ica 2011 hidden-proflie

Upload: shinji-ikari

Post on 10-Oct-2015

12 views

Category:

Documents


0 download

DESCRIPTION

PPT

TRANSCRIPT

  • Runninghead:HIDDENPROFILEPARADIGM 1Hidden-Profile Paradigm, a Tautology or an Uninteresting Argument at Best: Toward Complex

    and Dynamic Models of Group Information Processing

    Introduction

    Despite the prevalent belief that two heads are better than one, small group researchers

    have found a considerable amount of empirical evidence showing that groups frequently fail to

    outperform individuals in decision making. For example, since firstly formulized by Janis (1972),

    groupthink has been successfully replicated and developed by many follow-up studies (e.g.,

    Esser, 1998; McCauley, 1989; Schafer & Crichlow, 1996) and now became a classic theory that

    accounts for the failure of group decision-making. Also, group polarization studies constitute

    another stream of research on the failure of groups in decision making (e.g., Moscovici &

    Zavalloni, 1969), which has been extended into political settings (e.g., Sunstein, 2002; 2004).

    Although much empirical research has presented for the last several decades that groups

    are necessarily better than individuals in decision making, there has been no unanimous

    agreement upon the group mechanism that leads to the failures. For example, Janis (1972)

    pointed out a higher level of group cohesiveness as a main factor that groups tend to reduce the

    possibility of comprehensive examination of all potential alternatives by preventing

    contradictory views from being expressed. However, many other empirical studies suggested that

    the impacts of group cohesiveness on group performance, specifically decision quality is

    mediated or moderated by many other intervening variables (for overview, see Esser, 1998) .

    As an alternative theoretical explanation for the failure of groups in decision making,

    Stasser and Titus (1985; 1987) suggested an insightful model, called information sampling

    model. In their conception, group decision-making is regarded as or at least involves the process

    by which the information unevenly distributed among individuals is collected and aggregated so

  • Runninghead:HIDDENPROFILEPARADIGM 2that the group can establish a pool of information as the basis of its decision-making. Specifically,

    based on their simple and intuitive formula, they argued that decision-making groups tend to

    focus their discussions on information that their members already know (i.e., shared

    information) rather than what individual members uniquely possess (i.e., unshared

    information) (for overviews, see Stasser & Titus, 2003; Wittenbaum, Hollingshead, & Botero,

    2004, etc.).

    Furthermore, they attributed the failure of groups in decision making to the group

    tendency to exclude the information that group members uniquely possess (what they called

    overlooked gems: Stasser & Titus, 2003, p. 304) from the information pool. Such argument

    indeed shifted the focus of group decision-making research from the characteristics of groups

    and their members to the processes and mechanisms of group decision making, more specifically,

    information pooling: for example, comparison between the choosing-the-best and rank-order

    procedures or between computer-mediated and face-to-face discussions (Hollingshead, 1996). In

    short, Stasser and Tituss (1985; 1987) simple idea established the foundation of a new research

    area, called hidden-profile paradigm.

    For the last 25 years since Stasser and Titus proposed the information sampling model

    and developed it into the hidden-profile paradigm, many empirical studies have been conducted

    and provided consistent findings to some degree. A recent meta-analysis (Li, Yuan, McLeod, &

    Xia, 2010) found that (a) groups mentioned more common information than unique information;

    (b) hidden profile groups were seven times less likely to find the correct answer than

    information-all-shared groups; (c) whereas mentioning unique information indeed leads to better

    decision quality, discussion focus did not affect group decision quality; and (d) face-to-face

    groups did not differ from CMC groups in group decision quality. Despite the consistency of the

  • Runninghead:HIDDENPROFILEPARADIGM 3previous findings, a close logical examination of those findings reveals that many of them are no

    more than tautological statements or uninteresting arguments at best. Moreover, as Reimer,

    Reimer, and Czienskowski (2010) pointed out in their meta-analysis, a fundamental question of

    the hidden-profile research still remains as unanswered: How do groups manage to attenuate the

    group discussion bias toward shared information, and thereby How do groups establish a more

    comprehensive information pool? In other words, a large part of the questions concerning group

    information processing and its mechanisms have not been fully addressed.

    From this perspective, the present study attempts to point out the limitations of the

    previous studies on hidden-profiles due to both theoretically and methodologically problematic

    approaches, and to suggest a new approach appropriate for dealing with the complex and

    dynamic nature of group processes. For these purposes, a set of mathematical analyses using

    differential equations were carried out. Based on the findings of the analyses, general behaviors

    of decision making groups are discussed in the relation with the previous research. Finally, the

    study suggests the directions for further research.

