjects were staff, managers and executives, not faculty or stu-
dents. All groups consisted of people who had worked together previously—some
for up to 15 yrs—and contained members who worked together daily. All were
meeting and using a GSS to resolve real, salient and ongoing business problems. These
groups were not standing ‘‘committees’’ or ‘‘task forces’’ assigned to assess a particular
project; the groups were made up of almost all the members of their respective business
unit.
The first four groups (‘‘PE’’ 1—4) were departments belonging to the physical educa-
tion (PE) unit in a large university meeting to deal with budget cutbacks. Managers of the
438 S. C. HAYNE AND R. E. RICE

unit assigned members of the departments somewhat randomly (primarily on the

basis of availability) across the four groups (i.e. not assigning each department
completely to its own group), with sizes of 19, 23, 14 and 13. The remaining three
groups all came from organizations with approximately 300 employees each. The
‘‘Consult’’ group was composed of 16 senior/junior partners and staff in a con-
sulting engineering firm, who communicated as non-hierarchical peers. This en-
terprise was in the throes of the 1993 Canadian recession, trying to pick possible
strategies that might halt the downsizing and lead to growth. The ‘‘Pipe’’ group
was made up of 27 middle managers and staff from a pipeline corporation whose
department had just been administratively split into two. Several members of the original
department had been let go in the reorganization. Everyone felt the pressure from
wrenched social relationships, and had developed a considerable prior informal com-
munication network to discuss these issues and provide support. They were engaged in
a business process reengineering to try to further reduce costs yet increase market share.
The ‘‘Oil’’ group consisted of 17 senior executives, middle managers and staff from
a ‘‘junior’’ oil and gas company. The morale of the company was quite low from recent
announcements of increased work requirements (this in conjunction with recent industry
layoffs).
Thus, the general task for each group was highly salient—to determine methods for
improvement and redesign in response to demands to improve quality in the face of
imminent fiscal difficulties. The managers and executives wanted to gather comments
and ideas on current and new strategies, and while these groups are not traditional
small-sized teams or meeting groups, they are real, motivated and have prior interaction
histories.

4.2. GROUP SUPPORT SYSTEM

The GSS used in this research was Group Systems (Nunamaker et al., 1991), which is
often classified as a typical Level-I GSS (DeSanctis & Gallupe, 1987). The groups were
engaged in a very typical GSS session consisting of brainstorming (Idea Organizer),
rating and ranking (Voting) and implementation planning (Topic Commenter). The tool
used to gather anonymous comments was Topic Commenter which supports idea
sharing on user-selectable topics. Members could sit where they wished in the computer
center, and could not easily see a neighbor’s computer screen. The same person, trained
in GSS facilitation, facilitated each of the seven meetings.
Each of the four PE groups met for about 4 h in the same GSS room, on separate
days. The last three groups held day-long sessions (7 h) in the GSS facility. For the
four PE groups, managers had identified 13 primary issue areas during an earlier
session for the GSS groups to discuss. The other groups generated their own issue areas
in the previous 4 h. Each group’s topic comment session was one of several phases, and
lasted for 1 h and 15 min. This consisted of members’ entering comments under each
issue and reading other members’ comments, in whatever order they wished. There
were many ‘‘conversations’’ within the issue areas, where one member’s comment
would respond to another member’s comment. There was no formal indication via the
GSS features as to who entered which comment—i.e. the comments were ‘‘technically
anonymous’’.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 439

4.3. DATA
Three forms of data were collected. The first was a communication network roster
administered before each group began its session. This roster simply listed the names of
participants, and asked each participant to indicate the extent to which they communic-
ated with each other listed person about any topic, using any communication channel
(from 0"never, 1"once every few months, to 7"several times a day).
The second form of data were a sample of the actual comments participants entered.
Comments were restricted to those longer than 65 characters in length in order to reduce
comments like ‘‘Me, too!’’. Our intention was to provide comments with at least
a minimal amount of social cues.
Fifteen minutes before the end of the GSS session, one of the researchers used a simple
macro to select up to 50 comments from the issue areas with the most comments entered
into them at that time (potentially different across sessions), in order to maximize the
chance that the comments had been seen by everyone in the group. The comments were
randomly chosen in proportion to the numbers of total comments in each issue window,
yet the original order of the comments from Topic Commenter was preserved. Some of
the issues received more attention than did the others. We assumed that more comments
in an issue area might mean that more people had browsed the comments. Note that we
have no basis to assert that any particular comment was read by any particular member.
These selected comments were copied into a word-processing file and included in the
survey to create and print a ‘‘comment’’ questionnaire that was distributed immediately
at the end of the phase.
The comment questionnaire was the third form of data. First it asked several indi-
vidual-level questions about group membership and attitudes toward anonymity. These
were followed by a list of each of the participants with an assigned sequential letter, and
a listing of the selected comments for that group. Since Topic Commenter is designed to
maintain and guarantee anonymity, it could not record who actually entered each
comment. So, the participants were first asked to indicate which comments were theirs (if
any), by entering their own letter alongside them. Note that we have no basis to assert
that each such self-identification was correct. However, there are two bounds to this
assertion: (1) very few comments were claimed by more than one participant and
(2) very few people did not claim any comments (see Table 1, from 1 to 9). Then the

TABLE 1
Comment totals

Group Made by Selected Claimed by more Not claimed Mean percent

group than one person by anyone attributed

PE1 228 52 1 1 21
PE2 352 26 3 2 21
PE3 200 39 3 1 9
PE4 156 50 1 1 23
Consult 134 35 1 3 70
Pipe 159 40 1 9 34
Oil 122 39 1 1 17
440 S. C. HAYNE AND R. E. RICE

participants were asked to make their best guess as to which other participants contrib-
uted the listed comments, by entering the other participant’s letter alongside the com-
ment. The instructions indicated it was okay not to enter a person’s letter if one was
unsure. This questionnaire was completed in less than 20 min and returned before the
participants engaged in any other activities.

