ingroup bias experiments

An official website of the United States government

Official websites use .gov A .gov website belongs to an official government organization in the United States.

Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

• Publications
• Account settings
• Advanced Search
• Journal List

Ingroup bias in a social learning experiment

• Author information
• Article notes
• Copyright and License information

Corresponding author.

Received 2021 May 24; Revised 2022 Dec 18; Accepted 2022 Dec 20; Issue date 2023.

This article is made available via the PMC Open Access Subset for unrestricted research re-use and secondary analysis in any form or by any means with acknowledgement of the original source. These permissions are granted for the duration of the World Health Organization (WHO) declaration of COVID-19 as a global pandemic.

Does social learning and subsequent private information processing differ depending on whether the observer shares the same group identity as the predecessor whose action is observed? In this paper, we conduct a lab experiment to answer this question, in which subjects first observe a social signal and then receive a private signal. We find that subjects put greater weights on the social signal if they share with the predecessor the same group identity that is induced in the experimental environment. We also provide suggestive evidence that such an ingroup-outgroup difference cannot be explained by individuals’ beliefs of the predecessor’s rationality. Moreover, heterogeneous effects of group identity exist in weights given to the subsequent private signal: Compared to when the predecessor is an outgroup, those who have learned from an ingroup predecessor put a greater (smaller) weight on the private signal if it contradicts (confirms) the social signal. We conjecture that such group effects are consistent with the perspective that group identity works as a framing device and brings about certain decision heuristics in the social signal phase, which no longer exist in the private signal phase.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10683-022-09788-1.

Keywords: Ingroup bias, Social learning, Belief updating, Laboratory experiment

Introduction

Social learning is an important source for individuals to gather information to update their beliefs and make decisions. It has drawn growing research interest from economists since the early work pioneered by, for example, Banerjee ( 1992 ), Bikhchandani et al. ( 1992 ), Anderson and Holt ( 1997 ). In a digital era like ours, in which social network sites play important roles in information dissemination, it’s easy to think of situations where individuals observe behavior and opinions of others before they receive and process objective signals themselves, regarding various types of events and facts, such as the safety and effectiveness of Covid-19 vaccines, and the validity of investment opportunities. 1

Does social learning differ whether people learn from others who share the same or have different social identities with them? For example, does a Facebook post from a fellow alumna have the same effect on shaping one’s own beliefs and choices as a Reddit post from a stranger? How about a twitter comment from someone belonging to the same or opposing political party? If an ingroup-outgroup difference exists in processing the social signal, then the next question is whether the impact persists, or perhaps the difference will be offset once the person receives objective signals herself. Answering these questions may provide novel insights for understanding phenomena such as chamber effect and polarization (see Levy and Razin 2019 for a review), which are among the major themes in the public arena of our time. 2

In this paper, we investigate these research questions in a controlled lab experiment, with group identities induced in the experimental environment. We also simplify away from the numerous factors in the real world by adopting the existing paradigm in the social learning literature—asking subjects to make incentivized judgments regarding a neutral event (color of the jar) that requires no previous knowledge and invokes no prior opinions or intrinsic preferences. We conjecture that the case of such a minimal group and a neutral event probably provides us a conservative estimate of the ingroup-outgroup difference in social learning, if it ever exists. On the one hand, natural group identities might have stronger effects. On the other hand, if, for example, the belief is ego-relevant or identity-relevant, the ingroup bias might interact with a motivated belief-updating process, resulting in a larger ingroup-outgroup difference when processing the social signals. Specifically, we follow the experimental procedure used by De Filippis et al. ( 2022 ): The color of an urn is either black or white—the two events are, a priori , equally likely; subjects are asked to state their beliefs about the likelihood of the urn being white, and the belief-elicitation task is incentivized. A randomly chosen subject receives a noisy binary symmetric signal about the color of the urn, after which she is asked to state her belief. Then other subjects observe the first subject’s stated belief and make the same type of prediction themselves. Finally, each subject receives another conditionally independent signal and makes a new prediction.

As we aim to investigate whether ingroup bias exists in social learning, besides the standard social learning (SL) treatment as in De Filippis et al. ( 2022 ), we design group social learning (GSL) sessions, in which before the social learning phase we add a phase of group assignment and enhancement following the previous experimental work on group identity (Tajfel et al., 1971 ; Chen & Li, 2009 ). Depending on whether sharing the same group identity as the first subject, subjects in the GSL sessions are divided into the ingroup social learning (ING) treatment and the outgroup social learning (OUTG) treatment. We repeat the learning procedure for 30 rounds in all treatments. In the first 20 rounds of the GSL sessions, we randomly select the first subject among members of one group, and then switch to the other group for the last 10 rounds, with the order counter-balanced—In this way, we embed a within-subject design in the between-subject design. Besides, we also design an individual decision-making (IDM) treatment as De Filippis et al. ( 2022 ) did, to provide a benchmark for comparison. Instead of observing the first subject’s prediction and then a private signal, all subjects receive two independent private signals consecutively in the IDM treatment.

We hypothesize that when processing the social signal, an ingroup-outgroup difference exists. Our experimental results confirm this hypothesis, as we find that based on weights computed with the Bayesian updating structure, subjects in the OUTG treatment underweight the social signal compared to the Bayesian benchmark to a significantly greater extent than those in the ING treatment, whereas subjects in SL or IDM treatments are more close to Bayesian updating. Moreover, in round 21 of the experiment, we elicit subjects’ beliefs on the first subject’s rationality , that is, the likelihood that the first subject has reported beliefs regarding the color of the urn in the correct direction as the signal suggests. We find no statistically significant ingroup-outgroup differences in these elicited beliefs on the predecessor’s rationality; neither can these rationality beliefs explain the impact of the group identity on the weight placed on the social signal. By checking whether and how the ingroup-outgroup difference varies over the rounds, we find that it is more prominent in the first several rounds than in the later rounds. These results suggest that instead of working via a belief-updating process about the predecessor’s rationality, the ingroup-outgroup difference might reflect heuristics that prescribe caution and suspicion with others’ behavior, especially with the outgroup predecessors’, which are brought about by the group frame.

We further examine whether sharing the same group identity as the predecessor has an impact on the weights put on the subsequent private signal. We find that when the private signal contradicts (confirms) the social signal, subjects in the ING treatment put a greater (lower) weight on the private signal than those in the OUTG treatment. Since the weight for the private signal is calculated with the Bayesian updating structure using the stated belief after observing the social signal as the prior belief, we conjecture that these empirical patterns are, again, consistent with that group identity works as a framing device, the heuristics cued by which no longer exist in the private signal phase. In other words, because the ingroup-outgroup difference in processing the social signal no longer exists in the private signal phase when subjects process the social and private signals together, subjects behave as if they were putting a higher (lower) weight on the private signal if it contradicts (confirms) the social signal, given the way the weights are calculated in our analyses.

Finally, since our study adopts an experimental design similar to that used by De Filippis et al. ( 2022 ), we can check whether our results replicate theirs. They found that when processing the private signal, subjects weighed the signal as a Bayesian agent if the signal confirmed the social signal; subjects overweighted the signal if it contradicted the social signal. In contrast, in our SL and GSL treatments, after receiving a contradicting private signal, subjects weigh the signal similar to a Bayesian agent; subjects underweight the signal when it confirms the social signal. At first glance, the results are quite different between the two studies. However, the commonality lies in that subjects nevertheless put a higher weight on a contradicting private signal than on a confirming one. The intuition captured by the specific case of the Likelihood Ratio Test Updating (LRTU) model discussed by De Filippis et al. ( 2022 ), which reconciles the empirical patterns they found, can be used to explain our results as well, as long as we are willing to impose an ad hoc assumption that in our private signal phase, the updating benchmark is not Bayesian updating but some extent of underweighting. Intuitively, subjects hold confident prior beliefs about the predecessor’s rationality after observing the social signal. Then, a confirming private signal retains such confident beliefs; subjects update their event beliefs under a general tendency of underweighting. In contrast, a contradicting private signal makes one adjust downward her belief about the predecessor’s rationality, and thus she adjusts downward the weight she puts on the social signal. Because such an adjustment is not captured by our way of calculating the weight placed on the private signal, it will appear as if the subject puts more weight on the private signal—the greater weight then offsets the general tendency to underweight among our subjects, rendering the belief-updating consistent with Bayesian updating. We provide evidence supporting this interpretation by showing that the rationality belief we elicit in round 21 becomes significantly lower after observing a contradicting private signal than after observing a confirming one.

Our work contributes to the large and growing literature on the various aspects of group identity effects in experimental economics. Given the significant impact of group identity on lots of important outcomes such as distributive preferences, cooperation behavior, evaluations of task performances, and general attitudes (for example, Tajfel et al. 1971 ; Ryen and Kahn 1975 ; Brewer 1979 ; Bettencourt et al. 1992 ; Brown et al. 1999 ; Bernhard et al. 2006 ; Goette et al. 2006 ; Chen and Li 2009 ; Currarini and Mengel 2016 ; Xu et al. 2020 ; Li 2020 ,) it is reasonable to conjecture that social learning from an ingroup member could be different compared to that from an outgroup member. So far, research directly examining the impact of group identity on social learning is limited. To the best of our knowledge, only in two recent experimental studies, Ash and Van Parys ( 2015 ) and Berger et al. ( 2018 ) investigated how group identity affects the formation of information cascade. They both adopted the minimal group paradigm similar to the one used in the current paper to assign subjects to two groups (Tajfel et al., 1971 ; Chen & Li, 2009 ). After the group assignment stage, subjects engaged in repeated information cascade games based on Anderson and Holt ( 1997 ) in which each of them in a sequence knew all the prior players’ choices of the true state as well as her own private signal, and the group identities of the prior players were revealed. They found that information cascades are more likely to occur when learning from ingroup members compared to when learning from outgroup individuals. There are two major differences between the present study and theirs. First, we let subjects report their beliefs about the true state instead of making a binary decision regarding on which state to bet, which allows us to compute and compare the weights they put on the signals in different situations. Second, we collect subjects’ beliefs two times, one time after they observe a predecessor’s belief and the other time after they receive their own private signal, with which we can distinguish between their belief updating upon learning the two types of information. In contrast, Ash and Van Parys ( 2015 ) and Berger et al. ( 2018 ) let subjects make a decision after knowing the social and private information at the same time, focusing on the formation of information cascades. To summarize, we investigate the effect of group identity on social learning in a more general sense that is not constrained to the context of information cascade, and we also examine the implications beyond the social learning phase per se , that is, when new private information is available.

Our results also speak to an on-going debate in the broader literature on how to explain the identity effect in shaping behavior and cognition. First, the social preferences perspective has been popular among experimental economists, who represent the impact of group identity as changes in the values of prosocial parameters (Chen & Li, 2009 ; Müller, 2019 ; Robson, 2021 ). Second, some other researchers have focused on the role of beliefs. For example, Ockenfels and Werner ( 2014 ) examined the effect of second-order belief of group identity on ingroup favoritism in dictator games. Tanaka and Camerer ( 2016 ) showed that differences in outgroup discrimination can be explained by beliefs about the characteristics of potential outgroups. Grimm et al. ( 2017 ) found that beliefs about the behavior of other groups’ extent of ingroup favoritism matter. Third, an alternative tradition exists, viewing group identity as a framing device (Tajfel et al., 1971 ; Bacharach, 2006 ). Recent psychological research suggests that more similarity or experiences of more aligned interests among subjects may lead to a more “trusting mindset” (Gino et al., 2009 ; Levine et al., 2014 ; Bolton et al., 2016 ). More recently, Guala and Filippin ( 2017 ) and Filippin and Guala ( 2017 ) provided experimental evidence supporting the heuristics interpretation, by showing that the effect of group identity on distributive choices is highly dependent on contextual factors and that it can be easily displaced by changing the framing of decisions. As argued earlier, our findings can be viewed as supporting the framing perspective.

Last, this paper relates to the broader literature on other types of cognitive bias in information processing. For example, existing experimental evidence suggests that people tend to overweight their private information and underweight the social information (Nöth & Weber, 2003 ; Çelen & Kariv, 2004 ; Goeree et al., 2007 ; Weizsäcker, 2010 ; Ziegelmeyer et al., 2013 ). It is then an open question how the ingroup-outgroup difference, if existing, interacts with such a tendency to underweight social information. In our data, we do not find evidence supporting that people in general underweight social information compared to private information. Interestingly, we find that introducing the group frame makes people underweight the social signal, especially if the predecessor is an outgroup.

The remaining of the paper structures as follows: We introduce the experimental design in Sect. 2 , report the results in Sect. 3 , and then conclude as well as provide our discussions of the results and future research directions in Sect. 4 .

The experiment

The experiment consists of four treatments, SL, IDM, and GSL (including ING and OUTG). Our SL and IDM treatments are similar to those of De Filippis et al. ( 2022 ). 3 In the GSL treatments, we assign subjects to two groups and they observe the choices made by either an ingroup or an outgroup predecessor before making their own decisions. Then each subject receives an independent private signal. Details of experiments are as follows.

There are two parts in all treatments. In part 1 , subjects see paintings in pairs on the screen. One of the paintings is drawn by Paul Klee, and the other is drawn by Wassily Kandinsky. The artists’ names do not appear on the screen, and the positions of the paintings are randomly decided, that is, it could either be Klee on the left and Kandinsky on the right, or the opposite. Subjects are requested to indicate which one they prefer in each pair of paintings. There are 5 pairs in total. Then, according to their relative preferences over the two artists, subjects in the GSL treatments are assigned to the Klee group or the Kandinsky group. We use relative preferences instead of absolute preferences over the two artists for group assignment since we want to create two groups with the same or close number of members to control for the impact of group size within sessions. There were 14–22 participants in a GSL session. If the number of participants n is even, then both group sizes are n /2; otherwise, one group size is ( n - 1 ) / 2 and the other is ( n + 1 ) / 2 . There is no group assignment in the IDM and SL treatments, while subjects in these two treatments nevertheless make choice between the paintings.

Subjects in all treatments then perform a task of moving sliders to the middle points of slider bars. In the IDM and SL treatments a subject earns 5 points for every slider she moves to the middle point. In the GSL treatments a subject earns 0.5 points for every slider either she or any other group member moves to the middle point. Payoffs of the first part are not revealed until the end of the experiment. We introduce such common payoffs to enhance subjects’ perception of group identity (Charness et al., 2007 ). 4

Part 2 lasts for 31 rounds. In the beginning of each round, the computer randomly chooses one from two (virtual) urns, a white urn and a black urn, with a 50–50 chance. The white urn contains 7 white balls and 3 black balls, while the black urn contains 3 white balls and 7 black balls. Subjects won’t know which urn has been chosen until the end of each round. In round 1 of all treatments, the computer randomly draws a ball out of the chosen urn with replacement for each subject, and subjects observe whether it is a white ball or a black ball. Then, subjects report individually the probability that they believe the chosen urn is white in this round, ranging from 0 to 100 (up to two decimal points). The first round serves as a chance for practice.

In each of rounds 2–31 in the IDM treatment, subjects first observe a ball privately drawn from the chosen urn and report individually their subjective probability beliefs about the chosen urn being white. This process repeats twice in a round. In each of rounds 2–31 in the SL and GSL treatments, the computer first randomly chooses one subject as Player A and then randomly draws a ball out of the chosen urn with replacement. Only Player A observes the drawn ball and reports her subjective probability belief about the chosen urn. After Player A’decision, every other subject observes the subjective probability Player A reported ( social signal ), and reports individually his own belief about how likely the chosen urn is white. Last, each subject privately observes another randomly drawn ball ( private signal ) and reports his belief about the chosen urn again.

In the GSL treatments, the group identity of Player A is revealed. We introduce the ING and OUTG treatments within the same session, which means that Player A sends social signals to both groups at the same time. Such a design feature guarantees that any group identity effect we find in information processing is not because subjects expect Player A to act differently when sending social signals to ingroup versus outgroup. In particular, in a Klee-Kandinsky (Kandinsky-Klee) session, one subject is chosen randomly from the Klee (Kandinsky) group in each of rounds 2–21, and from the Kandinsky (Klee) group in each of rounds 22–31. Therefore, those in the Klee (Kandinsky) group in a Klee-Kandinsky session are in the ING (OUTG) treatment in rounds 2–21, but are in the OUTG (ING) treatment in rounds 22–31.

Subjects’ reported beliefs are incentivized with a quadratic scoring rule, that is, payoff = 150 - 0.015 × ( V - reported belief ) 2 . We adopt the quadratic scoring rule following De Filippis et al. ( 2022 ). 5 If the chosen urn is white, V = 100 ; otherwise, V = 0 . In rounds 2–31, the computer randomly draws one out of two reported beliefs to generate the payoff of the corresponding round. After round 31 ends, the computer randomly selects one round from rounds 1–10, 11–21, and 22–31 respectively. The sum of subjects’ earnings of the three selected rounds makes up the main part of their payoffs in the second part.

Moreover, we design two questions to elicit subjects’ beliefs about Player A’s rationality in round 21 of the SL and GSL treatments. Since beliefs about Player A’s rationality serve our secondary research objectives, we elicit them only in round 21, leaving the other rounds succinct. We ask subjects to report the probability they believe that Player A saw a white ball in the current round. The reported numbers also range from 0 to 100. They are asked twice, once after observing Player A’s action, and the other time after receiving the private signal. The answer is incentivized; the second-time question is phrased as a chance to modify the answer reported for the first time. Payoff = 50 - 0.005 × ( C - reported number ) 2 . If Player A did see a white ball, C = 100 ; otherwise, C = 0 . Subjects’ final earnings include the payoff from this extra question in round 21. Together with Player A’s action, such probability beliefs provide us with a measure of probability beliefs regarding Player A’s rationality. Specifically, if Player A’s action is larger than 50, which suggests a white ball, then the reported probability belief in round 21 that Player A has seen a white ball is equivalent to the probability belief that Player A has not made a mistake, or, that Player A is rational. If, instead, Player A’s action is smaller than 50, the probability that a subject thinks Player A has observed a white ball is the same as her probability belief that Player A has made a mistake. We use this simple and neutral belief question instead of directly asking about how likely subjects think Player A is or is not rational, or how likely that Player A has or has not made a mistake, to avoid potential confusion and minimize the influence of language connotation.

First, we aim at testing whether there is a direct impact of group identity on subjects’ social learning. We focus on α 1 , the weight that individuals put on a social signal calculated from the following formula:

where a 1 is the first belief, a A denotes Player A’s action, and q is the signal precision, which equals 0.7 in our experiment. If α 1 = 1 , it means that the subject updates his belief in a Bayesian way; if α 1 is larger (smaller) than 1, it means that he updates his belief as if a Bayesian belief updater has observed the signal more (fewer) than once.

In the existing experimental studies, Ash and Van Parys ( 2015 ) and Berger et al. ( 2018 ) found that information cascades were more likely to occur when learning from ingroup members than learning from outgroup members. Here, we provide our belief-based reasoning of an ingroup-outgroup difference in social learning under the Likelihood Ratio Test Updating (LRTU) rule model, a non-Bayesian updating model developed and empirically verified by De Filippis et al. ( 2022 ). Below, we describe the intuition; interested readers can find more formal analyses in Online Appendix C.1 .

According to the LRTU model, when an individual observes Player A’s action, she is initially ambiguous about Player A’s rationality, that is, whether Player A correctly takes an action a A > 50 ( a A < 50 ) if she observes a white (black) ball. She first selects her prior belief for Player A’s rationality, and then she updates her evaluation regarding the color of the urn. Specifically, she sets a critical value c and compares it to the likelihood ratio of observing a (set of) signal(s) conditional on the predecessor being rational. If the likelihood ratio is greater than or equal to c , then the individual will select the prior belief that is most confident about the predecessor’s rationality within the possible belief range. Otherwise, she will select the least confident prior belief. The value of c indicates how strong (or weak) the evidence needs to be to support the confidence in the predecessor’s rationality. When only one social signal is available, as in the first phase in each round of our experiment, the likelihood ratio of observing any given signal, conditional on the predecessor being rational, equals one. Let us discuss how group identity can make a difference in the three representative cases as follows.

Suppose that (a) the value of c differs depending on whether the predecessor is an ingroup or an outgroup; for the ingroup c < 1 , whereas for the outgroup c > 1 ; (b) the range of the rationality belief is identical for an ingroup and an outgroup. This means that it requires weaker evidence for the individual to have confidence about an ingroup’s rationality than about an outgroup’s rationality. Then, with the likelihood ratio equal one, the individual will select the most confident prior belief for the ingroup’s rationality while she will select the least confident prior belief for the outgroup’s rationality. Therefore, the individual will have a more confident belief about the ingroup’s rationality than the outgroup’s, and as a result, the weight calculated with the Bayesian updating structure, α 1 , will be higher if the predecessor is an ingroup than if he is an outgroup.

Suppose that (a) c has no ingroup-outgroup difference; c ≤ 1 ; (b) the most confident prior belief within the belief range is more confident for the ingroup than for the ougroup. In this case, the individual will select the most confident rationality belief within the respective range for both the ingroup and the outgroup. Because of the difference in the range, the same conclusion can be drawn as in the first case— α 1 would be higher if the predecessor is an ingroup than if he is an outgroup.

Suppose that (a) c has no ingroup-outgroup differences; c > 1 ; (b) the least confident prior belief within the belief range is more confident for the ingroup than for the ougroup. In this case, the individual will select the least confident rationality belief within the respective range for both the ingroup and the outgroup. Again, because of the difference in the range, the same conclusion can be drawn as in the first case— α 1 would be higher if the predecessor is an ingroup than if he is an outgroup.

In these cases above, the assumption that weaker evidence is required to have confidence in an ingroup’s rationality than that of an outgroup, as well as the assumption that either the most confident or least confident belief about the predecessor’s rationality is higher for an ingroup than for an outgroup, can find support in a broad sense in the existing literature. For example, Cacault and Grieder ( 2019 ) found that people are overconfident in their fellow group members’ intelligence. Summing it up, we propose the following hypotheses: 6

Subjects in the same group as Player A put a greater weight on the social signal than those in the other group do, that is,

α 1 , I N G > α 1 , O U T G .

Subjects in the same group as Player A have a higher belief regarding Player A being rational than those in the other group do. Once controlling for the rationality beliefs, the ingroup-outgroup differences in α 1 , if any, disappears, or at least, diminishes.

Moreover, we are interested in whether group identity has any impact on subsequent processing of private information after the social learning phase. We compute α 2 , the weight that an individual puts on his private signal using the following formula:

where a 2 is the second belief, and s = 1 ( 0 ) means that a white (black) ball was drawn. Similar to α 1 , when α 2 = 1 , it means that the subject updates his belief in a Bayesian way; α 2 deviating from 1 indicates over- or under-updating in comparison to Bayesian updating.