    Problems with the Information Sampling Model

    To answer the question of why groups often fail to outperform individuals in decision

    making, Stasser and Titus (1987) proposed a simple mathematical formula that describes the

    probability that unshared information that individual members uniquely possess is mentioned

    during group discussion as a function of group size (i.e., the number of participants). Specifically,

    provided that the probability that any participant mentions his or her own unique information is

    equal to p, and that each of n participants has an independent and identical chance to mention his

    or her own unique information (i.e., p = pi = pj for all i and j; the iid assumption), the probability

  • Runninghead:HIDDENPROFILEPARADIGM 4that the group as a whole does not mention the unique information (that is, the probability of the

    group mentioning common information only, P(Gshared)) can be written as following:

    1 1 (1)Where n is the number of participants and p is the probability that an individual participant

    mentions his/her own unique information. Since p is a probability, it is a real number between 0

    and 1. Therefore, as n increases, (1 p)n approaches 0, while P(Gshared) approaches 1 (Figure 1):

    that is,

    lim 1

    As shown in Figure 1, a group with a large number of participants tends to discuss shared

    information much more than unshared information. Even when the probability that unshared

    information is mentioned by an individual, p, is exactly equal to the probability that shared

    information is mentioned, (1 p), and/or even when the group size is quite small (e.g., a 2-

    person group), the group is more likely to discuss the shared information than the unshared

    information.

    [Insert Figure 1 about here]

    Such discussion bias favoring shared information is easily expected to result in the

    dominance of the information common among group members over the unique information. As

    this continues, the unique information will be squeezed out of the group discussion, and the

    group will reach a conclusion without a critical review of the unique information that might be

    important for making a better decision. Stasser and Titus (1985; 1987) argue that the failure in

    group decision-making is mainly due to the incompleteness of the information pool

    overwhelmed by the shared information. Based on this, it is easy to infer that a group would a

    worse decision or not a best decision at best without well-designed group discussion rules that

  • Runninghead:HIDDENPROFILEPARADIGM 5prevent the dominance of shared information and promote the exchanging of unshared

    information.

    With no doubt, Stasser and Tituss equation above precisely captures the probability of

    the dominance of shared information during group discussions, but only in terms of frequencies.

    To be sure, in a group with a large number of members the shared information is more likely

    (more precisely speaking, more frequently) to be discussed than the unshared information.

    However, this does not necessarily mean that the amount of the shared information has been

    accumulated more than that of the unshared information. By definition, shared information is

    what group members already possess in common. Therefore, no matter how many times or how

    frequently it is mentioned, the total amount of shared information can never be accumulated,

    simply because mentioning another piece of shared information is nothing other than just

    repeating what has been previously mentioned (recall the concept of information redundancy of

    Shannon, 1948). In contrast, no matter how infrequently it is mentioned, mentioning a piece of

    the unshared information will linearly increase its total amount, because any piece of the

    unshared information is completely new and not redundant at all to what has been previously

    mentioned. Thus, if a group makes a decision relying not upon simply the frequencies by which

    different types of information are mentioned but upon the amount of the information that has

    been accumulated so far, it is not impossible for a group with a large number of members to

    reach a conclusion that the unshared information designates.

    If the total amount of information does matter in group decision making, it might not be

    surprising that a group can choose the alternative that unshared information (or less dominant

    information in terms of frequency) supports. If it is the case, such possibility will lead us to

    suspect whether or not any particular group information processing mechanisms actually change

  • Runninghead:HIDDENPROFILEPARADIGM 6the direction of group decisions at all, because the change of the decisional direction may happen

    by chance. This is one of the most important reasons that previous research on hidden-profiles

    should be reexamined.

    Impacts of sharing unique information on group decisions

    In empirical research on hidden-profiles, more generally, studies on group information

    sampling, experimental groups are typically asked to reach a joint decision by choosing one out

    of a set of choice alternatives. Such tasks include choosing a company for a hypothetical

    financial investment, selecting a suspect in a murder mystery, selecting an individual for student

    body president, or developing a diagnosis for a medical case (for overviews, see Stasser &

    Birchmeier, 2003). In order for the groups to make a choice, they are given multiple sets of

    information items, each of which supports one choice alternative than any others. For example,

    in frequently cited experiment research conducted by Hollingshead (1996), a number of 3-person

    groups were asked to choose one out of three companies for a hypothetical financial investment,

    and they were provided with three different sets of information items, each of which supports

    each company. However, the information was not randomly distributed among the three

    members within each group; but rather, the sharedness of the information among the members

    was sophisticatedly manipulated in a systematic way. In particular, the information items that

    support one company, say A, (i.e., the correct choice) were given to single members separately

    (i.e., unshared information), while those that support the other companies, say B and C, (i.e., the

    incorrect choices) were given to all group members in common (i.e., shared information). Then,

    the researcher examined whether one discussion rule that apparently facilitates to share the

    unique information (i.e., the rank-order procedure) would lead to the correct decision (i.e.,

    company A) than the other rule that might inhibit exhaustive consideration of all information

  • Runninghead:HIDDENPROFILEPARADIGM 7items, specifically unshared ones (i.e., the choose-the-best procedure). Finally, the researcher

    concluded that the groups following the rank-order rule were more likely to choose the correct

    alternative than the others, because the rank-order procedure did facilitate sharing the unique

    information items. In short, when unique information is shared with all the group members, the

    group will make a right decision. Although the effect sizes of sharing unique information on

    group performance vary, previous research on hidden-profile tasks generally confirmed this

    finding (Lu, et al., 2010). Then, what is the hidden problem of this finding?