4.4. MEASURES
The data were used to create the following measures for each group: prior communica-
tion, accuracy of attribution and individual-level survey variables. Few other GSS
studies involving anonymity have developed any related measures associated with
anonymity, other than the ‘‘treatment’’ of anonymity itself. Therefore, most of our
measures of attribution accuracy and attitudes about anonymity were developed for this
study. The communication network measures, however, are typical operationalizations
of sociometric interaction (Rice, 1993).
Prior communication ( from pre-session communication roster).
1. The sum of the communication frequency for each member. This is the row total of
the square, asymmetric scalar matrix of communication frequency among all members of
a group, constructed from the communication roster on the pre-session survey. This is
simply a ‘‘who-to-whom’’ matrix indicating the level of interaction (from 0 to 7) person
i reported having with person j. As j might report a different level of interaction with i,
such matrices are inherently asymmetric (cell i, j of the matrix is not necessarily equal to j,
i). The row total ignores the diagonal (which is 0), and represents the total extent of each
individual’s reported communication with other members of the group.† The mean
communication frequency for each individual, across all the other group members (to
control for the different sizes of each group), was also calculated for descriptive purposes.
2. The standardized betweenness centrality of each person in that person’s group. As
derived by Freeman (1979), betweenness centrality measures the extent to which each
group member lies on the shortest communication paths between all other pairs in the
group. A communication path is the set of interactions needed to connect person i with
person j; the shortest such path is called a ‘‘geodesic’’. The resulting betweenness

† Self-reported measures of a wide range of behaviors—including responses to network questions about the
number and intensity of linkages with other individuals—often disagree with comparable measures of observed
(possibly ‘‘actual’’) behavior, including comparable computer-monitored measures, even when respondents are
surveyed on-line within minutes of actual system use (Bernard, Killworth & Sailer, 1982). Thus, we would be
reluctant to claim that this self-reported frequency measure necessarily accurately reflects prior communication
as opposed to perceived prior communication. It cannot be taken as a necessarily valid indicator of prior
exposure to impression development.
Thus, the column totals of the communication network matrix may be more reliable measures of exposure to
cues from one’s group members than the row totals. That is because one’s reported communication frequencies
with each other group member are not independent; each is systematically biased by the respondent’s recall or
other reporting errors (Bernard et al., 1982). But the total column frequency measures how frequently each
other person indicated he communicated with the respondent. This should average out individual biases,
increasing the reliability of this measure of exposure through communication frequency. To the extent that row
totals and column totals are highly correlated (see footnote ‡), however, not only would any effects be similar,
but our trust in the reliability of the respondent’s reported communication frequency, and the validity of both
measures, would increase.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 441

centrality is standardized (according to Freeman’s operationalization) to remove the

effect of differing group sizes. First the network matrices were symmetrized by putting
the average of the cell and cell into both cells, then dichotomizing those values at
i,j j,i
3 (communicating several times a month). That is, the average extent of communication
between two group members had to be at least several times a month to be considered
potentially influential. From this symmetric binary matrix, betweenness centrality for
each individual in the group was calculated as the average of all the geodesics (the
shortest path between each pair of group members) on which each individual was
located.

Accuracy of attributions ( from post-session survey).

3. The average normalized attribution accuracy score (row mean, ignoring the diag-
onal) for each member, based on the square, asymmetric matrix of normalized attribution
accuracy by each member of each other member’s comments.
Each cell in each group’s accuracy matrix is filled with the result from the equation

cell "(sum[(a )!(k )])/n

i,j ij j
where i, j is the number of participants, ie j, a the attribution by i that comment k was
ij
made by j (1 if made, 0 otherwise), k the comment k made by j (0 if true, 1 otherwise) and
j
n"k from 0 to n, the total number of attributions made by i. For each comment listed on
the comment questionnaire, if person i makes an attribution of authorship to person
j then that cell i, j gets a 1 if correct, 0 if incorrect. If i did not make an attribution for
a comment, the value is set to missing, so that nonattributions do not count either as
accurate or not accurate. That is, this formula does not presume that i should make
attributions to every comment; it only considers when i does make attributions. This is
done for each, and summed over all, of i’s attributions to comments made by j. This
number is then divided by the number of attributions i made. There was no correction for
random (as opposed to well-intentioned) guessing. Thus, the attribution accuracy score
for each i, j relation theoretically ranges from 0 to 100%.
4. Number of attributions made by each participant.
5. Percent attribution accuracy, the number of correct attributions made, divided by
the total number of attributions made, for each participant. This is set as missing for
those who made no attributions.

Individual-level variables ( from post-session survey).