De Filippis et al. ( 2022 ) found an asymmetry of belief updating upon receiving a private signal after a social signal, that is, people overweighted their private signal when it contradicted the social signal while they updated their belief in a Bayesian way when their private signal confirmed the social signal. The authors reconcile such patterns in their results with the Likelihood Ratio Test Updating (LRTU) rule model assuming that the parameter values satisfy certain conditions. Suppose that when receiving the social signal, the individual’s value of c is smaller than one, and thus the individual selects the most confident prior belief regarding the predecessor’s rationality. Then, a private signal confirming the social signal makes the individual stick to the same prior belief on the predecessor’s rationality, because a confirming private signal only increases the likelihood ratio. In contrast, a contradicting private signal will result in a much lower likelihood ratio, which can be smaller than c , making the individual select the least confident prior belief of the predecessor being rational within the range. As a result, the belief-updating of the events in both phases will look as if the individual overweights the private signal, because it is in the opposite direction of the social signal, the weight placed on which is now adjusted down because of lowered beliefs in the predecessor’s rationality.

Following the LRTU model while assuming that c < 1 regardless of the group identity of Player A, we generate hypotheses about the effect of group identity on the weight subjects put on their private signal in exemplary cases as follows. In general, when the private signal is confirming, the likelihood ratio becomes even larger—the belief about Player A’s rationality does not shift; it is then unlikely that the group identity exhibits any effect on the processing of the private signal. When the private signal contradicts the social signal, we discuss two cases below.

Suppose that (a) c is smaller for the ingroup than for the outgroup; (b) the range of the rationality belief is identical for an ingroup and an outgroup. In this case, it is possible that receiving the private signal induces a downward adjustment in beliefs of Player A’s rationality for the outgroup but not for the ingroup, if the updated likelihood ratio lies inbetween the two values of c . Thus, we expect to observe an ingroup-outgroup difference in α 2 after observing a contradicting private signal, since a more dramatic downward adjustment of Player A’s rationality will demonstrate itself in the form of a greater weight placed on the contradicting private signal, calculated with the Bayesian structure.

Suppose that (a) c and the most confident prior belief within the belief range are both the same; (b) the least confident prior belief is more confident for the ingroup than for the outgroup. In this case, it is possible that both for the ingroup and the outgroup, after observing a contradicting private signal, an adjustment of Player A’s rationality takes place and the individual adjusts the prior belief to the least confident one. As an ingroup-outgroup difference exists in the least confident prior belief, a group effect on α 2 as described in the first case will also be observed here.

Subjects in the same group as Player A put less (the same) weight on their private signal when it contradicts (confirms) Player A’s action than those in the other group do, that is,

α 2 , I N G < α 2 , O U T G if s = 1 ( a A < 50 ) & a A ≠ 50

α 2 , I N G = α 2 , O U T G if s = 1 ( a A > 50 ) & a A ≠ 50

The experiment was conducted in October and November, 2018 at Nankai University, at the China Centre for Experimental Social Sciences (CESS), which is a collaborative laboratory between Nankai University and the Nuffield College at University of Oxford. The subject pool of CESS China consists of undergraduate and graduate students in all disciplines at Nankai University. Experimental sessions were advertised within the subject pool of CESS China, and participants were recruited through a Wechat recruiting system (ancademy.org). Upon arrival, subjects were assigned to computers by randomly choosing one card from a pile of numbered cards. After subjects were seated in the lab, instructions of part 1 were distributed. The experimenter read instructions aloud in front of all subjects and then subjects answered understanding questions. After they answered all understanding questions correctly, the experiment started. Instructions of part 2 were distributed after part 1 ended. When the experiment was completed, subjects saw their final earnings on the screen privately and got paid on site. The experiment was computerized using oTree (Chen et al., 2016 ).

In total, 14 sessions were organized and 258 subjects participated in the experiment. On average each session lasted for one hour. Subjects earned 59.4 RMB yuan on average. The numbers of subjects and sessions as well as the means and standard deviations (in parentheses) of subjects’ characteristics are displayed by treatment in Table 1 .

Subjects’ characteristics by treatment

“Experience in experiments” indicates whether subjects have ever participated in an economics or psychology experiment or not. “Having learned prob theory” stands for whether subjects have learned probability theory or not. All variables except Age are dummy variables. None of the pairwise comparisons by treatment using rank-sum tests gives a statistically significant difference at 5 % level

Descriptive analysis and data preparation

Figure 1 shows that the distribution of Player A’s action after observing a white (black) ball has a peak around 70 (30), which is consistent with predictions of Bayes’ theorem. Also, a few Player As choose to report 50, which may be because of strong risk aversion. 7

Distribution of Player A’s actions. Notes : The figures are based on Player As’ actions in the 12 sessions of the SL and GSL treatments over rounds 2-31, that is 360 observations in total. The left and right panels each include 180 observations

Figure 2 displays other players’ beliefs conditional on Player A’s action. If Player A’s action is larger (smaller) than 50, other players should consider that a white (black) ball was drawn. We do find that the distribution of first beliefs has a peak around 70 (30) when Player A’s action is larger (smaller) than 50. Compared to Fig. 1 , a slightly higher percentage of subjects report 50.

Distribution of other players’ beliefs after observing Player A’s action. Notes : The figures are based on other players’ actions upon observing either Player A’s action > 50 or < 50 in the SL and GSL treatments over rounds 2-31. The left panel includes 2913 observations and the right panel includes 2648 observations

Figure 3 presents the distributions of players’ second beliefs, that is after observing a private signal, conditional on that the private signal confirms Player A’s action a A > 50 (a white ball was drawn) and contradicts a A > 50 (a black ball was drawn) respectively. 8 Given that most subjects report 70 as their first beliefs for a A > 50 , if a majority of subjects update their second beliefs in a Bayesian way, we would observe that the distribution of their second beliefs for a confirming (contradicting) private signal had a peak around 84.5 (50). The left panel shows that quite a few beliefs deviate saliently from 84.5 and the mean of the second beliefs is a bit lower than 80, which does not seem to support Bayesian belief updating. The right panel confirms that most people report 50 for a contradicting private signal, which is different from a very asymmetric distribution around 50 shown by De Filippis et al. ( 2022 ).

Distribution of other players’ second beliefs. Notes : The figures are based on other players’ actions upon observing either confirming (a white ball) or contradicting signals (a black ball) when a 1 > 50 , in the SL and GSL treatments over rounds 2-31. The left panel includes 1702 observations and the right panel includes 1211 observations

In Table 2 , we present the three quantiles of α 1 computed using Eq. ( 1 ) by treatment. Note that when a 1 = 0 or 100, the equation generates positive or negative infinity for α 1 respectively. We compute α 1 by approximating a 1 = 0 with 0.01 and a 1 = 100 with 100 - 0.01 . Table 2 shows that subjects display heterogeneity in belief updating upon observing the first signal. For example, some do not update at all ( α 1 = 0 ) while others update the same as a Bayesian agent would do ( α 1 = 1 ). We can see that median subjects in the SL and GSL treatments tend to underweight Player A’s action (the social signal) compared to a Bayesian agent ( α 1 < 1 ). Additionally, subjects in the OUTG treatment attach a lower weight to the social signal compared to the ING treatment. 9

Distributions of α 1 by treatment

Table 3 displays the quantiles of α 2 computed using Eq. ( 2 ), upon observing confirming and contradicting signals respectively, and by treatment. When a 2 = 0 or 100, or a 1 = 0 or 100, the equation generates positive or negative infinity for α 2 . We compute α 2 by approximating a 2 ( a 1 ) = 0 with 0.01 and a 2 ( a 1 ) = 100 with 100 - 0.01 .

Distributions of α 2 by treatment

As shown in Table 3 , in the SL treatment, the three quantiles of α 2 are all smaller upon observing a confirming signal than upon observing a contradicting signal. The median subject updates beliefs in a Bayesian way upon receiving a contradicting signal ( α 2 = 1 ), while she attaches less weight to the second signal compared to a Bayesian agent when observing a confirming signal ( α 2 = 0.64 ). The asymmetry of belief updating also holds for the ING and OUTG treatments, with even lower values of α 2 for observing a confirming private signal. We also compute α 2 for the IDM treatment, where confirming signals mean that two white or two black balls were drawn and contradicting signals mean that a white and a black ball were drawn. We find an asymmetry of belief updating upon receiving two signals in the IDM treatment but in the opposite direction, that is, subjects put more weight on the confirming signal than on the contradicting signal.

Group identity and belief updating

The focus of the present study is to investigate whether group identity has an impact on social learning and subsequent private belief updating. First, we look at whether group identity has an impact on α 1 . Figure 4 displays the cumulative probability distributions of α 1 for ingroup and outgroup. There exists a mild but significant difference between the two distributions with the Kolmogorov-Smirnov test ( p -value = 0.03). Table 2 also shows that the median value of α 1 is lower in the OUTG treatment than in the ING treatment.

CDFs of α 1 with groups. Notes : Observations satisfying - 1 ≤ α 1 ≤ 3 are included

We then include the observations from 2-31 rounds in all treatments and run random effects regressions of α 1 on exogenous variables Ingroup , SL , and IDM , controlling for round dummies, the first-round weight on a drawn ball, and subjects’ characteristics, excluding the extreme observations reporting a 1 = 0 or 100. Table 4 displays the outcomes. We find that the coefficients of Ingroup are statistically significant, which support our H1 . 10

Random effects regressions of α 1

Clustered standard error by session are in parentheses. Control contains age, gender, undergraduate or graduate students, having participated in any economics or psychology experiment, Han or minority, and having learned probability theory or not. ∗ p < 0.1 , ∗ ∗ p < 0.05 , ∗ ∗ ∗ p < 0.01

Result 1: Subjects put more weight on Player A’s action if they are in the same group than in different groups, that is, α 1 , I N G > α 1 , O U T G .

It should be noted that the group impact on α 1 becomes insignificant when including extreme values with a 1 = 0 or 100 in the linear random effects regressions ( p -value = 0.272). Figure 5 shows that the frequency of a 1 = 100 ( a 1 = 0 ) in the OUTG treatment is roughly double that in the ING treatment, given that Player A’s action is larger (smaller) than 50. An extreme value of a 1 consistent with Player A’s action generates α 1 = 10.87 with the approximation a 1 = 100 - 0.01 ( a 1 = 0.01 ). The third quantile of α 1 in either ING or OUTG is 1, which is substantially smaller than 10.87. Therefore, there are saliently more observations with a high value of α 1 = 10.87 in the OUTG treatment than in the ING treatment, which reduces the group impact on α 1 to nil.

Distributions of other players’ beliefs after observing Player A’s action in the ING and OUTG treatments. Notes : The figures are based on other players’ actions upon observing either Player A’s action > 50 or < 50 in the ING and OUTG treatments over rounds 2-31. The left panel includes 874 observations for ING and 983 observations for OUTG; while the right panel includes 895 observations for ING and 988 observations for OUTG

As the median estimates are less sensitive to the approximate weight for extreme observations reporting a 1 = 0 or 100, we run a quantile (median) regression of α 1 on the same independent variables as in Table 4 , with standard error clustered at the session level, including extreme values. Column (1) in Table A2 presents the results. The coefficient of Ingroup is still positive and significant ( p -value = 0.012), indicating that there exists ingroup bias in social learning.

Finally, we investigate whether the identity effect in α 1 differs between the early and later rounds. With different values of n (e.g., n = 4 , 7 ), we find that the identity effect in the first n rounds is significantly larger than in the later rounds (see Table A4).

Then we try to examine whether group identity has an impact on α 2 . Figure 6 presents the cumulative probability distributions of α 2 of ingroup and outgroup, for confirming and contradicting private signals respectively. There are mild differences in their distributions, especially for the case with a contradicting signal. The Kolmogorov-Smirnov test results show that for confirming signals (left panel), the two distributions are not significantly different ( p -value = 0.649); while for contradicting signals (right panel), the difference between the two distributions is marginally significant ( p -value = 0.071). According to Table 3 , the difference in α 2 between the contradicting and confirming cases is the greatest in the ING treatment, which implies that group identity increases the asymmetry in belief updating.

CDFs of α 2 with groups. Notes : Observations satisfying - 1 ≤ α 2 ≤ 3 are included

We also run random effects regressions of α 2 on exogenous variables Ingroup , SL , IDM , and their interaction terms with Confirm , controlling for round dummies, the first-round weight on a drawn ball, and subjects’ characteristics, excluding extreme observations reporting a 2 ( a 1 ) = 0 or 100. 11 Table 5 displays the results. First, the coefficients of Confirm are significantly negative. The coefficients of C o n f i r m + S L × C o n f i r m ( p -value = 0.075 ) and C o n f i r m + I n g r o u p × C o n f i r m ( p -value < 0.001 ) are also significantly negative. These outcomes together suggest that there exists an asymmetry of belief updating that subjects put less weight on the private signal when it confirms the social signal than when it contradicts the social signal, regardless of group identity. In addition, the coefficients of C o n f i r m + I D M × C o n f i r m are not significant ( p -value ≈ 0.5 ), that is, the asymmetry in the belief updating pattern upon receiving two signals is unique to the contexts involving social learning.

Regressions of α 2 with groups

Random effects models are applied. Clustered standard error by session are in parentheses. ∗ p < 0.1 , ∗ ∗ p < 0.05 , ∗ ∗ ∗ p < 0.01

However, the more detailed patterns we find are different from those of De Filippis et al. ( 2022 ). Table 3 shows that subjects in the SL and GSL treatments saliently underweight their private signal when it confirms Player A’s action compared to the Bayesian belief updating, while they update their beliefs upon their private signal when it contradicts Player A’s action in a more or less Bayesian way. In other words, the asymmetry in our experiment comes from the fact that subjects underweight the private signal when it confirms the social signal instead of the fact found by De Filippis et al. ( 2022 ) that subjects overweight the private signal when it contradicts the social signal. 12

Second, the coefficients of I n g r o u p × C o n f i r m are negative and significant. Given that the coefficients of Confirm are negative, it shows that subjects in the same group as Player A exhibit a greater asymmetry in belief updating than those in the other group do. More specifically, the coefficients of Ingroup are positive and significant, indicating that subjects in the same group as Player A put more weight on the private signal when it contradicts the social signal than those in the other group do. We also test the coefficients of I n g r o u p + I n g r o u p × C o n f i r m and obtain significantly negative results ( p -value < 0.005). It means that group identity decreases the weight on the private signal when it confirms the social signal. As a consequence, the asymmetry in the belief updating pattern increases with group identity as mentioned above. H3 is not supported.

Result 2a: Subjects in the same group as Player A put more weight on the private signal when it contradicts Player A’s action than those in the other group do, that is, α 2 , I N G > α 2 , O U T G if s = 1 ( a A < 50 ) & a A ≠ 50 .

Result 2b: Subjects in the same group as Player A put less weight on the private signal when it confirms Player A’s action than those in the other group do, that is, α 2 , I N G < α 2 , O U T G if s = 1 ( a A > 50 ) & a A ≠ 50 .

We also run a quantile (median) regression α 2 on the same independent variables as in Table 5 . Column (2) in Table A2 displays the results. The signs and significance of the estimated coefficients are the same as in Table 5 , except for the coefficient of I n g r o u p × C o n f i r m which is negative but not significant. Therefore, we still find an asymmetric pattern of belief updating. We also find some evidence for confirmation bias in the IDM treatment since the coefficient of C o n f i r m + I D M × C o n f i r m is positive and marginally significant ( p -value = 0.089). We do not reproduce the group effect on the asymmetry of belief updating, but the significantly positive coefficient of Ingroup supports Result 2a .

Last, we are interested in whether group identity affects learning efficiency. In particular, we examine whether a player’s payoffs of her two beliefs differ between learning from and an outgroup Player A. We conduct regressions of the payoffs of the first belief and second belief respectively. Table 6 displays the results—we do not find any salient effect of group identity on learning efficiency. 13

Random effects regressions of the payoffs of the two beliefs

Clustered standard error by session are in parentheses. π 1 ( π 2 ) stands for the payoffs of the first (second) belief. ∗ p < 0.1 , ∗ ∗ p < 0.05 , ∗ ∗ ∗ p < 0.01

Group identity and the beliefs about Player A’s rationality

With the documented ingroup bias in forming the two beliefs, we further explore whether the impact of group identity is due to that subjects perceive a higher rationality of Player A in the same group than in the other group, or it is a result of an intrinsic heuristic to follow ingroup members.

Based on subjects’ answers to the questions in round 21, we generate new variables B A e r r o r , 1 and B A e r r o r , 2 to capture their belief that A made a mistake after observing the first and second signal respectively. As explained in Sect. 2.1 , in the experiment, making a mistake means that Player A reported a A > 50 upon observing a black ball or a A < 50 upon observing a white ball. And for the questions in round 21, subjects are requested to report the probability that they think Player A observed a white ball. Therefore, if a A > 50 , the number a subject reported actually means the probability that she thinks Player A did not make a mistake, that is, B Aerror = 100 - reported number. On the other hand, if a A < 50 , the probability that a subject thinks Player A observed a white ball is the same as the probability that she thinks Player A made a mistake, that is, B Aerror = reported number. For example, a subject reported that the probability that she thought Player A observed a white ball is 80% (20%) when she saw a A > 50 ( a A < 50 ). It means that her belief that Player A made a mistake is 20%, that is B Aerror = 20 . We use B A e r r o r , 1 and B A e r r o r , 2 as the proxies for subjects’ beliefs about Player A’s rationality, which are relevant and not confusing. Figure 7 displays the means and confidence intervals (at 95% level) of B A e r r o r , 1 (left panel) and B A e r r o r , 2 (right panel) for ingroup and outgroup. We can see that subjects’ beliefs that Player A made a mistake are even a bit higher when Player A is an ingroup than when she is an outgroup, but the differences are not significant.

The means and confidence intervals of B A e r r o r , 1 and B A e r r o r , 2 with groups

We run OLS regressions of α 1 in round 21 on B A e r r o r , 1 as well as Ingroup and SL . Table 7 shows the result. First of all, the impact of group identity on α 1 is confirmed. Also, the coefficients of B A e r r o r , 1 are significantly negative, that is, a subject puts less weight on Player A’s action if her belief that A has made a mistake is stronger. But controlling for B A e r r o r , 1 does not decrease the ingroup bias in α 1 . These results imply that the impact of group identity on processing the first social signal is probably not driven by the beliefs that ingroup members have higher information-processing abilities or make more rational decisions than outgroup members, rejecting H2 . In other words, even though we find that subjects put more weight on Player A’s action if they are in the same group than in different groups, there is little evidence to suggest that this observed ingroup bias in social learning is driven by people’s different beliefs about the predecessor’s rationality, that is, the belief-based channel. Combining the outcomes displayed in Table A4 that the group identity effects on α 1 in the first 4 and 7 rounds is significantly stronger than in the later rounds, we conjecture that our findings are more consistent with the perspective that group identity works as a framing device and its effect can be easily displaced by changing the framing of decisions.

OLS regressions of α 1 in round 21

Clustered standard error by session are in parentheses. ∗ p < 0.1 , ∗ ∗ p < 0.05 , ∗ ∗ ∗ p < 0.01

Result 3: The group identity effect on α 1 can not be explained by different beliefs between ingroup and outgroup Player A’s rationality.

We also run OLS regressions of α 2 on B A e r r o r , 2 . Table A3 in Online Appendix A shows that B A e r r o r , 2 has no significant effects on α 2 for round 21. And it does not exhibit the group identity effects on α 2 as summarized in Result 2a and Result 2b either. Therefore, it is infeasible for us to conclude that the group effects on α 2 are driven by the ingroup-outgroup differences in B A e r r o r , 2 or not. However, we do find that a contradicting private signal results in a significantly higher B A e r r o r , 2 than a confirming private signal. The p -values from either the Mann-Whitney U test or the t-test are smaller than 0.001 ; results are the same if we use the difference between B A e r r o r , 2 and B A e r r o r , 1 instead.

Conclusion and discussion

In this paper, we investigate whether and how group identity influences the way people process social information as well as subsequent private information. We adopt the design paradigm of De Filippis et al. ( 2022 ) and conjecture how group identity may have an impact via the belief and the belief-updating process regarding the predecessor’s rationality under the theoretical framework of the LRTU model, which is also developed by De Filippis et al. ( 2022 ).

Our results show that there exist ingroup-outgroup differences, not only in the processing of the social signal, but also seem to be in that of the subsequent private signal. That the weight on the social signal is higher if it is from an ingroup predecessor compared to if it is from an outgroup predecessor, is broadly in line with the finding of ingroup favoritism in the experimental literature. In contrast, the patterns of α 2 imply that the effect of such ingroup favoritism on posterior beliefs is at least partly offset by the ingroup-outgroup difference in the private signal phase. Specifically, when the private signal accords (contradicts) the social signal of an ingroup predecessor, people appear to place a lower (higher) weight on the private signal compared to when the predecessor is an outgroup.

As to the underlying mechanism, we do not find evidence supporting the belief-based channel we have conjectured under the framework of the LRTU model, because the ingroup-outgroup difference in the social signal phase can not be explained by the elicited beliefs in the predecessor’s rationality, and no identity effects exist for the rationality beliefs. This result seems to be at odds with results from other studies in the literature: for example, Cacault and Grieder ( 2019 ) documented an ingroup bias in beliefs about others’ intelligence. We conjecture that the diverging findings might be driven by how challenging the cognitive task is. In our experimental context, the rational response is very obvious given the signal valence. In contrast, the intelligence test questions studied by Cacault and Grieder ( 2019 ) were quite challenging. Better understanding the circumstances in which group identification distorts people’s beliefs about the ability of their peers is an interesting question for future research. On a separate note, it is also unlikely that social preferences drive the results. If individuals’ reported event beliefs were shaped by their preference to have similar monetary payoffs from the belief-elicitation task as their ingroup members, such a preference should remain in the private signal phase, which is at odds with the fact that the effect of the group identity on posterior event beliefs is partly offset after the private signal phase.

We consider our results as supporting the alternative perspective in the literature that group identity works as a framing device (Tajfel et al., 1971 ; Bacharach, 2006 ; Guala & Filippin, 2017 ; Filippin & Guala, 2017 ), as this perspective can reconcile the observed patterns in both the two phases. In the first phase, information about the group identity of the predecessor may work as a cue of the group frame that triggers certain heuristics. Specifically, since subjects observing an outgroup predecessor place lower weights on the social signal than subjects observing an ingroup, whose weights are in turn lower than those in the SL treatment, we conjecture that the heuristics triggered by the group frame prescribe caution and suspicion with others’ actions and opinions, especially with those of the outgroup. The fact that such an ingroup-outgroup difference becomes less prominent in later rounds is consistent with the decaying nature of group identity as a framing device. In the second phase, the group frame can be considered removed since the subjects now receive their own private signal. Subjects then process the social signal and private signal together, without the influence of the heuristics as in the first phase. Since we calculate the weight subjects put on the private signal based on the beliefs before and after observing the private signal with the Bayesian structure, it just appears as if the subjects put a higher (lower) weight on a confirming (contradicting) private signal if the predecessor is an outgroup than if the predecessor is an ingroup.