    To illustrate the problem embedded in hidden-profile research, let us suppose a typical

    and general experimental setting for hidden-profile tasks used in previous studies. Groups are

    given a decision-making task that requires choosing one out of multiple alternatives (say,

    alternative A, B, C, ). In any cases, it is reasonable to assume that which alternative is better

    than any others is completely unknown, at least to the participants. If the best choice is already

    known in advance, the decision-making task itself is unnecessary. That is, a group of people do

    not have any practical reasons to waste their valuable resources, including time, in order to

    choose the best one already known.

    Further suppose that multiple sets of information are given, each of which supports one

    alternative exclusively (say, information , , , ). Then, the question is whether or not it is possible for the participants (either as individuals or as a whole group) to determine the

    significance of a particular information item in their decision-making: that is, in Stasser and

    Tituss terms, Is this information item an overlooked gem or not? The logical answer to this

    question is No. No matter how it looks significant during group discussion, the significance of

    an information item completely depends on the correctness of the alternative that the item

    supports, which is, in turn, completely unknown. For instance, an information item is

  • Runninghead:HIDDENPROFILEPARADIGM 8significant, only if alternative A that the item supports turns out to be the correct one; otherwise,

    the item is not significant at all. However, since the correctness of alternative A is undermined in advance, so is the significance of the information item . In short, the significance of a given information item is determined not by itself but by the correctness of the corresponding

    alternative, which is completely unknown. Logically speaking, thus, picking up and examining

    an information item during group discussion is a blind procedure to the group members.

    Let us go back to the typical findings of hidden-profile research when unique

    information is shared with all the group members, the group will make a right decision. As

    described above, the significant information items (known to researchers but unknown to

    participants) were given to individual members separately (i.e., unshared information). Then, the

    typical findings can be rephrased as When unshared information is significant, and when the

    unshared significant information is eventually shared through a certain discussion process, the

    group will make a correct choice. Obviously, this statement is a tautology or an uninteresting

    statement at best.

    Suppose that a researcher manipulates the sharedness of information in the opposite way

    to the previous research: that is, he or she distributes significant information to all the group

    members (i.e., shared information), whereas distributes insignificant information to individual

    members separately (i.e., unshared information). And he or she asks the groups to follow a

    certain discussion rule that facilitates sharing the unshared information. Two possible results can

    be conceived. First, if the groups make the wrong choice, this proves that the general finding of

    hidden-profile research is tautological, because the hypothetical experiment confirms that the

    quality of group decision depends on the significance of the unshared information. Second, the

    groups still make a correct choice, this proves that neither the group discussion rules nor the

  • Runninghead:HIDDENPROFILEPARADIGM 9manipulation of the sharedness of the information has no impacts on the quality of group

    decision. Therefore, the hidden-profile paradigm is a tautology or an uninteresting argument at

    best.

    Nonetheless, such tautological research design can be also useful when it is used to trace

    in what way a particular information item or a set of information items is disseminated,

    especially when the information is unshared prior to group discussion. It is not very interesting to

    merely address whether a group following a specific rule will make a correct decision, because it

    totally depends on the manipulation of the sharedness of information. Rather, it might be more

    interesting to examine how the relative proportion (not the relative frequency but the amount) of

    a particular type of information fluctuates over time through complex and dynamic group

    processes.

    Complex and Dynamic Models of Group Information Processing

    Group processes, either those by which groups generate new ideas or those by which they

    make joint decisions, are usually considered as to be complex and dynamic. Therefore, these two

    words are ones of the most frequently appear in the small group literature. However, few studies

    in the research area, if any, have attempted to conceptualize or to operationalize the notion of

    complexity and dynamics of group processes, unlike in other fields: e.g., Kolmogorov

    complexity in computer science (Kolmogorov, 1963). The lack of clear or at least agreeable

    definitions of complexity and dynamics results in inconsistent findings and more seriously

    prevents us from deepening our understanding of group processes, while it remains as so-called a

    black box. Sometimes, the complexity and dynamics of group processes are mentioned in

    order to justify the poor explanatory power of the research models or the inconsistent findings

    even within single studies. I do not believe that it is possible to investigate something complex

  • Runninghead:HIDDENPROFILEPARADIGM 10and dynamic without clear definitions of them. This is not a mere personal belief but a general

    principle of scientific research that is articulated in any social scientific research textbooks.

    This is, I suppose, in large part because of the a-century-long convention in psychological

    research, in general, in social scientific research. Unlike other sciences in which complex and

    dynamic phenomena are handled very well, social science has relied mainly upon correlational

    statistical analytic methods, which assume methodological individualism (i.e., the iid assumption)

    and linear relationships among variables (e.g., ANOVA family, various regression analysis).

    Although it is undeniable that the conventional statistical methods in social science have led to

    tremendous advance in the development of human knowledge for the last century, they are not

    suitable for the examination of complex and dynamic phenomena in behavioral and social

    sciences (for the detailed discussion, see Brown, 2007, Chapter 1). In brief, conventional

    statistical models that require programmable formulae in advance of any empirical tests, such as

    t-distribution for t-test, are not appropriate to complex and dynamic phenomena, those whose

    statistical behaviors are hard to specify in advance.