6. Duration of group membership (how many months they had been a member of this
particular group) used in many studies mentioned in Jessup and Valacich (1993).
7. Perception of anonymity (how anonymous they felt they actually were during the
Topic Commenter part of the meeting, from 1"not at all to 7"completely) (Jessup
& Tansik, 1991).
8. Importance of anonymity (how important it was to them for the Topic Commenter
part of this meeting to be anonymous, from 1"unimportant to 7"crucial) (Jessup
& Tansik, 1991).
9. Presence of cues as indicated by responses to an open-ended question asking
whether there were any events during the meeting that made it easy for the respondent to
identify others’ comments.
442 S. C. HAYNE AND R. E. RICE

TABLE 2
Duration in group, anonymity attitudes, attribution and accuracy

Months of Perceived Importance of Number of % correct

membership anonymity anonymity attributions attributions

Group N Mean S.D. Mean S.D. Mean S.D. Mean S.D. Mean S.D.

PE1 14 56.6 54.9 5.9 1.1 4.8 2.2 12.3 14.1 0.08 0.14
PE2 19 99.7 66.2 4.8 1.4 5.5 1.3 5.5 6.2 0.11 0.24
PE3 14 39.3 34.4 5.0 1.2 5.4 1.4 3.6 5.2 0.01 0.03
PE4 13 40.0 32.0 5.3 1.8 5.7 1.8 11.6 20.6 0.02 0.05
Consult 14 55.7 44.5 5.4 1.5 4.1 1.7 24.6 11.8 0.29 0.16
Pipe 26 51.6 46.7 5.8 1.1 6.0 1.1 10.2 9.0 0.23 0.28
Oil 17 38.9 32.9 5.3 1.5 5.2 1.6 6.7 9.1 0.01 0.03
Overall 117 55.9* 50.1 5.4 1.4 5.3 1.6 10.5* 12.6 0.12* 0.21

Note: ANOVA across 7 groups. *"p(0.05, S.D."Standard Deviation.

TABLE 3
Network measures

Sum individual Network Mean

communication betweenness individual
frequency centrality network strength

Group N Mean S.D. Mean S.D. Mean S.D.

PE1 15 56.0 11.2 0.03 0.02 4.00 0.80

PE2 21 60.3 25.6 0.03 0.04 3.00 1.28
PE3 14 40.9 14.4 0.04 0.05 3.15 1.11
PE4 13 42.7 17.5 0.04 0.04 3.56 1.46
Consult 16 46.8 12.5 0.04 0.03 2.92 1.00
Pipe 26 110.7 27.9 0.01 0.01 4.43 1.12
Oil 17 50.8 16.9 0.03 0.05 3.18 1.06
Overall 117 63.1 32.4 0.03* 0.04 3.52* 1.12

Note: ANOVA across 7 groups. *"p(0.05.

In the first PE group, two people refused to complete the comment questionnaire; they
were very concerned that their comments might find a way back to their managers
(implying anonymity was very important to them!). These were dropped from both the
network and the accuracy matrices before analyses were conducted. Tables 2 and 3
summarize the descriptive statistics for the individual-level variables, and the network
and accuracy variables, respectively.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 443

5. Results
The percent of all comments attributed by the group ranged from a low of 9% for PE3 to
a high of 70% for the Consult group, with a similar range in mean number of attributions
by individuals from 3.6 in PE3 to 24.6 in Consult. That is, groups ranged widely in the
percentage and frequency of attributions. People felt they were mostly anonymous
(M"5.4), and about equally concerned that these comments should be anonymous in
these meetings (M"5.3). There was no overall statistically significant difference among
the groups on these two perceptions. On average, members belonged to these groups for
almost 5 yr (M"55.9 months), with only PE2 having a significantly higher mean.
Very few people provided any indication (other than the content of the comments) that
there were events that could make it easy to identify comments: most either left the
question blank or wrote ‘‘no’’ or ‘‘none’’. Five people mentioned that it was occasionally
possible to see others’ screens. The following remarks came from the Consult particip-
ants: ‘‘certain comments have been made in the office and appear here’’; ‘‘fluency in
typing’’; ‘‘through joking discussion or the last person was still typing’’; ‘‘the person uses
many words not normally used by others’’; ‘‘recognize some people’s English or lack
thereof ’’ and one person in the Consult group apparently added his initials to his GSS
(not on the survey) comments, which two other respondents from the group mentioned.
As noted below, this group had the second-highest level of perceived anonymity,
a medium level of anonymity importance and was by far the most accurate group! The
remaining comments came from the PE groups (each of which, remember, included
people from the other three PE areas): ‘‘how people express themselves’’; ‘‘strong opinion
that serves their area’’; ‘‘people naming their areas’’ and ‘‘just the type of answers that
were given led me to assume I knew the person who answered’’. Each of these comments
about cues, except for being able to see another’s screen and seeing who’s left typing,
presumes prior knowledge of the other person from communication cues or associated
issue positions.
Notice that we are not able to determine what percentage would be below, equal to or
above chance guessing. Due to the nature of this commercial anonymous GSS, we
cannot capture the number of comments each participant contributed, and it would be
unwise to assume that the proportion of comments claimed by a group member would
hold for the totals in the session. Thus, we are not assessing guesses or errors of omission,
only the accuracy of attributions that are made.