To better illustrate our reasoning, we provide the following examples. Suppose that Player A reports 30 after observing a black ball. Then as affected by the heuristic in the social signal phase, a Bayesian subject reports a 1 = 30 after observing Player A’s action when she is in the same group as Player A and reports a 1 = 40 when she is in the other group. 14 In other words, the subject underweights Player A’s action if they are not in the same group. In the private signal phase, if the subject is NOT affected by the group framing any more, she will update her belief upon a A = 30 and s = 1 after observing a white ball which contradicts Player A’s action, that is, the subject will report the stated belief a 2 = 50 regardless of group identity. But the calculation of α 2 will be based on a 1 = 40 when the subject is not in the same group as Player A. Then for the subject we obtain α 2 , c o n t r a d i c t , I N G = 1 > α 2 , c o n t r a d i c t , O U T G ≈ 0.47 , which is consistent with Result 2a .

Aagain, suppose that Player A reports 70 after observing a white ball. Then as a consequence of ingroup bias, a Bayesian subject reports a 1 = 70 after observing Player A’s action when she is in the same group as Player A and reports a 1 = 60 when she is in the other group. In the private signal phase, if the subject is NOT affected by the group framing, she will update her belief upon a A = 70 and s = 1 after observing a white ball which confirms Player A’s action, that is, the subject will report a 2 = 84.5 regardless of the predecessor’s group identity. But the calculation of α 2 will be based on a 1 = 60 when the subject is not in the same group as Player A. Then for the subject we obtain α 2 , c o n f i r m , I N G = 1 < α 2 , c o n f i r m , O U T G ≈ 1.52 . which is consistent with Result 2b .

Comparing our results to those from De Filippis et al. ( 2022 ), we replicate an asymmetric pattern in the private signal phase in the sense that a private signal contradicting the social signal is given a much higher weight than a confirming private signal. The difference between results from the two studies is how the weights compare to the Bayesian benchmark. If we are willing to assume that our subjects have a general tendency to underweight in comparison to the Bayesian benchmark in the private signal phase, then the specific case of the LRTU model discussed by De Filippis et al. ( 2022 ) can explain the asymmetric pattern in our data as well. We provide evidence supporting such an interpretation by showing that beliefs about the predecessor’s rationality significantly differ upon receiving a contradicting and a confirming private signal. In other words, the cognitive process captured by the LRTU model probably exists among our subjects; however, the effect of the group identity does not seem to work via shifting parameters in this process.

Looking forward, it would be interesting for future researchers to further explore the relationship between social learning and group identity, viewing the case of a neutral event (the color of the jar) and an artefactual minimal group studied in the current paper as a first step. In the real world, there are many interesting cases in which certain beliefs are non-neutral in the sense that people have intrinsic preferences over the beliefs, and the identity of certain natural groups often plays a role in determining such intrinsic values. For example, believing that vaccines are ineffective in protecting us from viruses may directly bring the person utility by enhancing her sense of belonging to the social group she identifies herself with. The existence of such group identity-related utility has been proven by Hett et al. ( 2020 ) as they showed that individuals were willing to pay significant amounts of money to belong to certain groups. Another example is the belief about others’ prosocial or antisocial behavior. People probably prefer to believe that those within her social group are more prosocial than those outside the group. In general, such intrinsic preferences of beliefs can lead to motivated beliefs and bias the belief-updating process (see Gino et al. 2016 for a review of the broad literature on motivated reasoning.) Taking into consideration the social learning aspect thereof, there could be richer patterns to explore both theoretically and empirically. For instance, observing social signals that are against the predecessor’s identity-relevant utilities might be especially powerful in shifting beliefs, and if so, it can be utilized by policy-makers to bring about socially desirable changes. Another example is when the choice of information source is endogenous. While Duffy et al. ( 2019 ) documented distinct private and social information choosers but an overall tendency for choosing social information, how the inclusion of group identity may reshape such choices will be something interesting for future research.

Below is the link to the electronic supplementary material.

Acknowledgements

We acknowledge the financial support from the National Natural Science Foundation of China (Grant No. 71903100; Grant No. 72003101), the Asia Research Center in Nankai University (Grant No. G02J0209), and the Laboratory for Economic Behavior and Policy Simulation in Nankai University (Grant No. ZX20220201; Grant No. ZX20220202 .The replication material for the study is available at http://doi.org/10.17605/OSF.IO/DR7AK .

As of August 2017, two-thirds (67%) of Americans report that they get at least some of their news on social media—with a fifth doing so often (Shearer & Gottfried, 2017 ).

Levy and Razin ( 2019 ) identify two broad reasons behind echo chambers: One is the tendency for individuals to segregate with like-minded ones, which leads to the creation of chambers; The other is the echo, that is, behavioral biases including correlation neglect, selection bias, and confirmation bias that induce polarization when individuals exchange beliefs in these chambers. Here we conjecture a new mechanism that can result in polarization of beliefs and opinions even if there is no segregation—if individuals put a greater weight on social information from ingroup than outgroup, and members of the same group receive correlated signals, then in some sense, such ingroup bias in social learning creates a virtual chamber. In other words, even if you are able to force people with opposite beliefs and opinions to desegregate and listen to each other, if there exists ingroup bias in social learning, their minds might still diverge further away.

A difference between our design and that of De Filippis et al. ( 2022 ) is that in our experiment, in each round, everyone except Player A observes one social signal. In contrast, in their experiment, there were some subjects who observed more than one social signal when making their own predictions in the social signal phase. According to the authors, data on these subjects’ behavior were not included in the analyses of the specific article; they were designed to answer other research questions.

Subjects’ responses to the predecessor’s decision may depend on their judgment about the predecessor’s rationality or intelligence. Online chat, which is a popular group enhancing device used in the literature, may provide information for them to make inferences about the the predecessor’s rationality or intelligence. Hence, we instead use common payoff to enhance group identity.

It is worth noting that the quadratic scoring rule is not incentive compatible under risk aversion, which might be the reason why subjects’ beliefs often peak at 50% in our results. There exist other methods that are incentive compatible for risk-averse agents in a theoretical sense such as the binarized scoring rule (Hossain & Okui, 2013 ). However, in state-of-the-art methodological discussions of belief elicitation, researchers emphasize the importance of so-called behavioral incentive compatibility and demonstrate that both the binarized scoring rule and the quadratic scoring rule violate some weak conditions for behavioral incentive compatibility (Danz et al., 2022 ). As a result, Danz et al. ( 2022 ) advocate for less-precise measurements with simpler and starker incentives for truthful reporting. In our elicitation procedure, after introducing the quadratic scoring rule, we explain to the subjects that it is in their best interest to state their true belief, to maximize the amount of money they could expect to gain. Such a descriptive representation of the scoring rule to some extent resonates with the advocate of Danz et al. ( 2022 )’s.

The social preference perspective might also predict an ingroup-outgroup difference when processing the social signal. Specifically, if people are more inequality averse toward an ingroup member than an outgroup individual, they will place a higher weight on the social signal to better align one’s own belief with the predecessor’s, reducing inequality in the monetary payoff from the belief-elicitation task. Alternatively, people might have an intrinsic preference to comply with (contradict) the behavior of an ingroup (outgroup). Either case, the social preference perspective predicts our H 1 but not H 2 or H 3 . While our results support H 1 but not H 2 or H 3 , there are other patterns in the private signal phase that are difficult to be reconciled from the social preference perspective. In general, we do not put much emphasis on this perspective in the current study, as we find it a bit stretchy while lacking specific support in the existing literature.

Subjects’ first beliefs in the IDM treatment have similar distributions to Fig. 1 except that a higher percentage of subjects report 50.

The distributions of players’ second beliefs for a A < 50 conditional on confirming and contradicting private signals are very similar to the case of a A > 50 .

The three quantiles of α 1 are not saliently different between the IDM and the SL treatment. In other words, subjects do not put more weight on the (first) private signal than on the (first) social signal, which indicates that people process the (first) social and private information in similar ways.

We also have within-subject and between-subject analyses over the impact of group identity on α 1 . Both analyses produce significantly positive results, which are displayed in Table A1 in Online Appendix A .

In Online Appendix B, we analyse these extreme values separately.

In fact, De Filippis et al. ( 2022 ) also show that subjects underweight their private signal when it confirms the social signal, but they do not discuss this outcome.

We also run regressions with interaction terms between treatment dummy variables and Confirm and find no significant impacts of group identity on learning efficiency either.

Here we take a Bayesian subject as an example for the simplicity of expression. The reasoning also works for non-Bayesian subjects.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Wenbo Zou, Email: [email protected].

Xue Xu, Email: [email protected].

• Anderson, L. R., & Holt, C. A. (1997). Information cascades in the laboratory. American Economic Review , 87 (5), 847–862.
• Ash, E., & Van Parys, J. (2015). Group identity bias and information cascades.
• Bacharach, M. (2006). Beyond Individual Choice: Teams and Frames in Game Theory . Princeton University Press.
• Banerjee AV. A simple model of herd behavior. The Quarterly Journal of Economics. 1992;107(3):797–817. doi: 10.2307/2118364. [ DOI ] [ Google Scholar ]
• Berger, S., Feldhaus, C., & Ockenfels, A. (2018). A shared identity promotes herding in an information cascade game. Journal of the Economic Science Association , 4 , 63–72.
• Bernhard H, Fehr E, Fischbacher U. Group affiliation and altruistic norm enforcement. American Economic Review. 2006;96(2):217–221. doi: 10.1257/000282806777212594. [ DOI ] [ Google Scholar ]
• Bettencourt A, Brewer M, Croak M, Miller N. Cooperation and the reduction of intergroup bias: the role of reward structure and social orientation. Journal of Experimental Social Psychology. 1992;28(4):301–319. doi: 10.1016/0022-1031(92)90048-O. [ DOI ] [ Google Scholar ]
• Bikhchandani S, Hirshleifer D, Welch I. A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy. 1992;100(5):992–1026. doi: 10.1086/261849. [ DOI ] [ Google Scholar ]
• Bolton GE, Feldhaus C, Ockenfels A. Social interaction promotes risk taking in a stag hunt game. German Economic Review. 2016;17(3):409–423. doi: 10.1111/geer.12095. [ DOI ] [ Google Scholar ]
• Brewer M. In-group bias in the minimal intergroup situation: A cognitive-motivational analysis. Psychological Bulletin. 1979;86(2):307. doi: 10.1037/0033-2909.86.2.307. [ DOI ] [ Google Scholar ]
• Brown R, Vivian J, Hewstone M. Changing attitudes through intergroup contact: The effects of group membership salience. European Journal of Social Psychology. 1999;29(5–6):741–764. doi: 10.1002/(SICI)1099-0992(199908/09)29:5/6&#x0003c;741::AID-EJSP972&#x0003e;3.0.CO;2-8. [ DOI ] [ Google Scholar ]
• Cacault MP, Grieder M. How group identification distorts beliefs. Journal of Economic Behavior & Organization. 2019;164:63–76. doi: 10.1016/j.jebo.2019.05.027. [ DOI ] [ Google Scholar ]
• Çelen B, Kariv S. Distinguishing informational cascades from herd behavior in the laboratory. American Economic Review. 2004;94(3):484–498. doi: 10.1257/0002828041464461. [ DOI ] [ Google Scholar ]
• Charness G, Rigotti L, Rustichini A. Individual behavior and group membership. American Economic Review. 2007;97(4):1340–1352. doi: 10.1257/aer.97.4.1340. [ DOI ] [ Google Scholar ]
• Chen DL, Schonger M, Wickens C. otree–an open-source platform for laboratory, online, and field experiments. Journal of Behavioral and Experimental Finance. 2016;9:88–97. doi: 10.1016/j.jbef.2015.12.001. [ DOI ] [ Google Scholar ]
• Chen Y, Li S. Group identity and social preferences. American Economic Review. 2009;99(1):431–57. doi: 10.1257/aer.99.1.431. [ DOI ] [ Google Scholar ]
• Currarini S, Mengel F. Identity, homophily and in-group bias. European Economic Review. 2016;90:40–55. doi: 10.1016/j.euroecorev.2016.02.015. [ DOI ] [ Google Scholar ]
• Danz, D., Vesterlund, L., & Wilson, A. J. (2022). Belief elicitation and behavioral incentive compatibility. American Economic Review, 112 (9), 2851–2883.
• De Filippis R, Guarino A, Jehiel P, Kitagawa T. Non-Bayesian updating in a social learning experiment. Journal of Economic Theory. 2022;199:105188. doi: 10.1016/j.jet.2021.105188. [ DOI ] [ Google Scholar ]
• Duffy J, Hopkins E, Kornienko T, Ma M. Information choice in a social learning experiment. Games and Economic Behavior. 2019;118:295–315. doi: 10.1016/j.geb.2019.06.008. [ DOI ] [ Google Scholar ]
• Filippin A, Guala F. Group identity as a social heuristic: An experiment with reaction times. Journal of Neuroscience, Psychology, and Economics. 2017;10(4):153. doi: 10.1037/npe0000078. [ DOI ] [ Google Scholar ]
• Gino F, Norton MI, Weber RA. Motivated bayesians: Feeling moral while acting egoistically. Journal of Economic Perspectives. 2016;30(3):189–212. doi: 10.1257/jep.30.3.189. [ DOI ] [ Google Scholar ]
• Gino F, Shang J, Croson R. The impact of information from similar or different advisors on judgment. Organizational Behavior and Human Decision Processes. 2009;108(2):287–302. doi: 10.1016/j.obhdp.2008.08.002. [ DOI ] [ Google Scholar ]
• Goeree JK, Palfrey TR, Rogers BW, McKelvey RD. Self-correcting information cascades. The Review of Economic Studies. 2007;74(3):733–762. doi: 10.1111/j.1467-937X.2007.00438.x. [ DOI ] [ Google Scholar ]
• Goette L, Huffman D, Meier S. The impact of group membership on cooperation and norm enforcement: Evidence using random assignment to real social groups. American Economic Review (Papers & Proceedings) 2006;96(2):212–216. doi: 10.1257/000282806777211658. [ DOI ] [ Google Scholar ]
• Grimm V, Utikal V, Valmasoni L. In-group favoritism and discrimination among multiple out-groups. Journal of Economic Behavior & Organization. 2017;143:254–271. doi: 10.1016/j.jebo.2017.08.015. [ DOI ] [ Google Scholar ]
• Guala F, Filippin A. The effect of group identity on distributive choice: Social preference or heuristic? The Economic Journal. 2017;127(602):1047–1068. doi: 10.1111/ecoj.12311. [ DOI ] [ Google Scholar ]
• Hett F, Mechtel M, Kröll M. The structure and behavioral effects of revealed social identity preferences. The Economic Journal. 2020;130(632):2569–2595. doi: 10.1093/ej/ueaa055. [ DOI ] [ Google Scholar ]
• Hossain T, Okui R. The binarized scoring rule. Review of Economic Studies. 2013;80(3):984–1001. doi: 10.1093/restud/rdt006. [ DOI ] [ Google Scholar ]
• Levine SS, Apfelbaum EP, Bernard M, Bartelt VL, Zajac EJ, Stark D. Ethnic diversity deflates price bubbles. Proceedings of the National Academy of Sciences. 2014;111(52):18524–18529. doi: 10.1073/pnas.1407301111. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ]
• Levy G, Razin R. Echo chambers and their effects on economic and political outcomes. Annual Review of Economics. 2019;11:303–328. doi: 10.1146/annurev-economics-080218-030343. [ DOI ] [ Google Scholar ]
• Li, S. (2020). Group identity, ingroup favoritism, and discrimination. In K. Zimmermann (Ed.), Handbook of Labor, Human Resources and Population Economics . Cham, Switz.: Springer. 10.1007/978-3-319-57365-6_123-1
• Müller D. The anatomy of distributional preferences with group identity. Journal of Economic Behavior & Organization. 2019;166:785–807. doi: 10.1016/j.jebo.2019.09.009. [ DOI ] [ Google Scholar ]
• Nöth M, Weber M. Information aggregation with random ordering: Cascades and overconfidence. The Economic Journal. 2003;113(484):166–189. doi: 10.1111/1468-0297.00091. [ DOI ] [ Google Scholar ]
• Ockenfels A, Werner P. Beliefs and ingroup favoritism. Journal of Economic Behavior & Organization. 2014;108:453–462. doi: 10.1016/j.jebo.2013.12.003. [ DOI ] [ Google Scholar ]
• Ortoleva P. Modeling the change of paradigm: Non-Bayesian reactions to unexpected news. American Economic Review. 2012;102(6):2410–36. doi: 10.1257/aer.102.6.2410. [ DOI ] [ Google Scholar ]
• Robson M. Inequality aversion, self-interest and social connectedness. Journal of Economic Behavior & Organization. 2021;183:744–772. doi: 10.1016/j.jebo.2020.12.029. [ DOI ] [ Google Scholar ]
• Ryen AH, Kahn A. Effects of intergroup orientation on group attitudes and proxemic behavior. Journal of Personality and Social Psychology. 1975;31(2):302. doi: 10.1037/h0076283. [ DOI ] [ Google Scholar ]
• Shearer E, Gottfried J. News use across social media platforms 2017. Journalism and Media: Pew Research Center; 2017. [ Google Scholar ]
• Tajfel H, Billig M, Bundy R, Flament C. Social categorization and intergroup behaviour. European Journal of Social Psychology. 1971;1(2):149–178. doi: 10.1002/ejsp.2420010202. [ DOI ] [ Google Scholar ]
• Tanaka T, Camerer CF. Trait perceptions influence economic out-group bias: lab and field evidence from Vietnam. Experimental Economics. 2016;19(3):513–534. doi: 10.1007/s10683-015-9452-1. [ DOI ] [ Google Scholar ]
• Weizsäcker G. Do we follow others when we should? a simple test of rational expectations. American Economic Review. 2010;100(5):2340–60. doi: 10.1257/aer.100.5.2340. [ DOI ] [ Google Scholar ]
• Xu X, Potters J, Suetens S. Cooperative versus competitive interactions and in-group bias. Journal of Economic Behavior and Organization. 2020;179:69–79. doi: 10.1016/j.jebo.2020.08.041. [ DOI ] [ Google Scholar ]
• Ziegelmeyer A, March C, Krügel S. Do we follow others when we should? a simple test of rational expectations: Comment. American Economic Review. 2013;103(6):2633–42. doi: 10.1257/aer.103.6.2633. [ DOI ] [ Google Scholar ]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

• View on publisher site
• PDF (2.2 MB)
• Collections

Similar articles

Cited by other articles, links to ncbi databases.

• Download .nbib .nbib
• Format: AMA APA MLA NLM

Add to Collections

Ingroup bias in a social learning experiment

• Original Paper
• Published: 28 December 2022
• Volume 26 , pages 27–54, ( 2023 )

Cite this article

• Wenbo Zou 1 , 3 &
• Xue Xu   ORCID: orcid.org/0000-0003-3577-8560 2 , 3  

4126 Accesses

1 Altmetric

Explore all metrics

Does social learning and subsequent private information processing differ depending on whether the observer shares the same group identity as the predecessor whose action is observed? In this paper, we conduct a lab experiment to answer this question, in which subjects first observe a social signal and then receive a private signal. We find that subjects put greater weights on the social signal if they share with the predecessor the same group identity that is induced in the experimental environment. We also provide suggestive evidence that such an ingroup-outgroup difference cannot be explained by individuals’ beliefs of the predecessor’s rationality. Moreover, heterogeneous effects of group identity exist in weights given to the subsequent private signal: Compared to when the predecessor is an outgroup, those who have learned from an ingroup predecessor put a greater (smaller) weight on the private signal if it contradicts (confirms) the social signal. We conjecture that such group effects are consistent with the perspective that group identity works as a framing device and brings about certain decision heuristics in the social signal phase, which no longer exist in the private signal phase.

Similar content being viewed by others

Overlap in processing advantages for minimal ingroups and the self

A shared identity promotes herding in an information cascade game

Social learning strategies regulate the wisdom and madness of interactive crowds

Avoid common mistakes on your manuscript.

1 Introduction

Social learning is an important source for individuals to gather information to update their beliefs and make decisions. It has drawn growing research interest from economists since the early work pioneered by, for example, Banerjee ( 1992 ), Bikhchandani et al. ( 1992 ), Anderson and Holt ( 1997 ). In a digital era like ours, in which social network sites play important roles in information dissemination, it’s easy to think of situations where individuals observe behavior and opinions of others before they receive and process objective signals themselves, regarding various types of events and facts, such as the safety and effectiveness of Covid-19 vaccines, and the validity of investment opportunities. Footnote 1

Does social learning differ whether people learn from others who share the same or have different social identities with them? For example, does a Facebook post from a fellow alumna have the same effect on shaping one’s own beliefs and choices as a Reddit post from a stranger? How about a twitter comment from someone belonging to the same or opposing political party? If an ingroup-outgroup difference exists in processing the social signal, then the next question is whether the impact persists, or perhaps the difference will be offset once the person receives objective signals herself. Answering these questions may provide novel insights for understanding phenomena such as chamber effect and polarization (see Levy and Razin 2019 for a review), which are among the major themes in the public arena of our time. Footnote 2

In this paper, we investigate these research questions in a controlled lab experiment, with group identities induced in the experimental environment. We also simplify away from the numerous factors in the real world by adopting the existing paradigm in the social learning literature—asking subjects to make incentivized judgments regarding a neutral event (color of the jar) that requires no previous knowledge and invokes no prior opinions or intrinsic preferences. We conjecture that the case of such a minimal group and a neutral event probably provides us a conservative estimate of the ingroup-outgroup difference in social learning, if it ever exists. On the one hand, natural group identities might have stronger effects. On the other hand, if, for example, the belief is ego-relevant or identity-relevant, the ingroup bias might interact with a motivated belief-updating process, resulting in a larger ingroup-outgroup difference when processing the social signals. Specifically, we follow the experimental procedure used by De Filippis et al. ( 2022 ): The color of an urn is either black or white—the two events are, a priori , equally likely; subjects are asked to state their beliefs about the likelihood of the urn being white, and the belief-elicitation task is incentivized. A randomly chosen subject receives a noisy binary symmetric signal about the color of the urn, after which she is asked to state her belief. Then other subjects observe the first subject’s stated belief and make the same type of prediction themselves. Finally, each subject receives another conditionally independent signal and makes a new prediction.

As we aim to investigate whether ingroup bias exists in social learning, besides the standard social learning (SL) treatment as in De Filippis et al. ( 2022 ), we design group social learning (GSL) sessions, in which before the social learning phase we add a phase of group assignment and enhancement following the previous experimental work on group identity (Tajfel et al., 1971 ; Chen & Li, 2009 ). Depending on whether sharing the same group identity as the first subject, subjects in the GSL sessions are divided into the ingroup social learning (ING) treatment and the outgroup social learning (OUTG) treatment. We repeat the learning procedure for 30 rounds in all treatments. In the first 20 rounds of the GSL sessions, we randomly select the first subject among members of one group, and then switch to the other group for the last 10 rounds, with the order counter-balanced—In this way, we embed a within-subject design in the between-subject design. Besides, we also design an individual decision-making (IDM) treatment as De Filippis et al. ( 2022 ) did, to provide a benchmark for comparison. Instead of observing the first subject’s prediction and then a private signal, all subjects receive two independent private signals consecutively in the IDM treatment.