    Pointing out the limitations of the conventional statistical methods, many scholars have

    suggested new analytical tools, such as differential and difference equations, as alternatives for

    examining complex phenomena (e.g., Coleman, 1964; Brown, 1995; Rapoport, 1983; Simon,

    1957, etc.). Ironically, however, most small group researchers, those who (over)emphasize the

    importance of sharing unshared information for successful group performance, have ignored the

    unfamiliar approaches and exploited only familiar analytical tools: i.e., conventional statistical

    methods. As many small group theories predict, attachment only to conventional methods will be

    detrimental to the development of the small group research area.

  • Runninghead:HIDDENPROFILEPARADIGM 11

    From this perspective, the present study proposes a new model of group decision-making

    employing the new analytical approach, specifically differential equation techniques, as part of

    effort to introduce unshared information into the small group research area as a decision-

    making group with expectation that it will contribute to deepening our understanding of complex

    and dynamic group processes.

    A General Setting of Hidden-Profile Experiment

    In order to build a mathematical model of group information processing, this study adopts

    the manipulation of the sharedness of information used in Hollingsheads (1996) experiment as

    the simplest prototype to be extended into a general setting of hidden-profile experiment. In her

    experiment, she assigned every three participants into a group (i.e., 3-person groups) and asked

    either to choose one out of three companies (i.e., the choose-the-best procedure) or to rank all the

    three companies (i.e., the rank-order procedure) for a hypothetical financial investment. A total

    of six information items were given to each group members within a group: two of which favor

    one company, say A; the rest of which support the other two companies, say B and C,

    respectively. The first two information items were given to individual members uniquely (i.e.,

    unshared information); while the rest were given to the group members in common (i.e., shared

    information). The manipulation of the sharedness of the information described here is

    summarized in Figure 2.

    [Insert Figure 2 about here]

    It should be noted that while the total number of information items is equal to 18 (= 3

    persons * 6 items), the total amount of information is equal to 10 (= 2 * 3 unique information

    items + 4 * 1 common information) because the redundant information items are not counted as

    additional information. Provided that a group makes a decision based on the amount of

  • Runninghead:HIDDENPROFILEPARADIGM 12information rather than the number of information items available, when all the information

    items become available through group discussion (i.e., 6 items favoring company A > 4 items

    favoring either company B or C) the group will choose company A (assumed to be the correct

    choice); but when only the common information items become available (i.e., no item favoring

    company A < 4 items favoring either company B or C) as the information sampling model

    predicts, the group will choose either company B or C (assumed to be the wrong choices). In the

    case in which group members make their own decisions prior to group discussion, each member

    is more likely to choose either company B or C than A, since the amount of information favoring

    B or C is greater than A (i.e., 2 item favoring company A < 4 items favoring either company B or

    C).

    Let us extend this prototypical setting into a general one. The group size needs not to be

    limited to three; but rather we can extend it to n-person cases. Similarly, the number of

    information items given to individual members is not necessarily limited to 6, either; but can be

    extended to K. At the individual member level, the number of information items supporting one

    alternative (i.e., company A), which belong to each group member individually, p, should be less

    than that of information items supporting the other alternatives (i.e., companies B and C), which

    is common to all the members, q. So that when decisions are made individually, individual

    members are more likely to choose either B or C than A: that is, p < q. Accordingly, the sum of

    both types of information items that an individual member hold should be equal to K (= p + q).

    At the group level, the total number of information items is equal to Kn; whereas the total

    amount of information is equal to p + nq. In order for the group to choose company A with all

    information available, the amount of information supporting company A should be greater than

    that supporting either company B or C: i.e., p < nq. To avoid notational confusion, hereafter,

  • Runninghead:HIDDENPROFILEPARADIGM 13denote the total amount of information items supporting company A as X instead of p; and that of

    information items supporting either company B or C as Y instead of nq: hence, X < Y. The

    general setting described here is diagramed as Figure 3.

    [Insert Figure 3 about here]

    Finally, it is necessary to consider the scale of time unit. Unlike the previous empirical

    studies in which time is typically measured in minutes, in this study time is measured in a

    theoretical manner: that is, a unit of time is defined as the amount of time that should be spent for

    processing a single information item (i.e., mentioning and discussing it). Such a theoretical

    measure is not impossible, if every information item is perfectly manipulated as having an equal

    length and an equal level of readability and difficulty or if the variance of the time spent for

    processing the items is statistically controlled.