5.1. HYPOTHESIS TEST: INDIVIDUAL INFLUENCES

Mean accuracy ranged from 10% for PE3 and Oil group to 29% for the Consult group,
with an overall average of 12% accuracy (with the Consult and Pipe groups significantly
more accurate than the other groups). When our respondents indicated attempts at
authorship attribution, they were wrong almost 90% of the time. Hypothesis 1 must be
rejected (note, however, that they were right more than 10% of the time!). Furthermore,
there is no correlation between a group’s accuracy and the group size or the number of
comments selected for the survey.
The duration of group membership was significantly correlated with attribution
accuracy for three groups (see Table 4—all p(0.05), PE1 (r"0.85), PE4 (r"0.55) and
444 S. C. HAYNE AND R. E. RICE

TABLE 4
»ariable and matrix correlations with accuracy

Group (H2a) (H2b) (H2c) (H3)

Network Network Months Network
strength centrality membership matrix

PE1 0.14 0.43* 0.85**** 0.15**

PE2 0.03 !0.07 !0.37 !0.02
PE3 0.08 !0.09 0.27 0.01
PE4 !0.04 0.08 0.55** !0.40**
Consult 0.12 !0.13 !0.44 0.08
Pipe !0.36* !0.28 0.58**** 0.14
Oil 0.25 0.32 0.17 0.15**
Overall 0.16* !0.11 0.15 —

* p(0.1; ** p(0.05; *** p(0.01; **** p(0.001.

Pipe (r"0.58). Two other (nonsignificant) correlations were positive and two were
negative (PE2 and Consult). Hypothesis 2a received moderate support.
With respect to Hypothesis 2b (total prior communication and attribution accuracy),
only one group showed a significant correlation. The Pipe group members’ total
communication was negatively correlated r"!0.36 (p(0.1) with their percentage
accuracy. Note that this means that the more intensely communicative, the more likely
the group members were to make attributions that were not accurate! As noted in the
methods section, the members of this group may have developed exceptionally close
communication relations due to the emotionally gripping nature of their situation.
Indeed, from Table 3, their mean individual network strength (4.43) was significantly
higher than all other groups’ means except for PE1 (R2"0.21, F "5.2, p(0.001).
6,121
Other than the Pipe group, only the Oil group even approached a significant positive
correlation of network strength with attribution accuracy (r"0.25). Merging the seven
group data sets, the overall correlation was r"0.16 (p(0.1). Hypothesis 2b is not
supported.†
Hypothesis 2c predicted that location in the group’s communication network struc-
ture (centrality) would influence the accuracy of comment attribution. Network central-
ity was marginally significant and positively correlated only in PE1 (r"0.43, p(0.1).
The overall correlation was r"!0.11 and not significant. Hypothesis 2c is not
supported.
In addition, within any specific group, neither perceived anonymity nor importance of
anonymity was correlated with accuracy of attributions.

† The communication row (person i’s rated level of interaction with all j’s) and column (the level of
interaction person j receives from all i’s) totals are significantly correlated for all groups (r"0.56—0.88,
p(0.001) [the shared variance is in the same range reported in the prior studies of self-reported network
communication accuracy (see fn. †)] except for the Consult group (r"0.31, n.s.). So we would not expect much
difference in their effects. The row and column means, of course, equal each other, for each particular group.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 445

5.2. MULTIPLE REGRESSIONS

Because there were some significant correlations among the independent variables,
a final summary stepwise multiple regression was run for each group, including network
row total, network column total, network centrality, months in the group, perceived
anonymity and importance of anonymity (see Table 5).
For two groups (PE1 and PE4), number of months in the group was the only
significant positive predictor of accuracy. In two other groups (PE2 and Consult),
months approached significance as a predictor but was negatively associated with
accuracy. For the Pipe group, months was a significant positive predictor, and network
centrality was a significant negative predictor.

5.3. HYPOTHESIS TEST: NETWORK LEVEL INFLUENCES

The test for Hypothesis 3 compares each group’s comment accuracy matrix to the
group’s communication network matrix, based upon pairwise patterns of similarity
between i, j in each matrix. Traditional statistics are inappropriate for testing the
significance of network matrices, because the rows and columns of a who-to-whom
communication matrix are likely not independent. The generally accepted approach is to
first reorganize each matrix into a vector, and then compute the correlation between the
two vectors, to produce a familiar measure of the strength of association, the Pearson r.
Then the quadratic assignment procedure (QAP) (Hubert & Schultz, 1976) is used to
determine the significance of the association between the two matrices, by means of
a nonparametric gamma association significance test (derived from a distribution of
possible associations between one matrix and all permutations of the other matrix).
However, because there are asymmetric missing values in the accuracy matrix (when
j made a comment but i did not make an attribution), even this significance test will not

TABLE 5
Regressions of accuracy on network, duration and anonym-
ity variables

Equation parameters

Group Standardized betas Adj. R2 F-ratio

PE1 Months 0.85**** 0.69 25.0****

PE2 Months !0.37 n.s. 0.08 2.5 n.s.
PE3 — — — n.s.
PE4 Months 0.59* 0.21 3.5*
Consult Months !0.44 n.s. 0.13 2.9 n.s.
Pipe Months 0.65**** 0.46 10.9****
Net Cent !0.42***
Oil — — — n.s.

Note: Months"number of months as member in group.