We hypothesize that when processing the social signal, an ingroup-outgroup difference exists. Our experimental results confirm this hypothesis, as we find that based on weights computed with the Bayesian updating structure, subjects in the OUTG treatment underweight the social signal compared to the Bayesian benchmark to a significantly greater extent than those in the ING treatment, whereas subjects in SL or IDM treatments are more close to Bayesian updating. Moreover, in round 21 of the experiment, we elicit subjects’ beliefs on the first subject’s rationality , that is, the likelihood that the first subject has reported beliefs regarding the color of the urn in the correct direction as the signal suggests. We find no statistically significant ingroup-outgroup differences in these elicited beliefs on the predecessor’s rationality; neither can these rationality beliefs explain the impact of the group identity on the weight placed on the social signal. By checking whether and how the ingroup-outgroup difference varies over the rounds, we find that it is more prominent in the first several rounds than in the later rounds. These results suggest that instead of working via a belief-updating process about the predecessor’s rationality, the ingroup-outgroup difference might reflect heuristics that prescribe caution and suspicion with others’ behavior, especially with the outgroup predecessors’, which are brought about by the group frame.

We further examine whether sharing the same group identity as the predecessor has an impact on the weights put on the subsequent private signal. We find that when the private signal contradicts (confirms) the social signal, subjects in the ING treatment put a greater (lower) weight on the private signal than those in the OUTG treatment. Since the weight for the private signal is calculated with the Bayesian updating structure using the stated belief after observing the social signal as the prior belief, we conjecture that these empirical patterns are, again, consistent with that group identity works as a framing device, the heuristics cued by which no longer exist in the private signal phase. In other words, because the ingroup-outgroup difference in processing the social signal no longer exists in the private signal phase when subjects process the social and private signals together, subjects behave as if they were putting a higher (lower) weight on the private signal if it contradicts (confirms) the social signal, given the way the weights are calculated in our analyses.

Finally, since our study adopts an experimental design similar to that used by De Filippis et al. ( 2022 ), we can check whether our results replicate theirs. They found that when processing the private signal, subjects weighed the signal as a Bayesian agent if the signal confirmed the social signal; subjects overweighted the signal if it contradicted the social signal. In contrast, in our SL and GSL treatments, after receiving a contradicting private signal, subjects weigh the signal similar to a Bayesian agent; subjects underweight the signal when it confirms the social signal. At first glance, the results are quite different between the two studies. However, the commonality lies in that subjects nevertheless put a higher weight on a contradicting private signal than on a confirming one. The intuition captured by the specific case of the Likelihood Ratio Test Updating (LRTU) model discussed by De Filippis et al. ( 2022 ), which reconciles the empirical patterns they found, can be used to explain our results as well, as long as we are willing to impose an ad hoc assumption that in our private signal phase, the updating benchmark is not Bayesian updating but some extent of underweighting. Intuitively, subjects hold confident prior beliefs about the predecessor’s rationality after observing the social signal. Then, a confirming private signal retains such confident beliefs; subjects update their event beliefs under a general tendency of underweighting. In contrast, a contradicting private signal makes one adjust downward her belief about the predecessor’s rationality, and thus she adjusts downward the weight she puts on the social signal. Because such an adjustment is not captured by our way of calculating the weight placed on the private signal, it will appear as if the subject puts more weight on the private signal—the greater weight then offsets the general tendency to underweight among our subjects, rendering the belief-updating consistent with Bayesian updating. We provide evidence supporting this interpretation by showing that the rationality belief we elicit in round 21 becomes significantly lower after observing a contradicting private signal than after observing a confirming one.

Our work contributes to the large and growing literature on the various aspects of group identity effects in experimental economics. Given the significant impact of group identity on lots of important outcomes such as distributive preferences, cooperation behavior, evaluations of task performances, and general attitudes (for example, Tajfel et al. 1971 ; Ryen and Kahn 1975 ; Brewer 1979 ; Bettencourt et al. 1992 ; Brown et al. 1999 ; Bernhard et al. 2006 ; Goette et al. 2006 ; Chen and Li 2009 ; Currarini and Mengel 2016 ; Xu et al. 2020 ; Li 2020 ,) it is reasonable to conjecture that social learning from an ingroup member could be different compared to that from an outgroup member. So far, research directly examining the impact of group identity on social learning is limited. To the best of our knowledge, only in two recent experimental studies, Ash and Van Parys ( 2015 ) and Berger et al. ( 2018 ) investigated how group identity affects the formation of information cascade. They both adopted the minimal group paradigm similar to the one used in the current paper to assign subjects to two groups (Tajfel et al., 1971 ; Chen & Li, 2009 ). After the group assignment stage, subjects engaged in repeated information cascade games based on Anderson and Holt ( 1997 ) in which each of them in a sequence knew all the prior players’ choices of the true state as well as her own private signal, and the group identities of the prior players were revealed. They found that information cascades are more likely to occur when learning from ingroup members compared to when learning from outgroup individuals. There are two major differences between the present study and theirs. First, we let subjects report their beliefs about the true state instead of making a binary decision regarding on which state to bet, which allows us to compute and compare the weights they put on the signals in different situations. Second, we collect subjects’ beliefs two times, one time after they observe a predecessor’s belief and the other time after they receive their own private signal, with which we can distinguish between their belief updating upon learning the two types of information. In contrast, Ash and Van Parys ( 2015 ) and Berger et al. ( 2018 ) let subjects make a decision after knowing the social and private information at the same time, focusing on the formation of information cascades. To summarize, we investigate the effect of group identity on social learning in a more general sense that is not constrained to the context of information cascade, and we also examine the implications beyond the social learning phase per se , that is, when new private information is available.

Our results also speak to an on-going debate in the broader literature on how to explain the identity effect in shaping behavior and cognition. First, the social preferences perspective has been popular among experimental economists, who represent the impact of group identity as changes in the values of prosocial parameters (Chen & Li, 2009 ; Müller, 2019 ; Robson, 2021 ). Second, some other researchers have focused on the role of beliefs. For example, Ockenfels and Werner ( 2014 ) examined the effect of second-order belief of group identity on ingroup favoritism in dictator games. Tanaka and Camerer ( 2016 ) showed that differences in outgroup discrimination can be explained by beliefs about the characteristics of potential outgroups. Grimm et al. ( 2017 ) found that beliefs about the behavior of other groups’ extent of ingroup favoritism matter. Third, an alternative tradition exists, viewing group identity as a framing device (Tajfel et al., 1971 ; Bacharach, 2006 ). Recent psychological research suggests that more similarity or experiences of more aligned interests among subjects may lead to a more “trusting mindset” (Gino et al., 2009 ; Levine et al., 2014 ; Bolton et al., 2016 ). More recently, Guala and Filippin ( 2017 ) and Filippin and Guala ( 2017 ) provided experimental evidence supporting the heuristics interpretation, by showing that the effect of group identity on distributive choices is highly dependent on contextual factors and that it can be easily displaced by changing the framing of decisions. As argued earlier, our findings can be viewed as supporting the framing perspective.

Last, this paper relates to the broader literature on other types of cognitive bias in information processing. For example, existing experimental evidence suggests that people tend to overweight their private information and underweight the social information (Nöth & Weber, 2003 ; Çelen & Kariv, 2004 ; Goeree et al., 2007 ; Weizsäcker, 2010 ; Ziegelmeyer et al., 2013 ). It is then an open question how the ingroup-outgroup difference, if existing, interacts with such a tendency to underweight social information. In our data, we do not find evidence supporting that people in general underweight social information compared to private information. Interestingly, we find that introducing the group frame makes people underweight the social signal, especially if the predecessor is an outgroup.

The remaining of the paper structures as follows: We introduce the experimental design in Sect. 2 , report the results in Sect. 3 , and then conclude as well as provide our discussions of the results and future research directions in Sect. 4 .

2 The experiment

The experiment consists of four treatments, SL, IDM, and GSL (including ING and OUTG). Our SL and IDM treatments are similar to those of De Filippis et al. ( 2022 ). Footnote 3 In the GSL treatments, we assign subjects to two groups and they observe the choices made by either an ingroup or an outgroup predecessor before making their own decisions. Then each subject receives an independent private signal. Details of experiments are as follows.

There are two parts in all treatments. In part 1 , subjects see paintings in pairs on the screen. One of the paintings is drawn by Paul Klee, and the other is drawn by Wassily Kandinsky. The artists’ names do not appear on the screen, and the positions of the paintings are randomly decided, that is, it could either be Klee on the left and Kandinsky on the right, or the opposite. Subjects are requested to indicate which one they prefer in each pair of paintings. There are 5 pairs in total. Then, according to their relative preferences over the two artists, subjects in the GSL treatments are assigned to the Klee group or the Kandinsky group. We use relative preferences instead of absolute preferences over the two artists for group assignment since we want to create two groups with the same or close number of members to control for the impact of group size within sessions. There were 14–22 participants in a GSL session. If the number of participants n is even, then both group sizes are n /2; otherwise, one group size is \((n-1)/2\) and the other is \((n+1)/2\) . There is no group assignment in the IDM and SL treatments, while subjects in these two treatments nevertheless make choice between the paintings.

Subjects in all treatments then perform a task of moving sliders to the middle points of slider bars. In the IDM and SL treatments a subject earns 5 points for every slider she moves to the middle point. In the GSL treatments a subject earns 0.5 points for every slider either she or any other group member moves to the middle point. Payoffs of the first part are not revealed until the end of the experiment. We introduce such common payoffs to enhance subjects’ perception of group identity (Charness et al., 2007 ). Footnote 4

Part 2 lasts for 31 rounds. In the beginning of each round, the computer randomly chooses one from two (virtual) urns, a white urn and a black urn, with a 50–50 chance. The white urn contains 7 white balls and 3 black balls, while the black urn contains 3 white balls and 7 black balls. Subjects won’t know which urn has been chosen until the end of each round. In round 1 of all treatments, the computer randomly draws a ball out of the chosen urn with replacement for each subject, and subjects observe whether it is a white ball or a black ball. Then, subjects report individually the probability that they believe the chosen urn is white in this round, ranging from 0 to 100 (up to two decimal points). The first round serves as a chance for practice.

In each of rounds 2–31 in the IDM treatment, subjects first observe a ball privately drawn from the chosen urn and report individually their subjective probability beliefs about the chosen urn being white. This process repeats twice in a round. In each of rounds 2–31 in the SL and GSL treatments, the computer first randomly chooses one subject as Player A and then randomly draws a ball out of the chosen urn with replacement. Only Player A observes the drawn ball and reports her subjective probability belief about the chosen urn. After Player A’decision, every other subject observes the subjective probability Player A reported ( social signal ), and reports individually his own belief about how likely the chosen urn is white. Last, each subject privately observes another randomly drawn ball ( private signal ) and reports his belief about the chosen urn again.

In the GSL treatments, the group identity of Player A is revealed. We introduce the ING and OUTG treatments within the same session, which means that Player A sends social signals to both groups at the same time. Such a design feature guarantees that any group identity effect we find in information processing is not because subjects expect Player A to act differently when sending social signals to ingroup versus outgroup. In particular, in a Klee-Kandinsky (Kandinsky-Klee) session, one subject is chosen randomly from the Klee (Kandinsky) group in each of rounds 2–21, and from the Kandinsky (Klee) group in each of rounds 22–31. Therefore, those in the Klee (Kandinsky) group in a Klee-Kandinsky session are in the ING (OUTG) treatment in rounds 2–21, but are in the OUTG (ING) treatment in rounds 22–31.

Subjects’ reported beliefs are incentivized with a quadratic scoring rule, that is, payoff \(= 150-0.015\times (V-\) reported belief \()^2\) . We adopt the quadratic scoring rule following De Filippis et al. ( 2022 ). Footnote 5 If the chosen urn is white, \(V=100\) ; otherwise, \(V = 0\) . In rounds 2–31, the computer randomly draws one out of two reported beliefs to generate the payoff of the corresponding round. After round 31 ends, the computer randomly selects one round from rounds 1–10, 11–21, and 22–31 respectively. The sum of subjects’ earnings of the three selected rounds makes up the main part of their payoffs in the second part.

Moreover, we design two questions to elicit subjects’ beliefs about Player A’s rationality in round 21 of the SL and GSL treatments. Since beliefs about Player A’s rationality serve our secondary research objectives, we elicit them only in round 21, leaving the other rounds succinct. We ask subjects to report the probability they believe that Player A saw a white ball in the current round. The reported numbers also range from 0 to 100. They are asked twice, once after observing Player A’s action, and the other time after receiving the private signal. The answer is incentivized; the second-time question is phrased as a chance to modify the answer reported for the first time. \(\text{Payoff}= 50-0.005\times (C-\text{reported number})^2\) . If Player A did see a white ball, \(C=100\) ; otherwise, \(C = 0\) . Subjects’ final earnings include the payoff from this extra question in round 21. Together with Player A’s action, such probability beliefs provide us with a measure of probability beliefs regarding Player A’s rationality. Specifically, if Player A’s action is larger than 50, which suggests a white ball, then the reported probability belief in round 21 that Player A has seen a white ball is equivalent to the probability belief that Player A has not made a mistake, or, that Player A is rational. If, instead, Player A’s action is smaller than 50, the probability that a subject thinks Player A has observed a white ball is the same as her probability belief that Player A has made a mistake. We use this simple and neutral belief question instead of directly asking about how likely subjects think Player A is or is not rational, or how likely that Player A has or has not made a mistake, to avoid potential confusion and minimize the influence of language connotation.

2.2 Hypotheses

First, we aim at testing whether there is a direct impact of group identity on subjects’ social learning. We focus on \(\alpha _1\) , the weight that individuals put on a social signal calculated from the following formula:

where \(a_1\) is the first belief, \(a_A\) denotes Player A’s action, and q is the signal precision, which equals 0.7 in our experiment. If \(\alpha _1 = 1\) , it means that the subject updates his belief in a Bayesian way; if \(\alpha _1\) is larger (smaller) than 1, it means that he updates his belief as if a Bayesian belief updater has observed the signal more (fewer) than once.

In the existing experimental studies, Ash and Van Parys ( 2015 ) and Berger et al. ( 2018 ) found that information cascades were more likely to occur when learning from ingroup members than learning from outgroup members. Here, we provide our belief-based reasoning of an ingroup-outgroup difference in social learning under the Likelihood Ratio Test Updating (LRTU) rule model, a non-Bayesian updating model developed and empirically verified by De Filippis et al. ( 2022 ). Below, we describe the intuition; interested readers can find more formal analyses in Online Appendix C.1 .

According to the LRTU model, when an individual observes Player A’s action, she is initially ambiguous about Player A’s rationality, that is, whether Player A correctly takes an action \(a_A>50\) ( \(a_A<50\) ) if she observes a white (black) ball. She first selects her prior belief for Player A’s rationality, and then she updates her evaluation regarding the color of the urn. Specifically, she sets a critical value \(c\) and compares it to the likelihood ratio of observing a (set of) signal(s) conditional on the predecessor being rational. If the likelihood ratio is greater than or equal to \(c\) , then the individual will select the prior belief that is most confident about the predecessor’s rationality within the possible belief range. Otherwise, she will select the least confident prior belief. The value of \(c\) indicates how strong (or weak) the evidence needs to be to support the confidence in the predecessor’s rationality. When only one social signal is available, as in the first phase in each round of our experiment, the likelihood ratio of observing any given signal, conditional on the predecessor being rational, equals one. Let us discuss how group identity can make a difference in the three representative cases as follows.

Suppose that (a) the value of \(c\) differs depending on whether the predecessor is an ingroup or an outgroup; for the ingroup \(c<1\) , whereas for the outgroup \(c>1\) ; (b) the range of the rationality belief is identical for an ingroup and an outgroup. This means that it requires weaker evidence for the individual to have confidence about an ingroup’s rationality than about an outgroup’s rationality. Then, with the likelihood ratio equal one, the individual will select the most confident prior belief for the ingroup’s rationality while she will select the least confident prior belief for the outgroup’s rationality. Therefore, the individual will have a more confident belief about the ingroup’s rationality than the outgroup’s, and as a result, the weight calculated with the Bayesian updating structure, \(\alpha _1\) , will be higher if the predecessor is an ingroup than if he is an outgroup.

Suppose that (a) \(c\) has no ingroup-outgroup difference; \(c\le 1\) ; (b) the most confident prior belief within the belief range is more confident for the ingroup than for the ougroup. In this case, the individual will select the most confident rationality belief within the respective range for both the ingroup and the outgroup. Because of the difference in the range, the same conclusion can be drawn as in the first case— \(\alpha _1\) would be higher if the predecessor is an ingroup than if he is an outgroup.

Suppose that (a) \(c\) has no ingroup-outgroup differences; \(c>1\) ; (b) the least confident prior belief within the belief range is more confident for the ingroup than for the ougroup. In this case, the individual will select the least confident rationality belief within the respective range for both the ingroup and the outgroup. Again, because of the difference in the range, the same conclusion can be drawn as in the first case— \(\alpha _1\) would be higher if the predecessor is an ingroup than if he is an outgroup.

In these cases above, the assumption that weaker evidence is required to have confidence in an ingroup’s rationality than that of an outgroup, as well as the assumption that either the most confident or least confident belief about the predecessor’s rationality is higher for an ingroup than for an outgroup, can find support in a broad sense in the existing literature. For example, Cacault and Grieder ( 2019 ) found that people are overconfident in their fellow group members’ intelligence. Summing it up, we propose the following hypotheses: Footnote 6

Subjects in the same group as Player A put a greater weight on the social signal than those in the other group do, that is,

\(\alpha _{1,ING} > \alpha _{1,OUTG}\) .

Subjects in the same group as Player A have a higher belief regarding Player A being rational than those in the other group do. Once controlling for the rationality beliefs, the ingroup-outgroup differences in \(\alpha _1\) , if any, disappears, or at least, diminishes.

Moreover, we are interested in whether group identity has any impact on subsequent processing of private information after the social learning phase. We compute \(\alpha _2\) , the weight that an individual puts on his private signal using the following formula:

where \(a_2\) is the second belief, and \(s = 1\ (0)\) means that a white (black) ball was drawn. Similar to \(\alpha _1\) , when \(\alpha _2 = 1\) , it means that the subject updates his belief in a Bayesian way; \(\alpha _2\) deviating from 1 indicates over- or under-updating in comparison to Bayesian updating.

De Filippis et al. ( 2022 ) found an asymmetry of belief updating upon receiving a private signal after a social signal, that is, people overweighted their private signal when it contradicted the social signal while they updated their belief in a Bayesian way when their private signal confirmed the social signal. The authors reconcile such patterns in their results with the Likelihood Ratio Test Updating (LRTU) rule model assuming that the parameter values satisfy certain conditions. Suppose that when receiving the social signal, the individual’s value of \(c\) is smaller than one, and thus the individual selects the most confident prior belief regarding the predecessor’s rationality. Then, a private signal confirming the social signal makes the individual stick to the same prior belief on the predecessor’s rationality, because a confirming private signal only increases the likelihood ratio. In contrast, a contradicting private signal will result in a much lower likelihood ratio, which can be smaller than \(c\) , making the individual select the least confident prior belief of the predecessor being rational within the range. As a result, the belief-updating of the events in both phases will look as if the individual overweights the private signal, because it is in the opposite direction of the social signal, the weight placed on which is now adjusted down because of lowered beliefs in the predecessor’s rationality.

Following the LRTU model while assuming that \(c<1\) regardless of the group identity of Player A, we generate hypotheses about the effect of group identity on the weight subjects put on their private signal in exemplary cases as follows. In general, when the private signal is confirming, the likelihood ratio becomes even larger—the belief about Player A’s rationality does not shift; it is then unlikely that the group identity exhibits any effect on the processing of the private signal. When the private signal contradicts the social signal, we discuss two cases below.

Suppose that (a) \(c\) is smaller for the ingroup than for the outgroup; (b) the range of the rationality belief is identical for an ingroup and an outgroup. In this case, it is possible that receiving the private signal induces a downward adjustment in beliefs of Player A’s rationality for the outgroup but not for the ingroup, if the updated likelihood ratio lies inbetween the two values of \(c\) . Thus, we expect to observe an ingroup-outgroup difference in \(\alpha _2\) after observing a contradicting private signal, since a more dramatic downward adjustment of Player A’s rationality will demonstrate itself in the form of a greater weight placed on the contradicting private signal, calculated with the Bayesian structure.

Suppose that (a) \(c\) and the most confident prior belief within the belief range are both the same; (b) the least confident prior belief is more confident for the ingroup than for the outgroup. In this case, it is possible that both for the ingroup and the outgroup, after observing a contradicting private signal, an adjustment of Player A’s rationality takes place and the individual adjusts the prior belief to the least confident one. As an ingroup-outgroup difference exists in the least confident prior belief, a group effect on \(\alpha _2\) as described in the first case will also be observed here.

Subjects in the same group as Player A put less (the same) weight on their private signal when it contradicts (confirms) Player A’s action than those in the other group do, that is,

\(\alpha _{2,ING} < \alpha _{2,OUTG}\) if \(s = \mathbbm {1}(a_A<50)\) & \(a_A \ne 50\)

\(\alpha _{2,ING} = \alpha _{2,OUTG}\) if \(s=\mathbbm {1}(a_A>50)\) & \(a_A \ne 50\)

2.3 Procedures

The experiment was conducted in October and November, 2018 at Nankai University, at the China Centre for Experimental Social Sciences (CESS), which is a collaborative laboratory between Nankai University and the Nuffield College at University of Oxford. The subject pool of CESS China consists of undergraduate and graduate students in all disciplines at Nankai University. Experimental sessions were advertised within the subject pool of CESS China, and participants were recruited through a Wechat recruiting system (ancademy.org). Upon arrival, subjects were assigned to computers by randomly choosing one card from a pile of numbered cards. After subjects were seated in the lab, instructions of part 1 were distributed. The experimenter read instructions aloud in front of all subjects and then subjects answered understanding questions. After they answered all understanding questions correctly, the experiment started. Instructions of part 2 were distributed after part 1 ended. When the experiment was completed, subjects saw their final earnings on the screen privately and got paid on site. The experiment was computerized using oTree (Chen et al., 2016 ).

In total, 14 sessions were organized and 258 subjects participated in the experiment. On average each session lasted for one hour. Subjects earned 59.4 RMB yuan on average. The numbers of subjects and sessions as well as the means and standard deviations (in parentheses) of subjects’ characteristics are displayed by treatment in Table 1 .