    The general and theoretical setting of hidden-profile experiment developed in this section

    has several advantages over conventional experimental designs. First, it enables to overcome the

    limitations of laboratory settings usually reported at the end of research articles: for example,

    limited group sizes (typically 3 to 5), limited sample sizes (no more than 200), and sampling bias

    (e.g., college students). Second, it removes any possible measurement errors and allows to

    examine the pure effects of variables/manipulation of interest. As mentioned earlier, for example,

    the theoretical measure of time allows us to completely remove any possible noises related to

    empirical time measurement and to look at the pure effects of time on group decision-making

    processes. Third, the general and theoretical setting makes it possible to manipulate key

    (independent) variables that are impossible to in real world. For example, the following section

    will set four different rules of the sequence of information items to process: (1) independent, (2)

    self-reinforcing, (3) negative, and (4) positive mutual influences among different types of

  • Runninghead:HIDDENPROFILEPARADIGM 14information. Furthermore, in order to test the sensitivity of group processes to initial conditions,

    the impacts of all possible initial settings will be tested.

    Model 1: Independent model1

    The first model, labeled independent model, assumes the simplest situation, in which

    every group member has an equal chance to provide one out of K information items at a time,

    every single item is to be selected and discussed with an equal probability, and moreover the

    chance to select one type of information item is independent of what was discussed previously.

    Now, let x denote the amount of the information supporting alternative A at a time point

    of t. The increase rate of x at t can be expressed in a differential equation as:

    1 (2)

    where p is the probability of the information items supporting A to be selected, X is the total

    amount of information supporting A, and thus, 1 x/X is the proportion of the information items

    that have been not discussed until t. The solution of the differential equation is

    (3)

    where the initial value of x at t = 0 is equal to 0: that is, none of the information items supporting

    A has been discussed prior to group discussion.

    As shown in both equations (2) and (3), x monotonously increases over time: i.e., dx/dt >

    0 for all t. However, the increase rate itself decreases over time: i.e., d2x/d2t < 0 for all t. This is

    because additional items are likely to be redundant with the items that have been already

    discussed. Also, x tend to approach to its upper bound X over time: i.e., lim . 1Sofar,thegeneralandtheoreticalsettinghasassumeda3alternativesituation(i.e.,companiesA,B,andC).Intherestofthepaper,itwillbereducedintoatwoalternativesituationbymergingthelasttwoalternativesintoasingleone(i.e.,alternativeAvs.B).Thereasonistwofold:first,a3alternativesituationinvolvesaseriesofcomplicatedmathematicalcalculationofpartialdifferentialequationsthatmightexceedthepagelimit;second,theresultsofanalysesofa3alternativesituationishardtodisplayina2dimensionalplane.

  • Runninghead:HIDDENPROFILEPARADIGM 15

    In a similar way, the growth curve of the other type of information supporting B, say y,

    can be derived from its differential equation and solution:

    1 (3)

    (4)

    where q is the probability of information items supporting B to be selected, Y is the total amount

    of information supporting B, and thus, 1 y/Y is the proportion of the information items that

    have been not discussed until t.

    The growth curve for y shows the similar shape to that of x. However, the increasing rate

    of y is much smaller than that of x at early stage, while it is much larger than that of x at late

    stage, since p > q and X > Y. In fact, it is easily shown that y eventually exceeds x, say at a time

    point of t* by the intermediate value theorem. This implies that where any information items are

    freely exchanged by any members (i.e., equal chance to be mentioned), the shared information is

    dominant over the unshared information at early stage of group discussion (i.e., x > y at 0 < t <

    t*), but the unshared information eventually becomes dominant (i.e., x < y at t > t*).

    Figure 4 presents the time series plot of a numerical example in which X = 10, Y = 12, p

    = .6, and q = .4. As shown in the figure, both curves increase sharply at early stage, but the

    increasing rates slow down, and eventually approach to their upper bounds. The curve for x

    increases faster than the curve for y at early stage; but the cumulative amount of information y

    exceeds that of x at t = 46. That is, from the beginning up to t = 46, x is dominant over y. After

    the time point, however, y becomes dominant.

    [Insert Figure 4 about here]

  • Runninghead:HIDDENPROFILEPARADIGM 16

    From the findings of the simplest model, we can draw two important conclusions. First,

    no matter what size of a decision-making group, the amount of unshared/unique information

    eventually exceeds that of shared/common information specifically after t* because of

    information redundancy, which is opposite to what the information sampling theory predicts.

    Second, whether a group relies upon the shared or the unshared information is determined by the

    amount of time given for group discussion. However, in laboratory experiments, different groups

    following different discussion rules are given the equal amount of time. For example, in

    Hollingsheads (1996) study, four different groups (the choose-the-best vs. the rank-order

    procedure * face-to-face vs. CMC groups) were all given 20 minutes, but it was shown that the

    rank-order and FtF groups were more likely than the others to rely upon unshared information.

    Thus, it might be seen that time had no significant impacts on group processes. However, this

    previous finding can be interpreted to indicate that a unit amount of information could be

    processes in a shorter time under the condition of the rank-order rule and FtF than the others. For

    example, the one-to-one comparison between alternative companies may facilitate processing

    contrasting information items, and the nonverbal cues used in FtF communication may also

    reduce the information processing time. Therefore, it can be said that those groups spent more

    amount of the theoretical time than the others, even though they spent the same amount of actual

    time as the others.