Net cent"centrality in network of prior communication.
* p(0.1; ** p(0.05; *** p(0.01; **** p(0.001.
446 S. C. HAYNE AND R. E. RICE

work. As an informal comparison, correlations for matrices that had those values in all
the cells were significant at p(0.05 for values of r"0.15 and greater.
Table 4 shows that only PE4 exhibited a clearly significant association (r"!0.40)
between (less) pairwise prior communication and accuracy of attributions. Groups PE1,
Pipe and Oil showed a slightly (positive) association between more pairwise prior
communication and accuracy of attributions (r"0.15). These results reject Hypothesis 3.

6. Conclusions
Table 6 summarizes the results. Based on our preliminary attempt to see whether the
technical anonymity (the mere suppressing of information about the identity of comment
authors) can be overcome by users’ attempts to identify authors, and whether this is
influenced by different forms of prior communication among groups with history, we
may make several conclusions.
This study provides evidence that users of technically anonymous GSS may not feel
other users are completely socially anonymous, and may make attributions about the
identity of authors of anonymous text-based comments. However, even in this condition
not everyone made attributions, and not everyone made attributions about every
comment. On their own, people may not always be concerned about figuring out who
wrote which comment. We also ignore the issue of what, exactly, might be naturally
occurring attribution stimuli in anonymous GSS. We did not attempt to measure or
manipulate those stimuli. Such stimuli are either latent variables or perhaps something
that cannot be controlled for. That is, how would one be able to find out about silent
attributions without asking people if they made them?
However, when participants in these seven naturally occurring groups did attempt to
attribute authorship to technically anonymous comments, they were incorrect about
90% of the time. It is worth noting that these groups were working on a difficult problem
that was quite salient to them. Some of the comments selected were lengthy and complex
while others were quite short and simple. It may be that certain ‘‘path-breaking’’
comments—those that establish agendas for subsequent comments—might stand out
much better and be more accurately attributed. Further studies of anonymity might
examine attributions of ‘‘path-breaking’’ statements. Perhaps most of the ‘‘other’’ com-
ments are merely noise that the group filters out.

TABLE 6
Summary of results

Do individuals attempt attribution? Yes

Is an individual’s prior communication history associated with attribution accuracy? No

Are group members accurate in their attributions? No
Is duration of group membership a factor? Yes
Is total group communication a factor? No
Is the group’s network structure a factor? No
Do individual’s pairwise communication influence accuracy? No
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 447

Second, it may be that technical anonymity is fairly ‘‘effective’’—i.e. authorship is

separated from content and people cannot easily identify authorship. It may also be the
case that while we asked about the extent of prior communication based on all channels,
the GSS comments are of course textual, which represents only one of the channels used
in prior communication and, thus, can provide only a subset of possible social cues used
as the basis for attributions. Future studies might look at prior communication through
written media only. This potential threat to the validity of the results presumes, however,
great stylistic and cue differences for individuals between oral and written communica-
tion and that these are not correlated.
It is also plausible that some users would craft their comments to conceal their identity
and thus maintain anonymity. This might decrease the accuracy of attributions, but
would not affect attempts at attribution. Another possible contribution to Type-II error
might be that we asked participants to attribute statements to people they have worked
with but may not know that well. However, participants were asked to make attributions
of comments (not people), were asked only to make attributions if they thought they
knew who the author of the comment was, and had worked with each other for many
years. Further, the pairwise network matrix comparison specifically tests for differences
in strength of prior interactions.
Third, different naturally occurring groups may have different abilities to evaluate cues
from and attribute authorship to technically anonymous cues in a GSS. The results did
not reveal a pattern of relationships that was consistent across every group. It may be
that, while technical anonymity ‘‘works’’ (it separates content from authorship identity),
low social anonymity confounds, or reassembles, these components, leading people away
from an absence of any attributions to a presence of false ones. The more people
communicate with others, the more they may think they ‘‘know’’ others’ thoughts and
cues, leading them to overestimate their confidence in being able to attribute (mediated)
content to (technically anonymous) authors. In general, however, the small sample sizes
(in the statistical sense) of the separate groups increase the likelihood of Type-II error.
To the extent that prior communication with one’s group members does influence
one’s ability to detect and interpret cues about the identities of others making technically
anonymous GSS comments, specific pairwise relations and specific network locations
seem to be better predictors—both conceptually and empirically—than simple amount
of prior communication with the group. Thus, attempts to assess the effect of anonymous
comments should take a network-theoretic approach.
Apart from not being able to predict the level of attribution accuracy, there’s a more
disturbing issue involved here. It might seem that the technical anonymity of actual
identities in GSS groups is essentially preserved in spite of attempts to attribute, because
the users were largely inaccurate in their attributions. Note, however, that this is not the
same as maintaining technical anonymity, because users may still be making attributions.
That is, implementing technical anonymity is conceptually and pragmatically inde-
pendent from preventing attributions (whether accurate or inaccurate). It is plausible that
for groups with a salient task and prior history, GSS participants may well be quietly
attempting to identify authors of technically anonymous comments. They would be
applying status cues, etc., derived from prior communications (especially their location in
the communication network) and from on-line textual style and word choice, to the
evaluation of those comments themselves. However, not only is this counter to the
448 S. C. HAYNE AND R. E. RICE