3.1 Descriptive analysis and data preparation

Figure 1 shows that the distribution of Player A’s action after observing a white (black) ball has a peak around 70 (30), which is consistent with predictions of Bayes’ theorem. Also, a few Player As choose to report 50, which may be because of strong risk aversion. Footnote 7

Distribution of Player A’s actions. Notes : The figures are based on Player As’ actions in the 12 sessions of the SL and GSL treatments over rounds 2-31, that is 360 observations in total. The left and right panels each include 180 observations

Figure 2 displays other players’ beliefs conditional on Player A’s action. If Player A’s action is larger (smaller) than 50, other players should consider that a white (black) ball was drawn. We do find that the distribution of first beliefs has a peak around 70 (30) when Player A’s action is larger (smaller) than 50. Compared to Fig. 1 , a slightly higher percentage of subjects report 50.

Distribution of other players’ beliefs after observing Player A’s action. Notes : The figures are based on other players’ actions upon observing either Player A’s action \(>50\) or \(<50\) in the SL and GSL treatments over rounds 2-31. The left panel includes 2913 observations and the right panel includes 2648 observations

Figure 3 presents the distributions of players’ second beliefs, that is after observing a private signal, conditional on that the private signal confirms Player A’s action \(a_A>50\) (a white ball was drawn) and contradicts \(a_A>50\) (a black ball was drawn) respectively. Footnote 8 Given that most subjects report 70 as their first beliefs for \(a_A>50\) , if a majority of subjects update their second beliefs in a Bayesian way, we would observe that the distribution of their second beliefs for a confirming (contradicting) private signal had a peak around 84.5 (50). The left panel shows that quite a few beliefs deviate saliently from 84.5 and the mean of the second beliefs is a bit lower than 80, which does not seem to support Bayesian belief updating. The right panel confirms that most people report 50 for a contradicting private signal, which is different from a very asymmetric distribution around 50 shown by De Filippis et al. ( 2022 ).

Distribution of other players’ second beliefs. Notes : The figures are based on other players’ actions upon observing either confirming (a white ball) or contradicting signals (a black ball) when \(a_1>50\) , in the SL and GSL treatments over rounds 2-31. The left panel includes 1702 observations and the right panel includes 1211 observations

In Table 2 , we present the three quantiles of \(\alpha _1\) computed using Eq. ( 1 ) by treatment. Note that when \(a_1=0\) or 100, the equation generates positive or negative infinity for \(\alpha _1\) respectively. We compute \(\alpha _1\) by approximating \(a_1=0\) with 0.01 and \(a_1=100\) with \(100-0.01\) . Table 2 shows that subjects display heterogeneity in belief updating upon observing the first signal. For example, some do not update at all ( \(\alpha _1=0\) ) while others update the same as a Bayesian agent would do ( \(\alpha _1=1\) ). We can see that median subjects in the SL and GSL treatments tend to underweight Player A’s action (the social signal) compared to a Bayesian agent ( \(\alpha _1<1\) ). Additionally, subjects in the OUTG treatment attach a lower weight to the social signal compared to the ING treatment. Footnote 9

Table 3 displays the quantiles of \(\alpha _2\) computed using Eq. ( 2 ), upon observing confirming and contradicting signals respectively, and by treatment. When \(a_2=0\) or 100, or \(a_1 = 0\) or 100, the equation generates positive or negative infinity for \(\alpha _2\) . We compute \(\alpha _2\) by approximating \(a_2\ (a_1)=0\) with 0.01 and \(a_2\ (a_1)=100\) with \(100-0.01\) .

As shown in Table 3 , in the SL treatment, the three quantiles of \(\alpha _2\) are all smaller upon observing a confirming signal than upon observing a contradicting signal. The median subject updates beliefs in a Bayesian way upon receiving a contradicting signal ( \(\alpha _2=1\) ), while she attaches less weight to the second signal compared to a Bayesian agent when observing a confirming signal ( \(\alpha _2=0.64\) ). The asymmetry of belief updating also holds for the ING and OUTG treatments, with even lower values of \(\alpha _2\) for observing a confirming private signal. We also compute \(\alpha _2\) for the IDM treatment, where confirming signals mean that two white or two black balls were drawn and contradicting signals mean that a white and a black ball were drawn. We find an asymmetry of belief updating upon receiving two signals in the IDM treatment but in the opposite direction, that is, subjects put more weight on the confirming signal than on the contradicting signal.

3.2 Group identity and belief updating

The focus of the present study is to investigate whether group identity has an impact on social learning and subsequent private belief updating. First, we look at whether group identity has an impact on \(\alpha _1\) . Figure 4 displays the cumulative probability distributions of \(\alpha _1\) for ingroup and outgroup. There exists a mild but significant difference between the two distributions with the Kolmogorov-Smirnov test ( p -value = 0.03). Table 2 also shows that the median value of \(\alpha _1\) is lower in the OUTG treatment than in the ING treatment.

We then include the observations from 2-31 rounds in all treatments and run random effects regressions of \(\alpha _1\) on exogenous variables Ingroup , SL , and IDM , controlling for round dummies, the first-round weight on a drawn ball, and subjects’ characteristics, excluding the extreme observations reporting \(a_1=0\) or 100. Table 4 displays the outcomes. We find that the coefficients of Ingroup are statistically significant, which support our H1 . Footnote 10

Result 1: Subjects put more weight on Player A’s action if they are in the same group than in different groups, that is, \(\alpha _{1,ING} > \alpha _{1,OUTG}\) .

CDFs of \(\alpha _1\) with groups. Notes : Observations satisfying \(-1\le \alpha _1\le 3\) are included

It should be noted that the group impact on \(\alpha _1\) becomes insignificant when including extreme values with \(a_1=0\) or 100 in the linear random effects regressions ( p -value = 0.272). Figure 5 shows that the frequency of \(a_1 = 100\) ( \(a_1 = 0\) ) in the OUTG treatment is roughly double that in the ING treatment, given that Player A’s action is larger (smaller) than 50. An extreme value of \(a_1\) consistent with Player A’s action generates \(\alpha _1 = 10.87\) with the approximation \(a_1 = 100 - 0.01\) ( \(a_1 = 0.01\) ). The third quantile of \(\alpha _1\) in either ING or OUTG is 1, which is substantially smaller than 10.87. Therefore, there are saliently more observations with a high value of \(\alpha _1 = 10.87\) in the OUTG treatment than in the ING treatment, which reduces the group impact on \(\alpha _1\) to nil.

As the median estimates are less sensitive to the approximate weight for extreme observations reporting \(a_1=0\) or 100, we run a quantile (median) regression of \(\alpha _1\) on the same independent variables as in Table 4 , with standard error clustered at the session level, including extreme values. Column (1) in Table A2 presents the results. The coefficient of Ingroup is still positive and significant ( p -value = 0.012), indicating that there exists ingroup bias in social learning.

Distributions of other players’ beliefs after observing Player A’s action in the ING and OUTG treatments. Notes : The figures are based on other players’ actions upon observing either Player A’s action \(>50\) or \(<50\) in the ING and OUTG treatments over rounds 2-31. The left panel includes 874 observations for ING and 983 observations for OUTG; while the right panel includes 895 observations for ING and 988 observations for OUTG

Finally, we investigate whether the identity effect in \(\alpha _1\) differs between the early and later rounds. With different values of \(n\) (e.g., \(n=4, 7\) ), we find that the identity effect in the first \(n\) rounds is significantly larger than in the later rounds (see Table A4).

Then we try to examine whether group identity has an impact on \(\alpha _2\) . Figure 6 presents the cumulative probability distributions of \(\alpha _2\) of ingroup and outgroup, for confirming and contradicting private signals respectively. There are mild differences in their distributions, especially for the case with a contradicting signal. The Kolmogorov-Smirnov test results show that for confirming signals (left panel), the two distributions are not significantly different ( p -value = 0.649); while for contradicting signals (right panel), the difference between the two distributions is marginally significant ( p -value = 0.071). According to Table 3 , the difference in \(\alpha _2\) between the contradicting and confirming cases is the greatest in the ING treatment, which implies that group identity increases the asymmetry in belief updating.

We also run random effects regressions of \(\alpha _2\) on exogenous variables Ingroup , SL , IDM , and their interaction terms with Confirm , controlling for round dummies, the first-round weight on a drawn ball, and subjects’ characteristics, excluding extreme observations reporting \(a_2\ (a_1)=0\) or 100. Footnote 11 Table 5 displays the results. First, the coefficients of Confirm are significantly negative. The coefficients of \(Confirm + SL \times Confirm\) ( p -value \(= 0.075\) ) and \(Confirm + Ingroup \times Confirm\) ( p -value \(<0.001\) ) are also significantly negative. These outcomes together suggest that there exists an asymmetry of belief updating that subjects put less weight on the private signal when it confirms the social signal than when it contradicts the social signal, regardless of group identity. In addition, the coefficients of \(Confirm + IDM \times Confirm\) are not significant ( p -value \(\approx 0.5\) ), that is, the asymmetry in the belief updating pattern upon receiving two signals is unique to the contexts involving social learning.

However, the more detailed patterns we find are different from those of De Filippis et al. ( 2022 ). Table 3 shows that subjects in the SL and GSL treatments saliently underweight their private signal when it confirms Player A’s action compared to the Bayesian belief updating, while they update their beliefs upon their private signal when it contradicts Player A’s action in a more or less Bayesian way. In other words, the asymmetry in our experiment comes from the fact that subjects underweight the private signal when it confirms the social signal instead of the fact found by De Filippis et al. ( 2022 ) that subjects overweight the private signal when it contradicts the social signal. Footnote 12

Second, the coefficients of \(Ingroup \times Confirm\) are negative and significant. Given that the coefficients of Confirm are negative, it shows that subjects in the same group as Player A exhibit a greater asymmetry in belief updating than those in the other group do. More specifically, the coefficients of Ingroup are positive and significant, indicating that subjects in the same group as Player A put more weight on the private signal when it contradicts the social signal than those in the other group do. We also test the coefficients of \(Ingroup + Ingroup \times Confirm\) and obtain significantly negative results ( p -value < 0.005). It means that group identity decreases the weight on the private signal when it confirms the social signal. As a consequence, the asymmetry in the belief updating pattern increases with group identity as mentioned above. H3 is not supported.

Result 2a: Subjects in the same group as Player A put more weight on the private signal when it contradicts Player A’s action than those in the other group do, that is, \(\alpha _{2,ING} > \alpha _{2,OUTG}\) if \(s = \mathbbm {1}(a_A<50)\) & \(a_A \ne 50\) .

Result 2b: Subjects in the same group as Player A put less weight on the private signal when it confirms Player A’s action than those in the other group do, that is, \(\alpha _{2,ING} < \alpha _{2,OUTG}\) if \(s=\mathbbm {1}(a_A>50)\) & \(a_A \ne 50\) .

CDFs of \(\alpha _2\) with groups. Notes : Observations satisfying \(-1\le \alpha _2\le 3\) are included

We also run a quantile (median) regression \(\alpha _2\) on the same independent variables as in Table 5 . Column (2) in Table A2 displays the results. The signs and significance of the estimated coefficients are the same as in Table 5 , except for the coefficient of \(Ingroup \times Confirm\) which is negative but not significant. Therefore, we still find an asymmetric pattern of belief updating. We also find some evidence for confirmation bias in the IDM treatment since the coefficient of \(Confirm + IDM \times Confirm\) is positive and marginally significant ( p -value = 0.089). We do not reproduce the group effect on the asymmetry of belief updating, but the significantly positive coefficient of Ingroup supports Result 2a .

Last, we are interested in whether group identity affects learning efficiency. In particular, we examine whether a player’s payoffs of her two beliefs differ between learning from and an outgroup Player A. We conduct regressions of the payoffs of the first belief and second belief respectively. Table 6 displays the results—we do not find any salient effect of group identity on learning efficiency. Footnote 13

3.3 Group identity and the beliefs about Player A’s rationality

With the documented ingroup bias in forming the two beliefs, we further explore whether the impact of group identity is due to that subjects perceive a higher rationality of Player A in the same group than in the other group, or it is a result of an intrinsic heuristic to follow ingroup members.

Based on subjects’ answers to the questions in round 21, we generate new variables \(B_{Aerror,1}\) and \(B_{Aerror,2}\) to capture their belief that A made a mistake after observing the first and second signal respectively. As explained in Sect. 2.1 , in the experiment, making a mistake means that Player A reported \(a_A>50\) upon observing a black ball or \(a_A<50\) upon observing a white ball. And for the questions in round 21, subjects are requested to report the probability that they think Player A observed a white ball. Therefore, if \(a_A>50\) , the number a subject reported actually means the probability that she thinks Player A did not make a mistake, that is, \(B_{Aerror}=100\ -\) reported number. On the other hand, if \(a_A<50\) , the probability that a subject thinks Player A observed a white ball is the same as the probability that she thinks Player A made a mistake, that is, \(B_{Aerror}=\) reported number. For example, a subject reported that the probability that she thought Player A observed a white ball is 80% (20%) when she saw \(a_A>50\) ( \(a_A<50\) ). It means that her belief that Player A made a mistake is 20%, that is \(B_{Aerror}=20\) . We use \(B_{Aerror,1}\) and \(B_{Aerror,2}\) as the proxies for subjects’ beliefs about Player A’s rationality, which are relevant and not confusing. Figure 7 displays the means and confidence intervals (at 95% level) of \(B_{Aerror,1}\) (left panel) and \(B_{Aerror,2}\) (right panel) for ingroup and outgroup. We can see that subjects’ beliefs that Player A made a mistake are even a bit higher when Player A is an ingroup than when she is an outgroup, but the differences are not significant.

The means and confidence intervals of \(B_{Aerror,1}\) and \(B_{Aerror,2}\) with groups

We run OLS regressions of \(\alpha _1\) in round 21 on \(B_{Aerror,1}\) as well as Ingroup and SL . Table 7 shows the result. First of all, the impact of group identity on \(\alpha _1\) is confirmed. Also, the coefficients of \(B_{Aerror,1}\) are significantly negative, that is, a subject puts less weight on Player A’s action if her belief that A has made a mistake is stronger. But controlling for \(B_{Aerror,1}\) does not decrease the ingroup bias in \(\alpha _1\) . These results imply that the impact of group identity on processing the first social signal is probably not driven by the beliefs that ingroup members have higher information-processing abilities or make more rational decisions than outgroup members, rejecting H2 . In other words, even though we find that subjects put more weight on Player A’s action if they are in the same group than in different groups, there is little evidence to suggest that this observed ingroup bias in social learning is driven by people’s different beliefs about the predecessor’s rationality, that is, the belief-based channel. Combining the outcomes displayed in Table A4 that the group identity effects on \(\alpha _1\) in the first 4 and 7 rounds is significantly stronger than in the later rounds, we conjecture that our findings are more consistent with the perspective that group identity works as a framing device and its effect can be easily displaced by changing the framing of decisions.

Result 3: The group identity effect on \(\alpha _1\) can not be explained by different beliefs between ingroup and outgroup Player A’s rationality.

We also run OLS regressions of \(\alpha _2\) on \(B_{Aerror,2}\) . Table A3 in Online Appendix A shows that \(B_{Aerror,2}\) has no significant effects on \(\alpha _2\) for round 21. And it does not exhibit the group identity effects on \(\alpha _2\) as summarized in Result 2a and Result 2b either. Therefore, it is infeasible for us to conclude that the group effects on \(\alpha _2\) are driven by the ingroup-outgroup differences in \(B_{Aerror,2}\) or not. However, we do find that a contradicting private signal results in a significantly higher \(B_{Aerror,2}\) than a confirming private signal. The \(p\) -values from either the Mann-Whitney U test or the t-test are smaller than \(0.001\) ; results are the same if we use the difference between \(B_{Aerror,2}\) and \(B_{Aerror,1}\) instead.

4 Conclusion and discussion

In this paper, we investigate whether and how group identity influences the way people process social information as well as subsequent private information. We adopt the design paradigm of De Filippis et al. ( 2022 ) and conjecture how group identity may have an impact via the belief and the belief-updating process regarding the predecessor’s rationality under the theoretical framework of the LRTU model, which is also developed by De Filippis et al. ( 2022 ).

Our results show that there exist ingroup-outgroup differences, not only in the processing of the social signal, but also seem to be in that of the subsequent private signal. That the weight on the social signal is higher if it is from an ingroup predecessor compared to if it is from an outgroup predecessor, is broadly in line with the finding of ingroup favoritism in the experimental literature. In contrast, the patterns of \(\alpha _2\) imply that the effect of such ingroup favoritism on posterior beliefs is at least partly offset by the ingroup-outgroup difference in the private signal phase. Specifically, when the private signal accords (contradicts) the social signal of an ingroup predecessor, people appear to place a lower (higher) weight on the private signal compared to when the predecessor is an outgroup.

As to the underlying mechanism, we do not find evidence supporting the belief-based channel we have conjectured under the framework of the LRTU model, because the ingroup-outgroup difference in the social signal phase can not be explained by the elicited beliefs in the predecessor’s rationality, and no identity effects exist for the rationality beliefs. This result seems to be at odds with results from other studies in the literature: for example, Cacault and Grieder ( 2019 ) documented an ingroup bias in beliefs about others’ intelligence. We conjecture that the diverging findings might be driven by how challenging the cognitive task is. In our experimental context, the rational response is very obvious given the signal valence. In contrast, the intelligence test questions studied by Cacault and Grieder ( 2019 ) were quite challenging. Better understanding the circumstances in which group identification distorts people’s beliefs about the ability of their peers is an interesting question for future research. On a separate note, it is also unlikely that social preferences drive the results. If individuals’ reported event beliefs were shaped by their preference to have similar monetary payoffs from the belief-elicitation task as their ingroup members, such a preference should remain in the private signal phase, which is at odds with the fact that the effect of the group identity on posterior event beliefs is partly offset after the private signal phase.

We consider our results as supporting the alternative perspective in the literature that group identity works as a framing device (Tajfel et al., 1971 ; Bacharach, 2006 ; Guala & Filippin, 2017 ; Filippin & Guala, 2017 ), as this perspective can reconcile the observed patterns in both the two phases. In the first phase, information about the group identity of the predecessor may work as a cue of the group frame that triggers certain heuristics. Specifically, since subjects observing an outgroup predecessor place lower weights on the social signal than subjects observing an ingroup, whose weights are in turn lower than those in the SL treatment, we conjecture that the heuristics triggered by the group frame prescribe caution and suspicion with others’ actions and opinions, especially with those of the outgroup. The fact that such an ingroup-outgroup difference becomes less prominent in later rounds is consistent with the decaying nature of group identity as a framing device. In the second phase, the group frame can be considered removed since the subjects now receive their own private signal. Subjects then process the social signal and private signal together, without the influence of the heuristics as in the first phase. Since we calculate the weight subjects put on the private signal based on the beliefs before and after observing the private signal with the Bayesian structure, it just appears as if the subjects put a higher (lower) weight on a confirming (contradicting) private signal if the predecessor is an outgroup than if the predecessor is an ingroup.

To better illustrate our reasoning, we provide the following examples. Suppose that Player A reports 30 after observing a black ball. Then as affected by the heuristic in the social signal phase, a Bayesian subject reports \(a_1 = 30\) after observing Player A’s action when she is in the same group as Player A and reports \(a_1 = 40\) when she is in the other group. Footnote 14 In other words, the subject underweights Player A’s action if they are not in the same group. In the private signal phase, if the subject is NOT affected by the group framing any more, she will update her belief upon \(a_A = 30\) and \(s = 1\) after observing a white ball which contradicts Player A’s action, that is, the subject will report the stated belief \(a_2 = 50\) regardless of group identity. But the calculation of \(\alpha _2\) will be based on \(a_1 = 40\) when the subject is not in the same group as Player A. Then for the subject we obtain \(\alpha _{2,contradict,ING}=1 > \alpha _{2,contradict,OUTG} \approx 0.47\) , which is consistent with Result 2a .

Aagain, suppose that Player A reports 70 after observing a white ball. Then as a consequence of ingroup bias, a Bayesian subject reports \(a_1 = 70\) after observing Player A’s action when she is in the same group as Player A and reports \(a_1 = 60\) when she is in the other group. In the private signal phase, if the subject is NOT affected by the group framing, she will update her belief upon \(a_A = 70\) and \(s = 1\) after observing a white ball which confirms Player A’s action, that is, the subject will report \(a_2 = 84.5\) regardless of the predecessor’s group identity. But the calculation of \(\alpha _2\) will be based on \(a_1 = 60\) when the subject is not in the same group as Player A. Then for the subject we obtain \(\alpha _{2,confirm,ING}=1 < \alpha _{2,confirm,OUTG} \approx 1.52\) . which is consistent with Result 2b .

Comparing our results to those from De Filippis et al. ( 2022 ), we replicate an asymmetric pattern in the private signal phase in the sense that a private signal contradicting the social signal is given a much higher weight than a confirming private signal. The difference between results from the two studies is how the weights compare to the Bayesian benchmark. If we are willing to assume that our subjects have a general tendency to underweight in comparison to the Bayesian benchmark in the private signal phase, then the specific case of the LRTU model discussed by De Filippis et al. ( 2022 ) can explain the asymmetric pattern in our data as well. We provide evidence supporting such an interpretation by showing that beliefs about the predecessor’s rationality significantly differ upon receiving a contradicting and a confirming private signal. In other words, the cognitive process captured by the LRTU model probably exists among our subjects; however, the effect of the group identity does not seem to work via shifting parameters in this process.

Looking forward, it would be interesting for future researchers to further explore the relationship between social learning and group identity, viewing the case of a neutral event (the color of the jar) and an artefactual minimal group studied in the current paper as a first step. In the real world, there are many interesting cases in which certain beliefs are non-neutral in the sense that people have intrinsic preferences over the beliefs, and the identity of certain natural groups often plays a role in determining such intrinsic values. For example, believing that vaccines are ineffective in protecting us from viruses may directly bring the person utility by enhancing her sense of belonging to the social group she identifies herself with. The existence of such group identity-related utility has been proven by Hett et al. ( 2020 ) as they showed that individuals were willing to pay significant amounts of money to belong to certain groups. Another example is the belief about others’ prosocial or antisocial behavior. People probably prefer to believe that those within her social group are more prosocial than those outside the group. In general, such intrinsic preferences of beliefs can lead to motivated beliefs and bias the belief-updating process (see Gino et al. 2016 for a review of the broad literature on motivated reasoning.) Taking into consideration the social learning aspect thereof, there could be richer patterns to explore both theoretically and empirically. For instance, observing social signals that are against the predecessor’s identity-relevant utilities might be especially powerful in shifting beliefs, and if so, it can be utilized by policy-makers to bring about socially desirable changes. Another example is when the choice of information source is endogenous. While Duffy et al. ( 2019 ) documented distinct private and social information choosers but an overall tendency for choosing social information, how the inclusion of group identity may reshape such choices will be something interesting for future research.

As of August 2017, two-thirds (67%) of Americans report that they get at least some of their news on social media—with a fifth doing so often (Shearer & Gottfried, 2017 ).