    The independent model allows further examination of infinitely many different cases in

    which initial conditions are differently set up. For example, what if one type of information was

    already dominant over the other even before group discussion begins (i.e., x(t) > y(t) 0 or y(t) > x(t) 0). Such case can be also interpreted as the situation in which power relations among group members are not evenly distributed. That is, a unit of information provided by a powerful

  • Runninghead:HIDDENPROFILEPARADIGM 17member will be weighted more than that by a powerless member. For this, phase diagram

    techniques were used and summarized in Figure 5.

    The solid curve from the origin of the coordinate (i.e., x(0) = y(0) = 0: no preexisting

    information dominant) to the equilibrium point represents the trajectory of the relative

    proportions of the amounts of two types of information x and y. When the curve is below the

    straight line y = x, x is dominant over y, vice versa. By setting the initial condition differently

    (i.e., x(0) y(0) 0), it is possible to investigate the impacts of the initial conditions on the direction of trajectories and the area through which they pass. As shown in Figure 5, however,

    different initial conditions do not generate any complex behaviors during group discussion,

    compared to complex phenomena found in natural worlds: e.g., the Lorenz attractors and the

    competitive Lotka-Volterra model. In fact, all initial points are directed toward the equilibrium

    point at the top of the plane. Moreover, even when the complexity of group process rules

    increases, the impacts of different initial conditions remains less complex (discussed later). In

    turn, this implies that group processes may not be complex enough to justify the poor and/or

    inconsistent findings of previous research.

    [Insert Figure 5 about here]

    Model 2: Self-reinforcing model

    The independent growth model examined previously does not appear to be complex or

    dynamic enough to reflect real-world group processes. Therefore, the second model, self-

    reinforcing model, is suggested to capture self-reinforcing positive feedback such that makes the

    model more realistic as well as more complex and dynamic. As its name itself articulates, the

    self-reinforcing positive feedback refers to as a situation in which the more one type of

    information was mentioned previously, the more likely the same type of information is to be

  • Runninghead:HIDDENPROFILEPARADIGM 18mentioned, vice versa. This assumption reflects the violation of the independence assumption

    between the chances of information items being mentioned at different time points. This implies

    that conventional statistical methods are no longer suitable to test the self-reinforcing model.

    This model can be expressed in differential equations as following:

    1 (4)

    1 (5)

    and their solutions are:

    (6)

    (7)

    Figure 6 presents the time series plot of a numerical example of the model in which X =

    10, Y = 12, p = .6, and q = .4: the same values as the previous example. While the increase rates

    of the two curves in the independent model monotonously decrease (Figure 4), those in the self-

    reinforcing model fluctuate over time (Figure 6). At early stage of group discussion, both types

    of information increase slowly because neither of them have been mentioned enough. However,

    both curves increase sharply around the inflection points (i.e., the point at which the second-

    derivative equals 0). Compared to the independent model, the amounts of both types of

    information grow much more slowly on average. For example, x in the independent model

    almost reaches its upper bound around t = 100, while that in the self-reinforcing model reaches

    around t = 150. This is because the self-reinforcing process encourages the same type of

    information to be selected in the following rounds, and thereby increases the information

  • Runninghead:HIDDENPROFILEPARADIGM 19redundancy, too. Although the two curves eventually intersect with each other, that is, y exceeds

    x, it also takes much longer until the time point (t* = 151; cf. t* = 46 in the previous model).

    [Insert Figure 6 about here]

    The self-reinforcing model seems more complex than the previous one due to the

    fluctuating increase rates of both curves. However, a phase diagram analysis reveals almost the

    same pattern as the independent growth model (Figure 7). Even though the trajectory that begins

    from the origin (the solid curve in Figure 7) stays longer the area below the y = x line and passes

    through the deep right-bottom corner, it eventually pass through the y = x line and end up at the

    equilibrium point. The impacts of different initial conditions shows similar patters to the

    previous model, but the initial points around the center of the plane tend to change their phases at

    faster rates than those around the peripheral.

    Therefore, the two conclusions drawn previously are still held even when additional

    complexity and dynamics are added to group processes: (1) the amount of unshared/unique

    information eventually exceeds that of shared/common information because of information

    redundancy; (2) whether a group relies upon the shared or the unshared information is

    determined by the amount of time given for group discussion.

    Model 3.1: Interdependent model with negative mutual influences

    The previous two models the independent model and the self-reinforcing model do

    not take into account the possibilities of interdependent relationships between different types of

    information. Two possible scenarios can be considered: one is that the increase in one type of

    information suppresses that in the other (i.e., negative mutual influences); the other is that the

    increase in one type of information leads to the increases in the other type (i.e., positive mutual

  • Runninghead:HIDDENPROFILEPARADIGM 20influences). The first possible interdependent models can be expressed in a system of two

    differential equations with two unknowns as following2:

    1

    1

    (10)

    Figure 8 presents the time series plots of numerical examples of the two models in which

    X = 10, Y = 12, p = .6, and q = .4: the same values as the previous examples. As shown in Figure

    8, the negatively interdependent growth model shows almost an identical pattern with the

    independent growth model. The phase diagram also looks similar to both the independent growth

    model and the self-reinforcing model with slight differences (Figure 9)

    [Insert Figure 8 and 9 about here]

    The two conclusions drawn previously are still held even when eve more complexity and

    dynamics are added to group processes, i.e., the interdependence between the two different types

    of information: (1) the amount of unshared/unique information eventually exceeds that of

    shared/common information because of information redundancy; (2), whether a group relies

    upon the shared or the unshared information is determined by the amount of time given for group

    discussion.