presumed benefits of using the anonymity feature of a GSS (reducing social cues to
increase diversity of comments and to reduce a priori evaluations of comments based on
the identity of the author), but it appears that to the extent that participants do make
such identifications, they would be biasing their evaluations based on largely incorrect
authorship attributions. This is extremely disconcerting because we have little idea how
their judgments, and thus subsequent decisions, are being influenced by this mis-
attribution. Clearly, a topic for future research would be to analyse the decision-making
process and decision outcomes based on inaccurate vs. accurate, attributions, especially
when participants are socially contentious and the issues salient.
It may seem that GSS contexts are rather atypical for most groups, and thus the
disconcertingly low levels of accuracy, as well as the counterpredicted relationships of
prior communication with accuracy, could be dismissed as technological artefacts. After
all, few groups are anonymous when not using GSS to accomplish real tasks. However,
there are instances of naturally occurring, well-established groups with varying levels of
prior communication, that do generate and evaluate separate, anonymous comments
from specific authors. Many social environments that involve refereeing or making
judgments, such as classrooms or academic publishing, also involve power differentials.
If, in these environments, people are attempting to make attributions, it may suppress the
free exchange of ideas, bias those judgments or punish others based on incorrect or
atypical responses.
For example, the principle behind double-blind journal submissions is one motivating
the use of GSS noted earlier: the reduction of social, status and other cues that could bias
one’s judgment of the manuscript. However, anonymity of evaluations has some prob-
lems. The current peer review system, e.g. has been described as one in which power
relations have become dominant (Garfield, 1986; Blank, 1991). In such social systems, as
well as in anonymous conferencing systems, the anonymity of the author may in fact
encourage (unintentionally) irresponsible behavior. Other instances come to mind, such
as the familiar ‘‘suggestion box’’ in an organizational department, paper-based voting to
make decisions or graffiti on the wall of an office lavatory. However, our focus here is
the GSS context—i.e. the situation with which we are theoretically and empirically
concerned.
There are at least two possibly disconcerting implications of these (admittedly
inconsistent) results. First, most GSS designers, facilitators, researchers and users
seem to assume that technical anonymity implies social anonymity. This study
raises doubts about this assumption. From a facilitator’s perspective, it may be
helpful to advise anonymous GSS participants not to make attributions or to try
to change their writing styles. Further, comments could be randomized before
being presented to the whole group, with a note and a discussion that they have been
changed.
Second, one would like to think that social anonymity is influenced by contextual
variables such as cues derived from prior communication and on-line content. Here,
based on the hour or so of comments produced by other group members, we did not find
much evidence for this; rather, the comments generated mis-attributions. It may be that
CMC users, in general, do attempt to make judgments about others’ identities but have
more ongoing contextual cues such as cross-comment references, so that such attribu-
tions could be more accurate than those based on the fragmented nature of brainstorm-
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 449

ing in GSS. Consequently, the level of inaccuracy reported here may not be as high in
technically anonymous computer-mediated communication in general.
While our results are clearly not definitive, they do raise and clarify some previously
hidden assumptions about GSS anonymity, and arouse concerns about subtle con-
sequences of the use of GSS anonymity features. It is apparent from this preliminary
study that the social aspect of GSS anonymity needs to be explored more fully; it is not
the same thing as technical anonymity. We plan on attempting to resolve several
problems with this study in future experiments that achieve the following.

1. Develop ways for the researcher to accurately identify an author’s comments

without compromising the natural group’s comfort with the system.
2. Include other measures of prior relationships, such as time spent working with each
other member.
3. Assess the extent to which people think they informally attempt to identify authors
of comments.
4. Assess the extent to which they think their attempts are correct.

We feel that this initial study has uncovered some assumptions in prior GSS anonym-
ity research that should be reconsidered.

We would like to acknowledge the intellectual contribution and access to meetings provided by
Paul Licker (University of Calgary, Alberta), and the insightful comments provided by Joe Walther
(NorthWestern University, Illinois) and various anonymous reviewers on an earlier version of this
paper.

References
ARKES, H. (1981). Impediments to accurate clinical judgment and possible ways to minimize their
impact. Journal of Consulting and Clinical Psychology, 49, 323—330.
BERNARD, H., KILLWORTH, P. & SAILER, L. (1982). Informant accuracy in social-network data V.
An experimental attempt to predict actual communication from recall data. Social Science
Research, 11, 30—66.
BLANK, R. (1991). The effects of double-blind versus single-blind reviewing: experimental evidence
from ¹he American Economic Review. The American Economic Review, 81, 1041—1067.
CONNOLLY, T., JESSUP, L. & VALACICH, J. (1991). Effects of anonymity and evaluative tone on idea
generation in computer-mediated groups. Management Science, 36, 689—703.
CULNAN, M. & MARKUS, M. (1987). Information technologies: electronic media and intra-
organizational communication. In F. JABLIN, L. PUTNAM, K. ROBERTS & L. PORTER,
Eds. Handbook of Organizational Communication, pp. 420—443. Newbury Park, CA: Sage.
DAFT, R. & LENGEL, R. (1984). Information richness: a new approach to managerial behavior and
organization design. Research in Organizational Behavior, 6, 191—233.
DAFT, R. & LENGEL, R. (1986). Organizational information requirements, media richness and
structural design. Management Science, 32, 554—571.
DENNIS, A., GEORGE, J., JESSUP, L., NUNAMAKER, J. & VOGEL, D. (1988). Information technology
to support electronic meetings. MIS Quarterly, 12, 591—624.
DESANCTIS, G. & GALLUPE, B. (1987). A foundation for the study of group decision support
systems. Management Science, 33, 589—609.
FREEMAN, L. (1979). Centrality in social networks: conceptual clarification. Social Networks, 1,
215—239.
GARFIELD, E. (1986). Refereeing and peer review: part 1. Current Contents, 31, 3—11.
450 S. C. HAYNE AND R. E. RICE