Levy and Razin ( 2019 ) identify two broad reasons behind echo chambers: One is the tendency for individuals to segregate with like-minded ones, which leads to the creation of chambers; The other is the echo, that is, behavioral biases including correlation neglect, selection bias, and confirmation bias that induce polarization when individuals exchange beliefs in these chambers. Here we conjecture a new mechanism that can result in polarization of beliefs and opinions even if there is no segregation—if individuals put a greater weight on social information from ingroup than outgroup, and members of the same group receive correlated signals, then in some sense, such ingroup bias in social learning creates a virtual chamber. In other words, even if you are able to force people with opposite beliefs and opinions to desegregate and listen to each other, if there exists ingroup bias in social learning, their minds might still diverge further away.

A difference between our design and that of De Filippis et al. ( 2022 ) is that in our experiment, in each round, everyone except Player A observes one social signal. In contrast, in their experiment, there were some subjects who observed more than one social signal when making their own predictions in the social signal phase. According to the authors, data on these subjects’ behavior were not included in the analyses of the specific article; they were designed to answer other research questions.

Subjects’ responses to the predecessor’s decision may depend on their judgment about the predecessor’s rationality or intelligence. Online chat, which is a popular group enhancing device used in the literature, may provide information for them to make inferences about the the predecessor’s rationality or intelligence. Hence, we instead use common payoff to enhance group identity.

It is worth noting that the quadratic scoring rule is not incentive compatible under risk aversion, which might be the reason why subjects’ beliefs often peak at 50% in our results. There exist other methods that are incentive compatible for risk-averse agents in a theoretical sense such as the binarized scoring rule (Hossain & Okui, 2013 ). However, in state-of-the-art methodological discussions of belief elicitation, researchers emphasize the importance of so-called behavioral incentive compatibility and demonstrate that both the binarized scoring rule and the quadratic scoring rule violate some weak conditions for behavioral incentive compatibility (Danz et al., 2022 ). As a result, Danz et al. ( 2022 ) advocate for less-precise measurements with simpler and starker incentives for truthful reporting. In our elicitation procedure, after introducing the quadratic scoring rule, we explain to the subjects that it is in their best interest to state their true belief, to maximize the amount of money they could expect to gain. Such a descriptive representation of the scoring rule to some extent resonates with the advocate of Danz et al. ( 2022 )’s.

The social preference perspective might also predict an ingroup-outgroup difference when processing the social signal. Specifically, if people are more inequality averse toward an ingroup member than an outgroup individual, they will place a higher weight on the social signal to better align one’s own belief with the predecessor’s, reducing inequality in the monetary payoff from the belief-elicitation task. Alternatively, people might have an intrinsic preference to comply with (contradict) the behavior of an ingroup (outgroup). Either case, the social preference perspective predicts our \(H_1\) but not \(H_2\) or \(H_3\) . While our results support \(H_1\) but not \(H_2\) or \(H_3\) , there are other patterns in the private signal phase that are difficult to be reconciled from the social preference perspective. In general, we do not put much emphasis on this perspective in the current study, as we find it a bit stretchy while lacking specific support in the existing literature.

Subjects’ first beliefs in the IDM treatment have similar distributions to Fig. 1 except that a higher percentage of subjects report 50.

The distributions of players’ second beliefs for \(a_A < 50\) conditional on confirming and contradicting private signals are very similar to the case of \(a_A>50\) .

The three quantiles of \(\alpha _1\) are not saliently different between the IDM and the SL treatment. In other words, subjects do not put more weight on the (first) private signal than on the (first) social signal, which indicates that people process the (first) social and private information in similar ways.

We also have within-subject and between-subject analyses over the impact of group identity on \(\alpha _1\) . Both analyses produce significantly positive results, which are displayed in Table A1 in Online Appendix A .

In Online Appendix B, we analyse these extreme values separately.

In fact, De Filippis et al. ( 2022 ) also show that subjects underweight their private signal when it confirms the social signal, but they do not discuss this outcome.

We also run regressions with interaction terms between treatment dummy variables and Confirm and find no significant impacts of group identity on learning efficiency either.

Here we take a Bayesian subject as an example for the simplicity of expression. The reasoning also works for non-Bayesian subjects.

Anderson, L. R., & Holt, C. A. (1997). Information cascades in the laboratory. American Economic Review , 87 (5), 847–862.

Ash, E., & Van Parys, J. (2015). Group identity bias and information cascades.

Bacharach, M. (2006). Beyond Individual Choice: Teams and Frames in Game Theory . Princeton University Press.

Banerjee, A. V. (1992). A simple model of herd behavior. The Quarterly Journal of Economics, 107 (3), 797–817.

Article   Google Scholar  

Berger, S., Feldhaus, C., & Ockenfels, A. (2018). A shared identity promotes herding in an information cascade game. Journal of the Economic Science Association , 4 , 63–72.

Bernhard, H., Fehr, E., & Fischbacher, U. (2006). Group affiliation and altruistic norm enforcement. American Economic Review, 96 (2), 217–221.

Bettencourt, A., Brewer, M., Croak, M., & Miller, N. (1992). Cooperation and the reduction of intergroup bias: the role of reward structure and social orientation. Journal of Experimental Social Psychology, 28 (4), 301–319.

Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy, 100 (5), 992–1026.

Bolton, G. E., Feldhaus, C., & Ockenfels, A. (2016). Social interaction promotes risk taking in a stag hunt game. German Economic Review, 17 (3), 409–423.

Brewer, M. (1979). In-group bias in the minimal intergroup situation: A cognitive-motivational analysis. Psychological Bulletin, 86 (2), 307.

Brown, R., Vivian, J., & Hewstone, M. (1999). Changing attitudes through intergroup contact: The effects of group membership salience. European Journal of Social Psychology, 29 (5–6), 741–764.

Cacault, M. P., & Grieder, M. (2019). How group identification distorts beliefs. Journal of Economic Behavior & Organization, 164 , 63–76.

Çelen, B., & Kariv, S. (2004). Distinguishing informational cascades from herd behavior in the laboratory. American Economic Review, 94 (3), 484–498.

Charness, G., Rigotti, L., & Rustichini, A. (2007). Individual behavior and group membership. American Economic Review, 97 (4), 1340–1352.

Chen, D. L., Schonger, M., & Wickens, C. (2016). otree–an open-source platform for laboratory, online, and field experiments. Journal of Behavioral and Experimental Finance, 9, 88–97.

Chen, Y., & Li, S. (2009). Group identity and social preferences. American Economic Review, 99 (1), 431–57.

Currarini, S., & Mengel, F. (2016). Identity, homophily and in-group bias. European Economic Review, 90 , 40–55.

Danz, D., Vesterlund, L., & Wilson, A. J. (2022). Belief elicitation and behavioral incentive compatibility. American Economic Review, 112 (9), 2851–2883.

De Filippis, R., Guarino, A., Jehiel, P., & Kitagawa, T. (2022). Non-Bayesian updating in a social learning experiment. Journal of Economic Theory, 199, 105188.

Duffy, J., Hopkins, E., Kornienko, T., & Ma, M. (2019). Information choice in a social learning experiment. Games and Economic Behavior, 118, 295–315.

Filippin, A., & Guala, F. (2017). Group identity as a social heuristic: An experiment with reaction times. Journal of Neuroscience, Psychology, and Economics, 10 (4), 153.

Gino, F., Norton, M. I., & Weber, R. A. (2016). Motivated bayesians: Feeling moral while acting egoistically. Journal of Economic Perspectives, 30 (3), 189–212.

Gino, F., Shang, J., & Croson, R. (2009). The impact of information from similar or different advisors on judgment. Organizational Behavior and Human Decision Processes, 108 (2), 287–302.

Goeree, J. K., Palfrey, T. R., Rogers, B. W., & McKelvey, R. D. (2007). Self-correcting information cascades. The Review of Economic Studies, 74 (3), 733–762.

Goette, L., Huffman, D., & Meier, S. (2006). The impact of group membership on cooperation and norm enforcement: Evidence using random assignment to real social groups. American Economic Review (Papers & Proceedings), 96 (2), 212–216.

Grimm, V., Utikal, V., & Valmasoni, L. (2017). In-group favoritism and discrimination among multiple out-groups. Journal of Economic Behavior & Organization, 143, 254–271.

Guala, F., & Filippin, A. (2017). The effect of group identity on distributive choice: Social preference or heuristic? The Economic Journal, 127 (602), 1047–1068.

Hett, F., Mechtel, M., & Kröll, M. (2020). The structure and behavioral effects of revealed social identity preferences. The Economic Journal, 130 (632), 2569–2595.

Hossain, T., & Okui, R. (2013). The binarized scoring rule. Review of Economic Studies, 80 (3), 984–1001.

Levine, S. S., Apfelbaum, E. P., Bernard, M., Bartelt, V. L., Zajac, E. J., & Stark, D. (2014). Ethnic diversity deflates price bubbles. Proceedings of the National Academy of Sciences, 111 (52), 18524–18529.

Levy, G., & Razin, R. (2019). Echo chambers and their effects on economic and political outcomes. Annual Review of Economics, 11 , 303–328.

Li, S. (2020). Group identity, ingroup favoritism, and discrimination. In K. Zimmermann (Ed.), Handbook of Labor, Human Resources and Population Economics . Cham, Switz.: Springer. https://doi.org/10.1007/978-3-319-57365-6_123-1

Müller, D. (2019). The anatomy of distributional preferences with group identity. Journal of Economic Behavior & Organization, 166 , 785–807.

Nöth, M., & Weber, M. (2003). Information aggregation with random ordering: Cascades and overconfidence. The Economic Journal, 113 (484), 166–189.

Ockenfels, A., & Werner, P. (2014). Beliefs and ingroup favoritism. Journal of Economic Behavior & Organization, 108 , 453–462.

Ortoleva, P. (2012). Modeling the change of paradigm: Non-Bayesian reactions to unexpected news. American Economic Review, 102 (6), 2410–36.

Robson, M. (2021). Inequality aversion, self-interest and social connectedness. Journal of Economic Behavior & Organization, 183 , 744–772.

Ryen, A. H., & Kahn, A. (1975). Effects of intergroup orientation on group attitudes and proxemic behavior. Journal of Personality and Social Psychology, 31 (2), 302.

Shearer, E., & Gottfried, J. (2017). News use across social media platforms 2017 . Journalism and Media: Pew Research Center.

Google Scholar  

Tajfel, H., Billig, M., Bundy, R., & Flament, C. (1971). Social categorization and intergroup behaviour. European Journal of Social Psychology, 1 (2), 149–178.

Tanaka, T., & Camerer, C. F. (2016). Trait perceptions influence economic out-group bias: lab and field evidence from Vietnam. Experimental Economics, 19 (3), 513–534.

Weizsäcker, G. (2010). Do we follow others when we should? a simple test of rational expectations. American Economic Review, 100 (5), 2340–60.

Xu, X., Potters, J., & Suetens, S. (2020). Cooperative versus competitive interactions and in-group bias. Journal of Economic Behavior and Organization, 179, 69–79.

Ziegelmeyer, A., March, C., & Krügel, S. (2013). Do we follow others when we should? a simple test of rational expectations: Comment. American Economic Review, 103 (6), 2633–42.

Download references

Acknowledgements

We acknowledge the financial support from the National Natural Science Foundation of China (Grant No. 71903100; Grant No. 72003101), the Asia Research Center in Nankai University (Grant No. G02J0209), and the Laboratory for Economic Behavior and Policy Simulation in Nankai University (Grant No. ZX20220201; Grant No. ZX20220202 .The replication material for the study is available at http://doi.org/10.17605/OSF.IO/DR7AK .

Author information

Authors and affiliations.

Institute of State Economy, Nankai University, Weijin Road 94, Tianjin, 300071, China

School of Economics, Nankai University, Weijin Road 94, Tianjin, 300071, China

The Laboratory for Economic Behavior and Policy Simulation, Nankai University, Weijin Road 94, Tianjin, 300071, China

Wenbo Zou & Xue Xu

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Xue Xu .

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary file 1 (pdf 296 KB)

Rights and permissions.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Zou, W., Xu, X. Ingroup bias in a social learning experiment. Exp Econ 26 , 27–54 (2023). https://doi.org/10.1007/s10683-022-09788-1

Download citation

Received : 24 May 2021

Revised : 18 December 2022

Accepted : 20 December 2022

Published : 28 December 2022

Issue Date : March 2023

DOI : https://doi.org/10.1007/s10683-022-09788-1

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Ingroup bias
• Social learning
• Belief updating
• Laboratory experiment
• Find a journal
• Publish with us
• Track your research

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 03 November 2020

Overlap in processing advantages for minimal ingroups and the self

• Florence E. Enock 1 , 2 ,
• Miles R. C. Hewstone 1 ,
• Patricia L. Lockwood 1 , 3 , 4   na1 &
• Jie Sui 5   na1  

Scientific Reports volume  10 , Article number:  18933 ( 2020 ) Cite this article

2563 Accesses

8 Citations

9 Altmetric

Metrics details

• Human behaviour

Cognitive biases shape our perception of the world and our interactions with other people. Information related to the self and our social ingroups is prioritised for cognitive processing and can therefore form some of these key biases. However, ingroup biases may be elicited not only for established social groups, but also for minimal groups assigned by novel or random social categorisation. Moreover, whether these ‘ingroup biases’ are related to self-processing is unknown. Across three experiments, we utilised a social associative matching paradigm to examine whether the cognitive mechanisms underpinning the effects of minimal groups overlapped with those that prioritise the self, and whether minimal group allocation causes early processing advantages. We found significant advantages in response time and sensitivity ( d prime) for stimuli associated with newly-assigned ingroups. Further, self-biases and ingroup-biases were positively correlated across individuals (Experiments 1 and 3). However, when the task was such that ingroup and self associations competed, only the self-advantage was detected (Experiment 2). These results demonstrate that even random group allocation quickly captures attention and enhances processing. Positive correlations between the self- and ingroup-biases suggest a common cognitive mechanism across individuals. These findings have implications for understanding how social biases filter our perception of the world.

Similar content being viewed by others

Ingroup sources enhance associative inference

The neural representation of stereotype content

Ingroup favoritism overrides fairness when resources are limited

Introduction.

Information that is related to the self or a social ingroup to which one belongs is given high priority across a multitude of cognitive domains. Implicit self-associations tend to be positive 1 , 2 , people evaluate objects more favourably when they are given ownership of them 3 , 4 , 5 , participants are better at recalling adjective words that describe themselves 6 , and self-reference consistently enhances memory performance 7 , 8 , 9 . At the perceptual level, people show advantages in processing their own faces compared to those of other people 10 . Similarly, people generally prefer ingroups to outgroups 11 , with greater empathy shown towards ingroup members 12 , and implicit bias effects observed across a multitude of social contexts 13 . Group membership also influences lower-level perceptual judgements. Notably, there is a body of evidence demonstrating that social categorisation affects face processing. For example, biases are shown for own-race faces in recognition and memory 14 and these effects extend to other social contexts too 15 . Thus, the cognitive prioritisation of self- and ingroup-related information, known as self-bias and ingroup-bias, is ubiquitous. However, less is known about whether the same cognitive mechanisms underpin prioritisation for the self and social ingroups.

While belonging to social groups is adaptively beneficial in providing protection, support and the sharing of important resources 16 social categorisation is also a source of intergroup conflict and related issues such as prejudiced attitudes and discriminatory behaviours 17 . Understanding the cognitive basis of group membership and its relation to self identity therefore serves as an important foundation in guiding interventions to reduce, or even prevent, social conflict.

An associative matching paradigm developed by Sui and colleagues 18 established a method to study social effects on cognitive prioritisation whilst controlling for complete equivalence across the target stimuli, a potential confound in previous studies of processing advantages that have often used face and name stimuli. In this paradigm, participants learned to associate geometric shapes (e.g., square, circle and triangle) with social labels (e.g., ‘you’, ‘friend’ and ‘stranger’). Then, participants completed a computerised task in which they were asked to judge whether shape-label pairs presented on the screen conformed to the original pairings or not by responding with keys for yes and no. The stimuli were presented quickly at 100 ms with a short response window of 1100 ms, allowing for automatic levels of processing. A large prioritisation effect was found in response time, accuracy and sensitivity scores for self shapes compared to those of friend and stranger, and for friend compared to stranger shapes. These effects have been robustly replicated 19 , 20 , 21 , 22 , 23 , 24 and can operate even when associations are not instructed but are learnt over time 25 . Further, self-prioritisation exists even when familiarity is removed from all stimuli, such as when social labels were replaced with unfamiliar abstract symbols associated to the words (you, friend, stranger) before the experiment 26 . Initial studies have also demonstrated that ingroup associations can modulate processing for neutral stimuli in the same way 27 , 28 , 29 , 30 .

Whilst we tend to envision group membership as referring to static collectives imbued with historic meaning such as religions, nationalities and ethnicities, in fact the human drive for social classification is so great that ingroup favouritism is elicited even under ‘minimal group’ conditions, whereby individuals are assigned to novel groups randomly or on the ostensible basis of arbitrary information 31 , 32 , 33 , 34 , 35 . Interestingly, perceptual advantages for processing ingroup faces occur in minimal group contexts in the same way as for established groups 36 . This shows enhanced processing of own-group faces is not just rooted in greater exposure. There are competing theories as to how and why favouritism effects for minimal groups occur. According to Social Identity Theory 37 , individuals derive aspects of the self from the image of the ingroup, known as self-stereotyping . Ingroup favouritism occurs because when the group image is positive, self-esteem is enhanced by virtue of belonging. However, for self-stereotyping to occur, we would need adequate information about the group prototype, which would be unlikely in minimal group situations 38 . An alternative account for ingroup favouritism within the minimal group paradigm is known as self-anchoring 38 , 39 , which posits that rather than the self being assimilated to the group, it is the group that is assimilated to the self.

Studies that have examined the relations between self and ingroup biases have often used trait judgement paradigms which, for the most part, involve high-level evaluative decision-making processes 40 . To our knowledge, no studies have compared the relationship between self-bias and novel ingroup-bias for more automatic processes. Therefore it is unknown how these biases affect our processing of information in the world at an early stage. The primary aim of the present research was thus to test whether ingroup favouritism for minimal groups is demonstrated within low-level associative matching. To understand how group classification shapes earlier forms of processing has important implications for social biases at a multitude of levels. For example, biases in visual attention that are based in seemingly arbitrary criteria could provide the foundation for greater cognitive elaboration and in turn greater individuation of ingroup members. Correspondingly, lower attention to outgroups may lead to generalisation and deindividuation.

We examined whether neutral stimuli (lacking in any obvious social content) associated with novel ingroups are shown prioritisation compared to stimuli associated with novel outgroups. Secondly, we also examined how biases for novel ingroups relate to biases for the self. If self-anchoring is the means to favouritism for novel groups, then we may expect self-bias to be a reliable predictor of ingroup-bias. The paradigm utilised presently is beneficial in allowing for direct measurement and comparison of the magnitude of lower-level cognitive prioritisation for the self and novel ingroups under the same metric. This methodology provides a way of keeping complete equality between the familiarity of the target stimuli and the method of measuring responses to the stimuli.

Across three experiments we used a social associative matching paradigm to test for cognitive prioritisation of stimuli associated with novel ingroups, and to measure the relationship between these biases for the self and novel ingroups. In Experiment 1, all participants (within-subjects design) performed two separate matching tasks, one in which they learned to associate self and stranger labels with shapes (personal task), and the other in which they were allocated to novel teams and learned to associate ingroup and outgroup labels with different shapes (group task). In both tasks, participants had to respond to randomly presented shape-label pairs as correctly or incorrectly matched according to the learned associations. Experiment 2 aimed to examine the relationship between the self- and ingroup-prioritisation effects when the self, stranger, ingroup and outgroup associations were made within the same task. This allowed measurement of the relationship between the self- and ingroup-advantages when personal and group stimuli were salient in the same blocks of the same task, enabling us to test whether prioritisation for novel ingroups remains when the self is also salient. Experiment 3 measured the relationship between the self and ingroup advantages but this time when the personal and group social associations were made within different blocks of the same task. Performance advantages were always taken as the differences in response time and sensitivity ( d prime) between own associated (self/ingroup) shapes and other associated (stranger/outgroup) shapes. Note that bias effects for the self and novel ingroups are also referred to interchangeably as prioritisation and advantage effects. We refer to self and stranger as personal associations and ingroup and outgroup as group associations throughout. We refer to self and ingroup as ‘own’ and stranger and outgroup as ‘other’. In Experiment 1 ‘task type’ refers to whether the task involved personal or group associations. In Experiments 2 and 3, this becomes ‘association type’ as the tasks are no longer separated into two. ‘Shape’ always refers to the specific social concept (e.g., ‘self’) that it represents according to the association made at the start of the experiment. Mismatch trials are defined by shape such that mismatched trials for the self condition, for example, refer to a self shape presented with a stranger label. We analysed our data by shape condition (as opposed to label) because our primary interest is in how social meaning can be quickly tagged to novel stimuli (i.e., the shapes), rather than in biases for already learned social content (i.e., social labels).

We hypothesised that (1) A robust performance advantage for the novel ingroup would be demonstrated on the associative matching task in the same way as for the self and (2) If there is an overlap in the way we process information about the self and minimal ingroups, there should be a positive relationship between biases for the self and the novel ingroup.

Experiment 1—associative matching of self and novel ingroups

All studies (1–3) were approved by the Ministry of Defence Research Ethics Committee (MODREC) (approval number: 495/MODREC/13) and informed consent was obtained prior to the experiment according to approved ethical procedures. All studies were performed in accordance with the relevant guidelines and regulations.

Participants

Participants were recruited via a university online recruitment system. In a related previous research study 28 (Experiment 1), a sample of forty-two gave power of 0.99 to detect a significant effect of shape on response time and sensitivity (alpha 0.05) with partial eta 2 of ~ 0.60 in an similar repeated-measures design (power calculation conducted using MorePower 6.0.4). A similar sample size was planned in order for direct comparisons to be drawn. A total of fifty individuals took part (the additional eight participants were recruited to allow for data exclusion and drop out). Three were excluded on the basis of having chance accuracy for one or both tasks (M < 55%), leaving a total of forty-seven (24 female, range 19–34 years, mean age = 23.6, SD = 3.66). All were right-handed and had normal or corrected-to-normal vision.

Stimuli and tasks

Two associative matching tasks adapted from Sui et al. 18 were performed by each participant in a within-subjects design (Fig.  1 ). Each task took approximately twelve minutes. In the personal task, participants learned to associate two geometric shapes with ‘self’ and two with ‘stranger’ labels and then had to judge if randomly presented shape-label pairs were correctly or incorrectly matched. The group task took the same format, but participants learned to associate two geometric shapes with novel ingroup and outgroup labels (Team Green and Team Blue, counterbalanced). Although word length and category were not made equivalent, control experiments in previous work verified self bias effects were not rooted in word length, concreteness, usage frequency or familiarity 18 , 29 , 41 . Thus, we were confident we could choose the simplest words to identify own and other categories without impacting results.