    Model 3.2: Interdependent model with positive mutual influences (the ping-pong model)

    The final model of this study also involves the possibility of interdependent relationships

    between different types of information but in the opposing way to the previous model: i.e.,

    positive mutual influences. In other words, the model assumes that the increase in one type of

    information leads to the increases in the other type. This mechanism seems like playing a ping-

    2Duetothecomplexityofthesystemsofdifferentialquestions,theexplicitsolutionsforneitherofsystemscanbefound.Forthisreason,numericalintegrationmethodswereused.

  • Runninghead:HIDDENPROFILEPARADIGM 21pong game: the increase in one type of information causes the increase in the other type, and it

    will once again increases the first type of information. The model can be formulized as a system

    of two differential equations with two unknowns as following:

    1

    1

    (11)

    Figure 10 presents the time series plots of numerical examples of the two models in

    which X = 10, Y = 12, p = .6, and q = .4: the same values as the previous examples. Unlike the

    three previous models the independent model, the self-reinforcing model, and the negatively

    interdependent growth model, this final model shows the most striking patterns: both curves run

    parallel to each other until they begin to converge to their asymptotes.

    The phase diagram shows such distinctive patters more clearly (Figure 11). The trajectory

    that begins from the origin goes along with the y = x line, although it is slightly above the

    straight line, which means that both types of information increases at an almost equal rate for

    regardless of the proportions, relative amounts, or the upper bound of different types of

    information. In addition, the initial points around the straight lines appear to be quite static (i.e.,

    the short velocity arrows).

    The behavioral patterns of group processes shown in the positively interdependent

    models can be summarized as: (1) unshared/unique information is as likely to be mentioned as

    shared/common information, which is contrasting to both the prediction of the information

    sampling model and the previous conclusion of this study; and (2) whether a group tends to rely

    upon the shared or the unshared information is relatively independent of the amount of time

    spent for group discussion, which is also contrasting to the previous conclusions. In short, the

  • Runninghead:HIDDENPROFILEPARADIGM 22positively interdependent model allows groups to build their information pools in which different

    types of information increase at almost an equal rate. However, it is noted that the model does

    not guarantee better decisions, simply because the significance of the unshared information is

    unknown.

    Conclusion

    This study takes a close look at the major findings of hidden-profile research and the

    information sampling model (Stasser & Titus, 1985; 1985) from which hidden-profile paradigm

    has been originated. As a result, a set of critical limitations of hidden-profiles has been found.

    First, the information sampling model fails to pay attention to the information redundancy and

    therefore the amount of information, mainly focusing on the frequencies of information items

    mentioned during group discussion. Therefore, it predicts that groups tend to prefer the

    information that all group members share in common (i.e., shared information) over the

    information that individual members possess individually (i.e., unshared information). According

    to the probability model that they proposed (Equation 1), it is true that the share information will

    be mentioned more frequently than the unshared information. However, because of the

    informational redundancy, the shared information cannot be accumulated even when it is

    repeatedly mentioned. Therefore, the unshared information is not necessarily squeezed out of

    group information processing. Furthermore, the possibility of the survival of the unshared

    information without any intentional or pre-designed discussion rules implies that certain group

    discussion mechanisms, which were previously considered as critical for the survival of the

    unshared information, might not have significant impacts on the survival of the unshared

    information or the quality of group decision.

  • Runninghead:HIDDENPROFILEPARADIGM 23

    Second, the present study found that the key findings of the hidden-profile research are

    tautological statement. Logically speaking, the significance of a given piece of information,

    either shared or unshared, is completely unknown. Nevertheless, the previous research assumes

    that the unshared information is significant and critical for group decisions (overlooked gems),

    and examined whether or not sharing the unshared information would lead better decisions.

    However, such tautological research design might be very useful, if it is used for tracing certain

    types of information (e.g., shared or unshared; correct or incorrect, etc.) during group discussion

    processes. However, most studies failed to examine the group processes (Reimer, et al., 2010).

    This study pointed out the lack of appropriate analytical tools as the main reason that the

    important questions have long been unanswered. Using different equations as alternative

    analytical tools (Brown, 2007), the present study suggested four different models of group

    information processing: (1) the independent model, (2) the self-reinforcing model, (3) the

    negatively interdependent model, and (4) the positively interdependent model, each of which

    represents group discussion rules with different levels of complexity and dynamics, and detected

    the behavioral patterns of competitive dominance of different types of information within a

    group. The results from the first three models showed that the unshared information is not

    necessarily squeezed out of group information processing but rather become dominant over the

    shared information at late stage. Also, the results showed that whether groups rely upon the

    shared or the unshared information is determined by the amount of time spent for group

    discussion. On the other hand, the forth model, the positively interdependent model, suggested

    the most striking patterns. When two different types of information encourage the likelihood of

    each other to be mentioned, the unshared information is exchanged as much as the shared

    information, regardless of the amount of time spent.