GEORGE, J., EASTON, G., NUNAMAKER, J. & NORTHCRAFT, G. (1990). A study of collaborative
group work with and without computer-based support. Information Systems Research, 1,
394—415.
HACKMAN, J. & KAPLAN, R. (1974). Interventions into group process: an approach to improving
the effectiveness of groups. Decision Sciences, 5, 459—480.
HAMIL, R., WILSON, T. & NISBETT, R. (1979). Insensitivity to sample bias: generalizing from
atypical cases. Journal of Personality and Social Psychology, 39, 578—589.
HAYES, B. & HESKETH, B. (1989). Attribution theory, judgmental biases and cognitive behavior
modification: prospects and problems. Cognitive Therapy and Research, 13, 211—230.
HEIDER, F. (1958). ¹he Psychology of Interpersonal Relations. New York, NY: Wiley.
HENRION, M. (1980). Assessing Probabilities: A Review. Pittsburgh: Carnegie-Mellon University
Press.
HEWSTONE, M. & JASPARS, J. (1987). Covariation and causal attribution: a logical model of the
intuitive analysis of variance. Journal of Personality and Social Psychology, 53, 663—672.
HILTON, D. & JASPARS, J. (1987). The explanations of occurrences and non-occurrences: a test of
the inductive logic model of causal attribution. British Journal of Social Psychology, 26,
189—201.
HILTZ, S. & TUROFF, M. (1993). ¹he Network Nation, 2nd edn. Cambridge, MA: MIT Press.
HILTZ, S., TUROFF, M. & JOHNSON, K. (1989). Experiments in group decision making: disinhibi-
tion, de-individuation and group process in pen name and real name computer conferences.
Decision Support Systems, 5, 217—232.
HUBERT, L. & SCHULTZ, J. (1976). Quadratic assignment as a general data analysis strategy. British
Journal of Mathematical and Statistical Psychology, 29, 190—241.
IACOBUCCI, D. & McGILL, A. (1990). Analysis of attribution data: theory testing and effects
estimation. Journal of Personality and Social Psychology, 59, 426—441.
JESSUP, L., CONNOLLY, T. & GALEGHER, J. (1990a). Group process with an idea-generating task.
MIS Quarterly, 14, 313—321.
JESSUP, L., CONNOLLY, T. & TANSIK, D. (1990b). Toward a theory of automated group work: the
de-individuating effects of anonymity. Small Group Research, 21, 333—348.
JESSUP, L. & TANSIK, D. (1991). Decision making in an automated environment: the effects of
anonymity and proximity with a group decision support system. Decision Sciences, 22,
266—279.
JESSUP, L. & VALACICH, J., Eds. (1993). Group Support Systems: New Perspectives. New York, NY:
Macmillan.
JONES, E. & NISBETT, R. (1972). The actor and the observer: divergent perceptions of the causes of
behavior. In E. JONES, D. KANOUSE, H. KELLEY, R. NISBETT, S. VALINS & B. WEIDER, Eds.
Attribution: Perceiving the Causes of Behavior. Morristown, NJ: General Learning Press.
KAHNEMAN, D. & TVERSKY, A. (1973). Intuitive prediction. biases and corrective procedures. In
D. KAHNEMAN, P. SLOVIC & A. TVERSKY, Eds. Judgment ºnder ºncertainty: Heuristics
and Biases, New York, NY: Cambridge University Press.
KELLEY, H. (1955). ¹he Psychology of Personal Constructs, Vols. 1 and 2. New York, NY: Norton.
KELLEY, H. (1967). Attribution in social psychology. In D. LEVINE, Ed., Nebraska Symposium on
Motivation, Vol 15, pp. 192—238. Lincoln, NE: Lincoln University of Nebraska Press.
KELLEY, H. (1973). The process of causal attribution. American Psychologist, 28, 103—128.
KELTNER, J. (1989). Facilitation: catalyst for group problem-solving. Management Communication
Quarterly, 3, 8—31.
KERR, N. & BRUUN, S. (1981). Ringelman revisited: alternative explanations for the social loafing
effect. Personality and Social Psychology Bulletin, 7, 224—231.
KIESLER, S., SIEGEL, J. & McGUIRE, T. (1984). Social psychological aspects of computer-mediated
communication. American Psychologist, 39, 1123—1134.
KORIAT, A., LICHTENSTEIN, S. & FISCHOFF, B. (1980). Reasons for confidence. Journal of Experi-
mental Psychology: Human Learning and Memory, 6, 107—118.
KRAUSS, R., APPLE, W., MORENCY, N., WENZEL, C. & WINTON, W. (1981). Verbal, vocal and
visible factors in judgments of another’s affect. Journal of Personality and Social Psychology,
40, 312—320.
ATTRIBUTION ACCURACY IN GROUP SUPPORT SYSTEMS 451