Self and ingroup associative matching paradigm. A schematic representation of the matching task that was utilised across all three experiments. The shape-label pairings were always counterbalanced across participants. After a short learning phase of the initial shape-label associations, participants responded to randomly presented shape-label pairs shown on-screen as matched or mismatched by pressing assigned response keys. In Experiment 1, the personal (you and stranger) associations were made in a separate task to the group (ingroup and outgroup) associations and order of tasks were counterbalanced. In Experiment 2, all of the associations (personal and group) were learned and presented in the same blocks of the same tasks. In Experiment 3, the associations were learned as part of one task, but were presented and responded to in alternating blocks. Two shapes were associated with each label in Experiments 1 and 3 (as shown in the figure), but only one shape was associated with each label in Experiment 2.

For both tasks, the shape-label pairings were initially made by an on-screen instruction (e.g., ‘you are the square and the triangle and the stranger is the pentagon and the circle’). The associations were displayed at the same time and stayed on the screen until participants pressed the spacebar to continue to the experiment. By this design, the shape-label mappings are not learned during the course of the experimental task, but are rather explicitly instructed at the start. After these initial associations were made, 12 practice trials were completed and then a 2-s central fixation cross indicated the start of the main computer task. In each trial, a randomly-generated shape-label pair was shown on screen for 100 ms. Then, there was a 1100 ms blank response window in which participants indicated whether the pair had been correctly or incorrectly matched by pressing m or n , counterbalanced for yes or no . Feedback was given for 500 ms after the response for each trial. In total, 384 experimental trials were spread over 6 blocks (64 trials per block). Accuracy was presented at the end of each block followed by an 8-s break. There were four conditions: two match conditions, (matched/mismatched); and two shape conditions (self and stranger), with 96 trials per condition. The task descriptions are based on methods described in Enock et al. 28 for Experiment 1.

The associations across both tasks between the eight shapes (circle, triangle, square, pentagon, octagon, cross, diamond, rectangle) and the four written labels (self, stranger; Team Blue, Team Green) were counterbalanced across participants. Order of task completion was also counterbalanced such that half the participants completed the personal task first and half completed the group task first. In both tasks, the geometric shapes (each 118 × 118 pixels, 4.16 × 4.16 cm, corresponding to a visual angle of around 4.8° × 4.8° as viewed 50 cm from the screen) were presented randomly above a white fixation cross (0.8° × 0.8°) against a grey background at the centre of the screen, with a matched or mismatched label presented simultaneously below the central fixation point. The experiment was run on a PC using E-Prime software (version 2.0) and displayed on a 17-inch monitor.

Upon arriving to the experimental session, informed consent was obtained in accordance with approved ethical procedures and participants were briefed with regards to what the computer tasks involved. Then, participants were verbally allocated to either Team Green or Team Blue by being told:

‘For this study, each participant is randomly assigned to one of two teams: Team Green or Team Blue. You are on Team (Green). Please remember your team affiliation throughout the study’.

Those allocated to Team Blue received the corresponding instruction. Team allocation was made such that odd-numbered participant numbers were allocated to Team Green, while even-numbered participant numbers were allocated to Team Blue. After group allocation, some brief demographic information was taken and participants began the first of the two computer tasks. The first screen showed the shape-label pairings and once participants felt confident they had learned them, they pressed the spacebar to continue to the first block. When participants had completed the first task, they were given a short break while the researcher prepared the computer for the next part, and when they were ready, they began the second task. At the end, participants were debriefed and remunerated for their time at a rate of £10 per hour.

Statistical analyses

For all experiments, data were analysed in SPSS 24 (Armonk, New York: IBM Corp). For all response time analyses, only correct responses were included, and those higher or lower than 2.5 standard deviations from the mean response time for each participant under each condition were excluded as a standard way of removing outliers 42 . Less than 5% of each dataset was excluded and each analysis was performed on the remaining trials. Data were also analysed using signal detection theory (dprime) 43 which is a useful measure that allows for the distinction between effects that can be attributed to changes in perceptual sensitivity (i.e., the ability to discriminate between targets) and effects that are related to response biases (i.e., whether participants have a preferred type of response throughout the task). The estimation of parameters included requires the classification of each response type as either a correct identification of a target (a ‘hit’ response), a correct rejection of a non-target (a ‘correct rejection’), a misidentification of a non-target as a target (a ‘false alarm’) and misjudging a target as a non-target (a ‘miss’). In the current experimental design, a response was considered as a hit when a matching shape-label pair appeared and participants responded ‘yes’; a correct rejection when the shape and label did not match and participants responded ‘no’; false alarm when a shape and label did not match and participants responded ‘yes’; and a miss when a shape and label matched but a ‘no’ response was given. The sensitivity index ( d prime) represents the ability to distinguish matching and shape-based mismatching stimuli, and the response criterion represents a bias towards judging mismatching stimuli as matching, or the opposite. Matched and mismatched trials for each shape were combined to give a measure of d prime using the following formula, where H denotes correct hits and F denotes false alarms 43 :

and the response criterion (RC) for each participant was also calculated as follows 44

For each experiment, we examined biases for the self and ingroup using repeated measures ANOVAs on response time and sensitivity ( d prime) scores. There were two task types (personal: self and stranger associations; and group: ingroup and outgroup associations), two shape conditions (own: self and ingroup; and other: stranger and outgroup); and two match conditions (matched/mismatched). Note: ± always denotes standard error of the mean.

Results (Experiment 1)

Own-associated biases in response time data.

There were three within-subject variables: task type (personal: self/stranger or group: ingroup/outgroup associations); shape category (own: self/ingroup, or other: stranger/outgroup) and match condition (matched/mismatched). A 2 (task: personal/group) × 2 (shape: own/other) × 2 (matched/mismatched) repeated-measures analysis of variance (ANOVA) was used to test for the effect of shape on response time scores.

There was no significant effect of task type, F (1, 46) = 0.90, p  = 0.347, η p 2  = 0.019, showing that response time was similar across the personal and group tasks. There was a significant effect of shape, F (1, 46) = 29.62, p  < 0.001, η p 2  = 0.392, with faster response times for own ( M  = 745 ± 8.59) than other ( M  = 766 ± 8.64) shapes. There was also a significant effect of match condition, F (1, 46) = 242.73, p  < 0.001, η p 2  = 0.841, with faster response times for matched ( M  = 725 ± 8.54) than for mismatched ( M  = 786 ± 8.68) trials.

There was no significant interaction between task and shape, F (1, 46) = 1.01, p  = 0.319, η p 2  = 0.022, nor between task and match, F (1, 46) = 1.54, p  = 0.222, η p 2  = 0.032. There was, however, a significant interaction between shape and match condition F (1, 46) = 58.09, p  < 0.001, η p 2  = 0.558, and there was also a significant three-way interaction between task, shape and match, F (1, 46) = 8.24, p  = 0.006, η p 2  = 0.152. To decompose these interactions, separate 2 (task type) × 2 (shape) ANOVAs were performed for matched and mismatched trials. For matched trials, there was no significant main effect of task type, F (1, 46) = 1.55, p  = 0.220, η p 2  = 0.033, but there was a significant main effect of shape, F (1, 46) = 62.51, p  < 0.001, η p 2  = 0.576, and a significant interaction between the two, F (1, 46) = 6.53, p  = 0.014, η p 2  = 0.124. Self shapes ( M  = 699 ± 10.05) were responded to significantly faster than stranger shapes ( M  = 763 ± 8.44), t (46) = 7.99, p  < 0.001, d  = 1.16, and ingroup shapes ( M  = 698 ± 11.78) were responded to significantly faster than outgroup shapes ( M  = 741 ± 11.23), t (46) = 5.57, p  < 0.001, d  = 0.81.

For mismatched trials, there was no significant main effect of task type, F (1, 46) = 0.33, p  = 0.571, η p 2  = 0.007, but there was a significant main effect of shape condition, F (1, 46) = 5.32, p  = 0.026, η p 2  = 0.104, with other shapes responded to significantly faster than own shapes. There was no significant interaction between task type and shape condition, F (1, 46) = 0.33, p  = 0.298, η p 2  = 0.024. In summary, the effect of shape on response time was similar across both tasks. Figure  2 a shows response times for matched and mismatched trials for each shape association.

Self and ingroup information is prioritised and processed with greater sensitivity compared to strangers and outgroups. Mean response time (RT) scores as a function of shape (self vs. stranger, ingroup vs. outgroup), match (matched vs. mismatched) and task (self vs. group) (a) along with mean sensitivity ( d prime) scores as a function of shape and task (b) for Experiment 1. Error bars represent standard errors. *** denotes p  < .001; * denotes p  < .01; n.s. denotes non-significant.

Own-associated biases in dprime (perceptual sensitivity)

A 2 (shape: own/other) × 2 (task: personal/group) repeated measures ANOVA revealed a significant main effect of shape, F (1, 46) = 46.63, p  < 0.001, η p 2  = 0.503 and of task type, F (1, 46) = 24.44, p  < 0.001, η p 2  = 0.347. Sensitivity was greater for own ( M  = 1.99 ± 0.12) than for other ( M  = 1.40 ± 0.11) shapes, and for the personal ( M  = 2.02 ± 0.11) than the group ( M  = 1.37 ± 0.13) task.

There was a significant interaction between shape and task, F (1, 46) = 7.70, p  = 0.008, η p 2  = 0.143. Self shapes were responded to with greater sensitivity than stranger shapes, t (46) = 3.68, p  < 0.001, d  = 0.54, and ingroup shapes were responded to with greater sensitivity than outgroup shapes, t (46) = 6.72, p  < 0.001, d  = 0.98, with the latter effect almost twice the magnitude. Additionally, response criterion scores were significantly lower for self than stranger shapes and for ingroup than outgroup shapes (both p  < 0.01). Similar to response time data, the d prime scores demonstrated performance advantages for self compared to stranger, and for ingroup compared to outgroup. Figure  2 b shows sensitivity scores for each shape association.

Discussion (Experiment 1)

The first experiment aimed to establish whether there was an ingroup-advantage in associative matching within the minimal group paradigm. For matched trials, self shapes were responded to significantly faster than stranger shapes and ingroup shapes were responded to significantly faster than outgroup shapes. For mismatched trials, stranger shapes were responded to significantly faster than self shapes, but there was no difference in response time between ingroup and outgroup shapes. Slower responses for self than stranger on mismatched trials is a common finding within this paradigm and the self-bias effect is usually only observed within matching trials 18 . D prime scores were enhanced for self over stranger and for ingroup over outgroup shapes. Further, there was a lower response criterion for self and ingroup than stranger and outgroup shapes, showing participants tended to more readily indicate that the stimuli were own- than other-associated. In line with our main hypothesis, these results demonstrate that an ingroup-bias in attentional prioritisation is manifested even when group affiliations are new, arbitrary and randomly assigned, and in the absence of other group members.

These results complement the body of literature demonstrating the effects of minimal group allocation on both higher-level decision-making 31 , 32 , 35 , and also on more implicit processes 45 , 46 , 47 . The present results show novel group allocation affects the earlier processing of previously neutral, non-face stimuli. This experiment benefitted from a unique approach to measuring minimal group-biases by using random ingroup allocation combined with an associative matching paradigm, useful in allowing for comparison across social associations whilst keeping the measurement and visual features of the stimuli exactly the same.

Experiment 2 included self, stranger, ingroup and outgroup associations within the same task blocks so we could test whether self- and ingroup-biases remain constant when both types of stimuli are salient at the same time. As this paradigm included combinations of shape-label pairings for self and ingroup together, this also allowed us to examine the overlap between self and novel ingroup representation. For example, higher error rates and slower response times for self shapes paired with ingroup labels than self shapes paired with outgroup labels would indicate a cognitive overlap between self and ingroup stimuli.

Experiment 2—competition between self and ingroup

As before, ethics was approved by MODREC (approval number: 495/MODREC/13) and all participants gave informed consent. Relevant guidelines and regulations were followed.

Participants were recruited via the university online system. Sample size was again planned according to a related previous research study 28 (Experiment 2) in which an N of thirty-one gave power of 0.97 to detect a significant effect of shape on response time and sensitivity (alpha 0.05) with partial eta 2 of 0.346 for response time (0.484 for d prime) in a similar repeated-measures design. Using these effect sizes as estimates, a sample size calculation showed eighteen participants would be sufficient to detect the effect of shape on response time and sensitivity with power of 0.8 and alpha 0.05. Therefore, although this sample size was smaller than that of Experiment 1, a larger sample size was not necessary to establish the effects of interest. The power and sample size calculations were conducted using MorePower 6.0.4. A total of 22 individuals (14 female, age range 19–28 years, mean age = 23.3, SD  = 2.71) took part, all right-handed and with normal or corrected-to-normal vision.

Stimuli, task and procedure

The task was very similar to those in Experiment 1, but this time the personal and group associations were made within the same blocks of the same matching task rather than separately in two tasks. Other than that, exactly the same experimental parameters were used as for Experiment 1. In an initial on-screen instruction, participants learned to pair Self, Stranger, Team Green and Team Blue labels with different geometric shapes (circle, square, triangle, pentagon), fully counterbalanced across participants. All shape-label associations were displayed together and remained on the screen until participants pressed spacebar to continue. In each trial, participants again judged whether shape-label pairings presented on the screen were correctly or incorrectly matched. The task consisted of six blocks of 96 trials per block (excluding 12 practice trials at the beginning). There were eight conditions within the task: four shape conditions (self, stranger, ingroup, outgroup) and two match conditions (matched/mismatched), with 72 trials per condition. The procedure followed exactly that of Experiment 1, except participants completed one rather than two tasks.

In addition to the analyses we conducted in Experiment 1, we compared specific combinations of personal and group mismatched pairs in response time and accuracy in order to test whether participants found it more difficult to reject self and ingroup paired stimuli than self and outgroup or ingroup and stranger. A greater difficulty in rejecting self and ingroup as correct would indicate an overlap in the cognitive representation of newly learned self and ingroup associations. Paired samples t-tests were used to examine differences in response time and accuracy scores for shape  + label mismatches of self  + ingroup and self  + outgroup, along with ingroup  + self and ingroup  + stranger.

Results (Experiment 2)

A 2 (association type: personal/group) × 2 (shape: own/other) × 2 (matched/mismatched) repeated measures ANOVA found a significant effect of association type, with responses faster for the personal ( M  = 682 ± 10.89) than the group ( M  = 705 ± 11.01) shapes, F (1, 21) = 25.06, p  < 0.001, η p 2  = 0.544. There was a marginal effect of shape, F (1, 21) = 4.23, p  = 0.052, η p 2  = 0.167, with faster response times for own ( M  = 687 ± 10.86) than for other ( M  = 698 ± 10.91) shapes. There was a significant effect of match condition, F (1, 21) = 79.28, p  < 0.001, η p 2  = 0.791, with responses faster for matched ( M  = 663 ± 9.60) than for mismatched ( M  = 725 ± 12.71) pairs.

There was no significant interaction between association type and shape, F (1, 21) = 2.70, p  = 0.116, η p 2  = 0.114. However, there was a significant interaction between association type and match condition, F (1, 21) = 25.01, p  < 0.001, η p 2  = 0.544, and between shape and match condition, F (1, 21) = 18.56, p  < 0.001, η p 2  = 0.469. There was also a significant three-way interaction between association type, shape and match condition, F (1, 21) = 7.38, p  = 0.013, η p 2  = 0.260.

To decompose this, separate 2 (association type) × 2 (shape) ANOVAs were performed for matched and mismatched trials. For matched trials, there was a significant main effect of association type, F (1, 21) = 38.39, p  < 0.001, η p 2  = 0.646, with faster responses for the personal ( M  = 637 ± 9.96) than the group ( M  = 688 ± 10.90) task. There was also a significant main effect of shape, F (1, 21) = 17.18, p  < 0.001, η p 2  = 0.450, with faster responses for own ( M  = 646 ± 10.41) compared to other ( M  = 679 ± 10.38) shapes. Importantly, a significant interaction between association type and shape, F (1, 21) = 6.57, p  = 0.018, η p 2  = 0.238, showed that for matched trials, responses were significantly faster for self than for stranger shapes ( p  < 0.001) but not for ingroup than outgroup shapes ( p  = 0.356).

For mismatched trials, there was no significant effect of association type, F (1, 21) = 0.49, p  = 0.494, η p 2  = 0.023, with similar response times for personal ( M  = 727 ± 13.67) and group ( M  = 722 ± 12.42) trials. There was a significant main effect of shape type, F (1, 21) = 7.89, p  = 0.011, η p 2  = 0.273, with faster responses for other ( M  = 717 ± 12.92) than own ( M  = 733 ± 13.17) shapes. There was no significant interaction between association type and shape condition F (1, 21) = 1.63, p  = 0.216, η p 2  = 0.072. In summary, responses were significantly faster for self than stranger shapes (but not for ingroup than outgroup) on matched trials, while for mismatched trials, responses were overall faster for other than own associations. Figure  3 a shows response times for matched and mismatched trials under each shape condition.

Self, but not ingroup, information is prioritised and processed with greater sensitivity. Mean response time (RT) scores as a function of shape (self vs. stranger, ingroup vs. outgroup), match (matched vs. mismatched) and task (self vs. group) (a) along with mean sensitivity ( d prime) scores as a function of shape and task (b) for Experiment 2. Error bars represent standard errors. *** denotes p  < .001; * denotes p  < .01; n.s. denotes non-significant.

A 2 (association type: personal/group) × 2 (shape: own/other) repeated measures ANOVA found no significant effect of association type, F (1, 21) = 3.22, p  = 0.087, η p 2  = 0.133, with comparable scores for personal ( M  = 2.19 ± 0.19) and group ( M  = 1.98 ± 0.19) associations. There was a significant effect of shape, F (1, 21) = 8.20, p  = 0.009, η p 2  = 0.281, with greater sensitivity for own ( M  = 2.27 ± 0.18) than for other ( M  = 1.91 ± 0.20) shapes. A significant interaction, F (1, 21) = 5.63, p  = 0.027, η p 2  = 0.211, indicated the effect of shape was dependent on the association type, with enhanced perceptual sensitivity for self shapes relative to stranger shapes (p  = 0.005), but not for ingroup relative to outgroup shapes, ( p  = 0.417).

Similar to response time data, sensitivity was enhanced for self relative to stranger, but not for ingroup relative to outgroup stimuli. Additionally, the response criterion was significantly lower for self shapes than for stranger shapes ( p  = 0.002), but there was no difference in response criterion between ingroup and outgroup shapes ( p  = 0.124). Figure  3 b shows sensitivity scores for each shape condition.

Analysis of mismatched pairs

Response time data.

Although there was no significant advantage for the novel ingroup compared to the novel outgroup, performance for specific mismatched shape-label combinations tested an overlap between the representation of the self and novel ingroup. If the novel ingroup is anchored to the self, then we would expect slower response times when responding ‘no’ to a self shape mismatched with an ingroup label than a self shape mismatched with an outgroup label, as the latter would be more obviously incongruent. Paired-samples t-tests were conducted to determine if there were differences in response time scores between self shapes paired with ingroup labels and self shapes paired with outgroup labels. There were 24 trials for each mismatched condition.

There were no significant differences in response times for self shapes paired with ingroup labels than for self shapes paired with outgroup labels, t (21) = 1.22, p  = 0.235, d  = 0.26. However, ingroup shapes paired with self labels were responded to significantly faster than ingroup shapes paired with stranger labels, t (21) = 2.19, p  = 0.041, d  = 0.46. That participants were faster to correctly reject ingroup shapes paired with self labels is opposite to the predicted effect, and merely suggests generally faster responses for self labels than stranger labels.

Accuracy data

Accuracy scores were significantly higher for self shapes paired with outgroup labels than self shapes paired with ingroup labels, t (21) = 2.18, p  = 0.041, d  = 0.47. Accuracy scores were not significantly higher for ingroup shapes paired with stranger labels compared to ingroup shapes paired with self labels t (21) = 1.87, p  = 0.075, d  = 0.40. That participants found it easier to reject self +  outgroup than self +  ingroup may indicate an overlap in the way that newly learned self and novel ingroup stimuli are represented. Table 1 shows the mean response times and proportions of correct responses (accuracy scores) for each mismatched pair of interest.

Discussion (Experiment 2)

Experiment 2 tested whether self- and novel ingroup-advantages in associative matching are present when shape-label pairs were shown within the same blocks of the same task. This design allowed for comparison of self and ingroup biases when both kinds of social stimuli were present at the same time. A strong self-bias effect was found, with significantly faster responses for self than for stranger shapes on matched trials (but not on mismatched trials) and significantly enhanced d prime scores. However, there was no response time or d prime advantage for ingroup compared to outgroup shapes. Similarly, there was a response criterion bias for self shapes compared to ingroup shapes, with participants more likely to indicate stimuli as self- than stranger-associated, but with no such bias for ingroup compared to outgroup shapes. These results contrast to those from Experiment 1, which demonstrated a robust ingroup-advantage for new and arbitrarily assigned groups. Similarly, the results contrast to prior work in which a significant ingroup-advantage (with a similar sample size) in the context of well-established teams was found even when the self and ingroup pairings were presented within the same task blocks 28 . The results are, however, consistent with a recent study showing minimal ingroup prioritisation was lower in magnitude than self-prioritisation when the two kinds of social stimuli were salient at the same time, and barely present when other ingroup members were not known at all 27 , as was the case in the present study. These findings suggest that novel ingroup-biases do not follow exactly the same patterns as those for established groups, particularly when competing with the saliency of the self.

The results imply that when self and novel ingroup stimuli are presented at the same time, the self takes priority, with the ingroup-bias effect disappearing. Thus, in a more cognitively demanding task, self significance plays a more important role in performance than minimal group affiliation. Interestingly, this is not the case for established groups, in which both kinds of social biases remained constant even in the presence of one another 28 . Turner’s principle of functional antagonism, which is an important part of self-categorisation theory and closely linked to the social identity perspective 48 , holds that whilst many social identities exist within each individual, as the salience of one increases, the saliency of the other will correspondingly decrease. Thus if the self is salient in comparison to the minimal group, then attention may be shifted entirely to the self. With regards to the notion of self-anchoring, it may be that when explicit self stimuli are present, the projection of the self onto novel ingroup associations is less important and the experimental stimuli become categorised more as ‘self’ or ‘non-self’.

Although there was no ingroup-advantage, responses to specific combinations of shape-label mismatches provided some evidence for greater difficulty in distinguishing self from ingroup stimuli. Error rates were lower for self shapes paired with outgroup labels than with ingroup labels which indicates a closer association between self and the minimal ingroup, than self and the minimal outgroup. This may suggest a self-anchoring mechanism in which the novel ingroup is incorporated to the self 39 . Results from Experiment 2 were not consistent with those of Experiment 1 and did not support our main hypothesis that minimal ingroups would be prioritised over outgroups in the associative matching task. Experiment 3 aimed to test this discrepancy in a similar sample size to Experiment 2 by separating self and ingroup associations into alternating blocks to reduce competition between self and ingroup whilst still controlling for practice and order effects.