  • Runninghead:HIDDENPROFILEPARADIGM 24

    In this study, one of the simplest forms of differential equations was used as the

    unshared information to the group of researchers to examine the emerging patterns of group

    information processing and to suggest a new direction of hidden-profile paradigm and small

    group research, in general. However, as discussed above, the unshared information is not always

    significant than the shared information. Therefore, the suggested direction should be developed

    into a set of hypotheses against which empirical evidence can be tested.

  • Runninghead:HIDDENPROFILEPARADIGM 25References

    Brown. C. (l995). Chaos and catastrophe theories. Thousand Oaks, CA: Sage.

    Coleman, J. S. (1964). Introduction to mathematical sociology. New York: Free Press.

    Esser, J. K. (1998). Alive and well after 25 years: A review of groupthink research.

    Organizational behavior and human decision process, 73(2/3), 116-141.

    Hollingshead, A. B. (1996). The rank order effect in group decision making. Organizational

    Behavior and Human Decision Processes,68, 181193

    Janis, I. (1972). Victims of groupthink; a psychological study of foreign-policy decisions and

    fiascoes. MI: Houghton.

    Kolmogorov, A. N. (1963). On tables of random numbers. Theoretical Computer Science, 207(2),

    387--395.

    Lu, L., Yuan, Y. C., McLeod, P., and Xia, L. (2010). Twenty-five years of hidden profiles in

    group decision making: A meta-analysis. Paper presented at the 96 Annual Convention

    of National Communication Association, San Francisco.

    McCauley, Clark. (1989). The nature of social influence in groupthink: compliance and

    internalization. Journal of Personality and Social Psychology. 57(2), 250-260

    Moscovici, S., and Zavalloni, M. (1969). The group as a polarizer of attitudes. Journal of

    Personality and Social Psychology, 12, 125-135.

    Rapoport, A. (1983). Mathematical models in the social and behavioral sciences. New York:

    Wiley.

    Reimer, T., Reimer, A., and Czienskowski, U. (2010). Decision-making groups attenuate the

    discussion bias in favor of shared information: A meta-analysis. Communication

    Monographs, 77(1), 121-142

  • Runninghead:HIDDENPROFILEPARADIGM 26Schafer, M. and Crichlow, S. (1996). Antecedents of groupthink: a quantitative study. The

    Journal of Conflict Resolution, 40(3), 415435

    Schulz-Hardt, S., Brodbeck, F., Mojzisch, A., Kerschreiter, R., & Frey, D. (2006). Group

    decision making in hidden profile situations: Dissent as a facilitator for decision quality.

    Journal of Personality and Social Psychology, 6, 1080-1093.

    Shannon, C. (1948). A Mathematical Theory of Communication. The Bell System Technical

    Journal, 27, 379423

    Simon, H. A. (1957). Models of man: Social and rational. New York: Wiley.

    Stasser, G. & Titus, W. (2003). Hidden profiles: A brief history. Psychological Inquiry, 14, 304-

    313.

    Stasser, G., and Titus, W. (1985). Pooling of unshared information in group decision making:

    Biased information sampling during discussion. Journal of Personality and Social

    Psychology, 48, 1467-1478

    Sunstein, C. (2002). The law of group polarization. The Journal of Political Philosophy, 10(2),

    175-195.

    Sunstein, C. (2004). Group polarization and 12 angry men. Negotiation Journal, 23(4), 443-447.

  • Runninghead:HIDDENPROFILEPARADIGM 27

    Figures

    Figure 1. A numerical example of the information sampling model with p = .5

    Figure 2. A typical research design for hidden-profile task

  • Runninghead:HIDDENPROFILEPARADIGM 28

    Figure 3. The general and theoretical setting of hidden-profile experiments

  • Runninghead:HIDDENPROFILEPARADIGM 29

    Figure 4. Growth curves of the independent model, where X = 10, Y = 12, p = .6, and q = .4

    Figure 5. A phase diagram of the independent model, where X = 10, Y = 12, p = .6, and q = .4

  • Runninghead:HIDDENPROFILEPARADIGM 30

    Figure 6. Growth curves of the self-reinforcing model, where X = 10, Y = 12, p = .6, and q = .4

    Figure 7. A phase diagram of the self-reinforcing model, where X = 10, Y = 12, p = .6, and q = .4

  • Runninghead:HIDDENPROFILEPARADIGM 31

    Figure 8. Growth curves of the negatively interdependent model, where X = 10, Y = 12, p = .6,

    and q = .4

    Figure 9. A phase diagram of the negatively interdependent model, where X = 10, Y = 12, p = .6,

    and q = .4

  • Runninghead:HIDDENPROFILEPARADIGM 32

    Figure 10. Growth curves of the positively interdependent model, where X = 10, Y = 12, p = .6,

    and q = .4

    Figure 11. A phase diagram of the positively interdependent model, where X = 10, Y = 12, p = .6,

    and q = .4