LEA, M. & SPEARS, R. (1991). Computer-mediated communication, de-individuation and group

decision-making. International Journal of Man—machine Studies, 34, 283—301.
LICKER, P. (1992a). Knowing ourselves and knowing others: exposing anonymity in GSS. Proceed-
ings of the S¼ Decision Sciences Institute Conference, San Antonio, TX, pp. 184—186.
LICKER, P. (1992b). You said what? Understanding the role and function of anonymity in GSS.
Proceedings of the Administrative Sciences Association of Canada Conference, Quebec City.
LIM, F. & BENBASAT, I. (1991). A communication-based framework for group interfaces in
computer-supported collaboration. Proceedings of the 24th Hawaii International Conference on
Systems Sciences, Kauai, Hawaii, pp. 610—620.
MCARTHUR, L. (1972). The how and what of why: some determinants and consequences of causal
attributions. Journal of Personality and Social Psychology, 22, 171—193.
MCGILL, A. (1989). Context effects in judgments of causation. Journal of Personality and Social
Psychology, 57, 189—200.
MCGRATH, J. & HOLLINGSHEAD, A. (1993). Putting the ‘group’ back in group support systems:
some theoretical issues about dynamic processes in groups with technological enhancements.
In L. M. JESSUP & J. S. VALACICH, Eds. Group Support Systems: New Perspectives, pp. 78—96.
New York, NY: Macmillan.
McLEOD, P. (1992). An assessment of the experimental literature on electronic support of group
work: results of a meta-analysis. Human—Computer Interaction, 7, 257—278.
MENNECKE, B., HOFFER, J. & WYNNE, B. (1992). The implications of group development and
history for group support system theory and practice. Small Group Research, 23, 524—550.
MOORE, C. (1987). Group techniques for idea building. In Nominal Group ¹echnique. Newbury
Park, CA: Sage.
MOSVICK, R. & NELSON, R. (1987). ¼e’ve Got to Start Meeting ¸ike ¹his! A Guide ¹o Successful
Business Meeting Management. Glenview, IL: Scott, Foresman.
NISBETT, R. & ROSS, L. (1980). Human Inference: Strategies and Shortcomings of Information
Judgment, Englewood Cliffs, NJ: Prentice-Hall.
NUNAMAKER, J., DENNIS, A., VALACICH, J., VOGEL, D. & GEORGE, J. (1991). Electronic meeting
systems to support group work. Communications of the ACM, 34, 40—61.
PINSONNEAULT, A. & KRAEMER, K. (1990). The effects of electronic meetings on group processes
and outcomes: an assessment of the empirical research. European Journal of Operational
Research, 46, 143—161.
POLLARD, C. & HAYNE, S. (1996). The reality of meetings and the use of automated meeting tools,
Proceedings of the Americas Information Systems Conference, pp. 250—253.
RICE, R. (1984). Mediated group communication. In R. E. RICE & Associates. ¹he New Media:
Communication, Research and ¹echnology, pp. 129—154. Beverly Hills, CA: Sage.
RICE, R. (1987). Computer-mediated communication and organizational innovation. Journal of
Communication, 37, 65—94.
RICE, R. (1993). Using network concepts to clarify sources and mechanisms of social influence. In
W. RICHARDS, Jr. & G. BARNETT, Eds. Advances in Communication Network Analysis,
pp. 43—52. Norwood, NJ: Ablex.
RICE, R. & LOVE, G. (1987). Electronic emotion: socio-emotional content in a computer-mediated
communication network. Communication Research, 14, 85—108.
RICE, R. E. & SHOOK, D. (1990). Relationships of job categories and organizational levels to use of
communication channels, including electronic mail: a meta-analysis and extension. Journal of
Management Studies, 27, 196—229.
SHORT, J., WILLIAMS, E. & CHRISTIE, B. (1976). ¹he Social Psychology of ¹elecommunications.
London: Wiley.
SIEGAL, J., DUBROVSKY, V., KIESLER, S. & McGUIRE, T. (1986). Group processes in computer
mediated communication. Organizational Behavior and Human Decision Process, 37, 157—187.
SPROULL, L. & KIESLER, S. (1986). Reducing social context cues: electronic mail in organizational
communication. Management Science, 32, 1492—1512.
TAYLOR, S. & FISKE, S. (1978). Salience, attention and attribution: top of the head phenomena. In
L. BERKOWITZ, Ed. Advances in Experimental Social Psychology, Vol. 11, pp. 249—288. New
York, NY: Academic Press.
452 S. C. HAYNE AND R. E. RICE

TVERSKY, A. & KAHNEMAN, D. (1974). Judgment under uncertainty: heuristics and biases. Science,
185, 1124—1131.
VALACICH, J., JESSUP, L., DENNIS, A. & NUNAMAKER, J. (1992). A conceptual framework of
anonymity in group support systems. Proceedings of the 25th Hawaii International Conference
on Systems Sciences, Kauai, Hawaii: IEEE Press, pp. 101—112.
WALTHER, J. (1992). Interpersonal effects in computer-mediated interaction: a relational perspect-
ive. Communication Research, 19, 52—90.
WALTHER, J. (1993). Impression development in computer-mediated interaction. Western Journal
of Communication, 57, 381—398.
WEINER, B. (1985). Spontaneous causal search. Psychological Bulletin, 97, 74—84.
ZIGURS, I., POOLE, M. & DESANCTIS, G. (1988). A study of influence in computer-mediated
communication. MIS Quarterly, 12, 625—644.