Experiment 3—blocked associative matching design

As before, ethics was approved by MODREC (approval number: 495/MODREC/13), all participants gave informed consent and relevant guidelines and regulations were followed.

As for Experiments 1 and 2, participants were recruited via the online system. Based on the sample size calculation described for Experiment 2, and for a comparable sample size, a total of 22 individuals took part (12 female, age range 19–35 years, mean age = 25.2, SD  = 4.5). All of the participants were right-handed and had normal or corrected-to-normal vision.

The task was very similar to that for Experiment 2, but this time the personal and group associations were made within alternating blocks of the same task and participants learned to associate two shapes per label rather than one. Other than that, exactly the same experimental parameters were used. In an initial on-screen instruction, participants learned to pair two geometric shapes each with Self, Stranger, Team Green and Team Blue labels, again counterbalanced. All shape-label associations were displayed together and remained on the screen until participants pressed spacebar to continue. In each trial, participants had to judge whether shape-label pairings subsequently presented were correctly or incorrectly matched. There were six blocks of 80 trials per block (excluding 12 practice trials at the beginning) and personal and group blocks were alternated such that there were 3 of each. Half of the participants were presented with a personal block first and half with a group block first. Participants were given a break in between each block during which they were reminded on-screen of the shape-label associations. There were eight conditions: four shapes (self, stranger, ingroup, outgroup) each either matched or mismatched, with 60 trials per condition. The procedure was the same as for Experiment 2.

Results (Experiment 3)

A 2 × 2 × 2 repeated measures ANOVA with association type (personal/group), shape condition (own/other) and match condition (matched/mismatched) as the three within-subject variables was used to test for the effect of shape on response time scores. There was no effect of task type on response time, F (1, 21) = 0.04, p  = 0.847, η p 2  = 0.002, showing that response times were generally similar for personal and group associations. There was a significant effect of shape on response time, F (1, 21) = 24.17, p  < 0.001, η p 2  = 0.535, with faster responses for own (self and ingroup) ( M  = 722 ± 14.9) than other (stranger and outgroup) ( M  = 756 ± 14.2) shapes. There was also a significant effect of match condition, F (1, 21) = 64.09, p  < 0.001, η p 2  = 0.753, with responses to matched pairs ( M  = 714 ± 14.0) significantly faster than responses to mismatched pairs ( M  = 764 ± 14.9).

There was no significant interaction between task and shape, F (1, 21) = 0.004, p  = 0.950, η p 2  = 0.000, but there was a significant interaction between task and match condition, F (1, 21) = 4.35, p  = 0.049, η p 2  = 0.171, and between shape and match condition, F (1, 21) = 68.78, p  < 0.001, η p 2  = 0.766, with own-associated shapes responded to faster than other-associated shapes on matched ( p  < 0.001) but not mismatched ( p  = 0.059) trials (Fig.  4 a). There was no significant three-way interaction between task, shape and match, F (1, 21) = 0.39, p  = 0.540, η p 2  = 0.018.

Self and ingroup stimuli are prioritised over stranger and outgroup and responded to with greater sensitivity. Mean response time (RT) scores as a function of shape, match and task (a) and mean sensitivity ( d prime) scores as a function of shape and task (b) for Experiment 3. Error bars represent standard errors. *** denotes p  < .001; n.s. denotes non-significant.

A 2 (association type: personal/group) × 2 (shape: own/other) repeated measures ANOVA found a significant effect of shape on sensitivity scores, F (1, 21) = 43.95, p  < 0.001, η p 2  = 0.677, with sensitivity significantly enhanced for own (self and ingroup) ( M  = 2.19 ± 0.20) over other (stranger and outgroup) ( M  = 1.40 ± 0.15) shapes (Fig.  4 b). There was no significant effect of association type on sensitivity, F (1, 21) = 0.92, p  = 0.349, η p 2  = 0.042, and no significant interaction between association type and shape, F (1, 21) = 0.15, p  = 0.710, η p 2  = 0.007. Overall, sensitivity was significantly enhanced for self relative to stranger shapes and for ingroup relative to outgroup shapes. As before, a significantly lower response criterion was adopted for own compared to other shapes ( p  < 0.01).

Discussion (Experiment 3)

Experiment 3 examined self and minimal ingroup advantage effects when the social associations were separated into alternating blocks within the same task. The ingroup advantage in response time and sensitivity that was observed in Experiment 1, but not in Experiment 2, was re-established and there were similar effect sizes for the self and ingroup response time and sensitivity advantages. These results were consistent with those from Experiment 1 and were also in line with our main hypothesis of a minimal group advantage effect in the associative matching paradigm. These results again complement the wider literature showing that minimal groups are prioritised across a range of cognitive processes. The self and ingroup bias effects in response time were observed only in the matched trials which, as noted in the discussion for Experiment 1, is consistent with previous work using this paradigm.

Although the self and ingroup associations were present within the same task as for Experiment 2, the separation into alternating blocks appeared to sufficiently reduce competition for attention between the self and novel group. This suggests individuals are able to quickly switch attentional priority between self and novel ingroup relevance when they are required to only focus on one at a time. Overall, this experiment provides further evidence for an ingroup-advantage for novel social categories.

Is there overlap between self and ingroup biases?

Analysis plan.

Results from Experiments 1 and 3 showed that response times (for matched pairs) were enhanced for own (self and ingroup) shapes compared to other (stranger and outgroup) shapes. Correlation analyses were used to measure the relationship between the self- and ingroup-biases in order to elucidate whether shared mechanisms drive these prioritisation effects. As d prime scores combine matched and mismatched trials and the main bias effects occurred only for matched trials, only response time data was used for the subsequent analyses.

The data from Experiments 1 and 3 were combined because the pattern of results was identical. Difference in response times between self and stranger and between ingroup and outgroup were taken as a measure of the magnitude of the self- and ingroup-biases. Only matched trials were included in calculating the bias scores for response time because the self- and ingroup-advantages in response time were only observed within these. Scores for own shapes were deducted from scores for other shapes (as lower scores indicate better performance) and the relationships between the self- and ingroup-advantages were then measured using Pearson bivariate correlations.

Two data-points were identified as showing extremely low self- and ingroup-bias scores (below 2.5SD from the group mean) in response time. These outliers were excluded for the following analysis 49 , 50 . The total N for the subsequent analyses was 67.

There was a significant positive relationship between the self- and ingroup-advantages in response time, r (65) = 0.467 [0.275, 0.636], p  < 0.001. When split into the two separate experiments, this effect held for both Experiment 1, r (44) = 0.368 [0.097, 0.599], p  = 0.012, and also Experiment 3, r (19) = 0.612 [0.323, 0.819], p  = 0.003 (Fig.  5 ).

Overlap in processing advantages for the self and social ingroups. Pearson bivariate correlations showed a positive relationship between self and ingroup advantages in RT across Experiments 1 and 3 combined (a) ; Experiment 1 (b) and Experiment 3 (c) .

General discussion

Across three experiments, we examined the effects of minimal group allocation on low-level biases in associative matching. A robust advantage in response time and sensitivity for stimuli associated with the self and the ingroup was found when self and ingroup associations were responded to in separate tasks (Experiment 1) and in separate blocks of the same task (Experiment 3). However, when self and ingroup associations were presented within the same task blocks, the prioritisation effects were only observed for stimuli associated with the self (Experiment 2). Participants also tended to adopt a lower response criterion for own shapes compared to other shapes, showing they were generally more likely to indicate a shape was own- than other-associated in Experiments 1 and 3, although this bias only held for personal associations in Experiment 2. There was no ingroup-advantage effect when self and ingroup stimuli were present in the same task blocks, but higher error rates for self shapes paired with ingroup than with outgroup labels indicated the possibility of a shared cognitive representation of newly-learned self and (novel) ingroup stimuli and show a closer association between self and novel ingroup than self and novel outgroup. Further to this, there were positive correlations between the self and ingroup response time biases in both Experiments 1 and 3, suggesting the possibility of a self-anchoring mechanism driving the prioritisation effects for new social ingroups.

As noted throughout these experiments, the prioritisation effects found for novel ingroup stimuli within the matching task complement several studies that find ingroup favouritism effects in the minimal group paradigm across a range of outcome measures 51 . Notably, the results extend previous research showing processing advantages for novel ingroup faces 36 . Our results demonstrate that minimal group-bias extends even when no obvious social content (e.g., faces) is present. It may be that this fundamental bias in categorising social information may serve as the foundation for higher-level ingroup bias effects (e.g., a perceptual model of intergroup relations 52 ). For example, greater cognitive resources allocated to ingroup stimuli may then cause better differentiation of ingroup and outgroup members in facial processing. The present findings significantly broaden these empirical findings and suggest even minimal group allocation provides a motivating factor in directing attention and enhancing cognitive performance.

Experiments 1 and 3 demonstrated a minimal group effect but this was eliminated when self stimuli were present within the same task blocks (Experiment 2). An ingroup-bias was demonstrated previously in the same design as Experiment 2 using an established group manipulation of college rowers 28 . This discrepancy indicates a possible difference in the mechanisms behind novel and established ingroup-prioritisation. It is likely that established social groups contain inherently more meaning and personal significance than novel ones which enables them to remain a priority even when self stimuli are also present. Novel groups, however, may not hold enough weight to retain attention when task demands increase and the self competes for attention. This is supported in a recent study 27 , which found significantly reduced ingroup-prioritisation compared to self-prioritisation when groups were randomly and arbitrarily assigned with no information given about the other group members. This point highlights an important methodological consideration: if novel group-biases do not always follow the same patterns as established group-biases, then studies utilising minimal group paradigms hold the potential of showing a different pattern of responses. It would be useful for further research to delve into the specific mechanisms behind the processing of established and novel ingroup stimuli.

Both Experiments 1 and 3 provided evidence for a positive relationship between the self- and novel ingroup-advantage in response time, indicating the possibility of a shared process driving the advantage effects. Although conclusions that can be drawn are limited in that data is correlational only, this is consistent with experimental evidence and theory finding ingroups are represented as part of the self 48 , 53 , 54 , 55 . This dovetails with work on self-anchoring 38 , 39 , which suggests ingroup favouritism arises, at least in some cases and particularly in the case of minimal groups, by the projection of the self onto the group, contrasting to the traditional social identity approach, which explains ingroup-bias in terms of self-stereotyping 37 . It is important to note that prior work found no relationship between biases for the self and high reward or positive valence 56 , 57 when measured by an associative matching paradigm in a similar way. Thus, it is not likely that results are explained simply by individual differences in effectiveness of prioritisation systems more generally as there do not appear to be relationships between self-bias and other goal-relevant categories. The relationship observed presently, then, implies a conceptual connection between self- and novel ingroup-prioritisation.

Original research for self-anchoring utilised a trait judgement paradigm in which it was found that participants ascribed characteristics to novel ingroups based on information about the self 39 . The present study also implies that ingroup-bias arises from the projection of self significance. This may serve as a cognitive heuristic that allows a gap to be filled from the known (the self) to the unknown (the novel ingroup) 38 .

The three experiments benefitted from replicating previous findings using a minimal group paradigm in which self- and ingroup-prioritisation measures were exactly equal. However, there are some limitations to the design. It must be noted that in response time data for Experiments 1 and 3, it is difficult to completely disentangle effects of shape and label for mismatching pairs. This is because, for example, a stranger mismatched shape is by definition presented with a self label and a self mismatched shape is presented with a stranger label. Thus, it could be argued that faster responses for stranger than self mismatched shapes could be interpreted simply as faster responses for self than stranger labels. However, our dprime data strengthen the importance of interpreting results by shape (not label), as here match and mismatch trials by shape are combined to give a measure of overall sensitivity towards the shape. Additionally, in Experiment 2, mismatching shapes may have been presented with any one of the possible three mismatching labels. Prior research too provides grounds for focussing on the importance of shape category over label. For example, one study found biases for ingroup shapes when they were presented on sequential screens to label cues, ensuring responses were to shapes and not labels 29 . Another more recent study found self bias effects in the complete absence of familiar labels. Here, labels were replaced with abstract symbols before the experimental task started 26 , showing self bias effects did not depend on familiar social labels. Future work could employ these approaches to further separate the role of social labels in minimal group designs.

Conclusions

Here we provide novel evidence for minimal group effects on the processing of neutral stimuli. Further, positive relations between self- and novel ingroup-biases provide support for self-anchoring: the projection of the self onto novel social groups to fill a cognitive gap between the known and the unknown. These findings could have important implications for understanding how social importance biases our perception of the world. Even arbitrary group allocation can quickly alter subsequent attentional allocation of stimuli associated with those new groups, and this could have consequences for many real-world social contexts in which categorisation occurs.

Data availability

Data will be made available on request.

Farnham, S. D., Greenwald, A. G. & Banaji, M. R. Implicit self-esteem. in Social Identity and Social Cognition 230–248 (Blackwell Publishing, 1999).

Pelham, B. W., Carvallo, M. & Jones, J. T. Implicit egotism. Curr. Dir. Psychol. Sci. 14 , 106–110 (2005).

Article   Google Scholar  

Beggan, J. K. On the social nature of nonsocial perception: The mere ownership effect. J. Pers. Soc. Psychol. 62 , 229–237 (1992).

Knetsch, J. L. The endowment effect and evidence of nonreversible indifference curves. Am. Econ. Rev. 79 , 1277–1284 (1989).

Google Scholar  

Knetsch, J. L. & Sinden, J. A. Willingness to pay and compensation demanded: Experimental evidence of an unexpected disparity in measures of value. Q. J. Econ. 99 , 507–521 (1984).

Rogers, T. B., Kuiper, N. A. & Kirker, W. S. Self-reference and the encoding of personal information. J. Pers. Soc. Psychol. 35 , 677 (1977).

Article   CAS   PubMed   Google Scholar  

Cunningham, S. J., Turk, D. J., Macdonald, L. M. & Neil Macrae, C. Yours or mine? Ownership and memory. Conscious. Cogn. 17 , 312–318 (2008).

Symons, C. S. & Johnson, B. T. The self-reference effect in memory: A meta-analysis. Psychol. Bull. 121 , 371–394 (1997).

Turk, D. J., Cunningham, S. J. & Macrae, C. N. Self-memory biases in explicit and incidental encoding of trait adjectives. Conscious. Cogn. 17 , 1040–1045 (2008).

Article   PubMed   Google Scholar  

Keyes, H., Brady, N., Reilly, R. B. & Foxe, J. J. My face or yours? Event-related potential correlates of self-face processing. Brain Cogn. 72 , 244–254 (2010).

Perdue, C. W., Dovidio, J. F., Gurtman, M. B. & Tyler, R. B. Us and them: Social categorization and the process of intergroup bias. J. Pers. Soc. Psychol. 59 , 475 (1990).

Johnson, J. D. et al. Rodney King and O. J. revisited: The impact of race and defendant empathy induction on judicial decisions. J. Appl. Soc. Psychol. 32 , 1208–1223 (2002).

Nosek, B. A., Greenwald, A. G. & Banaji, M. R. The implicit association test at age 7: A methodological and conceptual review. 28.

Hugenberg, K., Young, S. G., Bernstein, M. J. & Sacco, D. F. The categorization-individuation model: An integrative account of the other-race recognition deficit. Psychol. Rev. 117 , 1168–1187 (2010).

Shriver, E. R., Young, S. G., Hugenberg, K., Bernstein, M. J. & Lanter, J. R. Class, race, and the face: Social context modulates the cross-race effect in face recognition. Pers. Soc. Psychol. Bull. 34 , 260–274 (2008).

Correll, J. & Park, B. A model of the ingroup as a social resource. Personal. Soc. Psychol. Rev. Off. J. Soc. Personal. Soc. Psychol. Inc 9 , 341–359 (2005).

Hewstone, M., Rubin, M. & Willis, H. Intergroup bias. Annu. Rev. Psychol. 53 , 575–604 (2002).

Sui, J., He, X. & Humphreys, G. W. Perceptual effects of social salience: Evidence from self-prioritization effects on perceptual matching. J. Exp. Psychol. Hum. Percept. Perform. 38 , 1105–1117 (2012).

Frings, C. & Wentura, D. Self-priorization processes in action and perception. J. Exp. Psychol. Hum. Percept. Perform. 40 , 1737–1740 (2014).

Macrae, C. N., Visokomogilski, A., Golubickis, M., Cunningham, W. A. & Sahraie, A. Self-relevance prioritizes access to visual awareness. J. Exp. Psychol. Hum. Percept. Perform. 43 , 438 (2017).

Mattan, B., Quinn, K. A., Apperly, I. A., Sui, J. & Rotshtein, P. Is it always me first? Effects of self-tagging on third-person perspective-taking. J. Exp. Psychol. Learn. Mem. Cogn. 41 , 1100 (2015).

Sui, J., Sun, Y., Peng, K. & Humphreys, G. W. The automatic and the expected self: Separating self- and familiarity biases effects by manipulating stimulus probability. Atten. Percept. Psychophys. 76 , 1176–1184 (2014).

Sui, J., Liu, M., Mevorach, C. & Humphreys, G. W. The salient self: The left intraparietal sulcus responds to social as well as perceptual-salience after self-association. Cereb. Cortex N. Y. N 1991 (25), 1060–1068 (2015).

Yin, S., Sui, J., Chiu, Y.-C., Chen, A. & Egner, T. Automatic prioritization of self-referential stimuli in working memory. Psychol. Sci. https://doi.org/10.1177/0956797618818483 (2019).

Lockwood, P. L. et al. Neural mechanisms for learning self and other ownership. Nat. Commun. 9 , 4747 (2018).

Article   ADS   CAS   PubMed   Google Scholar  

Woźniak, M. & Knoblich, G. Self-prioritization of fully unfamiliar stimuli. Q. J. Exp. Psychol. 72 , 2110–2120 (2019).

Constable, M. D., Elekes, F., Sebanz, N. & Knoblich, G. Relevant for us? We-prioritization in cognitive processing. J. Exp. Psychol. Hum. Percept. Perform. 45 , 1549–1561 (2019).

Enock, F., Sui, J., Hewstone, M. & Humphreys, G. W. Self and team prioritisation effects in perceptual matching: Evidence for a shared representation. Acta Psychol. (Amst.) 182 , 107–118 (2018).

Moradi, Z., Sui, J., Hewstone, M. & Humphreys, G. W. In-group modulation of perceptual matching. Psychon. Bull. Rev. 22 , 1255–1277 (2015).

Moradi, Z. et al. The rival doesn’t catch my eyes: In-group relevance modulates inhibitory control over anti-saccades. Vis. Cogn. (2017).

Allen, V. L. & Wilder, D. A. Categorization, belief similarity, and intergroup discrimination. J. Pers. Soc. Psychol. 32 , 971–977 (1975).

Brewer, M. B. & Silver, M. Ingroup bias as a function of task characteristics. Eur. J. Soc. Psychol. 8 , 393–400 (1978).

Mullen, B., Brown, R. & Smith, C. Ingroup bias as a function of salience, relevance, and status: An integration. Eur. J. Soc. Psychol. 22 , 103–122 (1992).

Tajfel, H. Experiments in intergroup discrimination. Sci. Am. 223 , 96–103 (1970).

Article   ADS   Google Scholar  

Tajfel, H., Billig, M. G., Bundy, R. P. & Flament, C. Social categorization and intergroup behaviour. Eur. J. Soc. Psychol. 1 , 149–178 (1971).

Bernstein, M. J., Young, S. G. & Hugenberg, K. The cross-category effect: Mere social categorization is sufficient to elicit an own-group bias in face recognition. Psychol. Sci. 18 , 706–712 (2007).

Tajfel, H. & Turner, J. An integrative theory of intergroup conflict. in The Social Psychology of Intergroup Relations (eds. Austin, W. G. & Worchel, S.) 33–47 (Brooks/Cole Pub. Co, 1979).

van Veelen, R., Otten, S., Cadinu, M. & Hansen, N. An integrative model of social identification: Self-stereotyping and self-anchoring as two cognitive pathways. Personal. Soc. Psychol. Rev. 20 , 3–26 (2016).

Cadinu, M. & Rothbart, M. Self-anchoring and differentiation processes in the minimal group setting. J. Pers. Soc. Psychol. 70 , 661–677 (1996).

Otten, S. & Epstude, K. Overlapping mental representations of self, ingroup, and outgroup: Unraveling self-stereotyping and self-anchoring. Pers. Soc. Psychol. Bull. 32 , 957–969 (2006).

Moradi, Z. Z., Manohar, S., Duta, M., Enock, F. & Humphreys, G. W. In-group biases and oculomotor responses: Beyond simple approach motivation. Exp. Brain Res. 236 , 1347–1355 (2018).

Ratcliff, R. Methods for dealing with reaction time outliers. Psychol. Bull. 114 , 510–532 (1993).

Green, D. M. & Swets, J. A. Signal Detection Theory and Psychophysics Vol. 1 (Wiley, New York, 1966).

Macmillan, N. A. Signal detection theory as data analysis method and psychological decision model. (1993).

Ashburn-Nardo, L., Voils, C. I. & Monteith, M. J. Implicit associations as the seeds of intergroup bias: How easily do they take root?. J. Pers. Soc. Psychol. 81 , 789–799 (2001).

Otten, S. & Wentura, D. About the impact of automaticity in the minimal group paradigm: Evidence from affective priming tasks. Eur. J. Soc. Psychol. 29 , 1049–1071 (1999).

Sheng, F. & Han, S. Manipulations of cognitive strategies and intergroup relationships reduce the racial bias in empathic neural responses. NeuroImage 61 , 786–797 (2012).

Turner, J. C., Hogg, M. A., Oakes, P. J., Reicher, S. D. & Wetherell, M. S. Rediscovering the Social Group: A Self-Categorization Theory . (Basil Blackwell, 1987).

Barnett, V. & Lewis, T. Outliers in Statistical Data . (Wiley, 1978).

Lovie, P. & Lovie, A. D. (Eds). New Developments in Statistics for Psychology and the Social Sciences : Vol 2. (Routledge, London, 1991.).

Diehl, M. The minimal group paradigm: Theoretical explanations and empirical findings. Eur. Rev. Soc. Psychol. 1 , 263–292 (2011).

Xiao, Y. J., Coppin, G. & Bavel, J. J. V. Clarifying the role of perception in intergroup relations: Origins of bias, components of perception, and practical implications. Psychol. Inq. 27 , 358–366 (2016).

Coats, S., Smith, E. R., Claypool, H. M. & Banner, M. J. Overlapping mental representations of self and in-group: Reaction time evidence and its relationship with explicit measures of group identification. J. Exp. Soc. Psychol. 36 , 304–315 (2000).

Smith, E. R., Coats, S. & Walling, D. Overlapping mental representations of self, in-group, and partner: Further response time evidence and a connectionist model. Pers. Soc. Psychol. Bull. 25 , 873–882 (1999).

Smith, E. R. & Henry, S. An in-group becomes part of the self: Response time evidence. Pers. Soc. Psychol. Bull. 22 , 635–642 (1996).

Stolte, M., Humphreys, G., Yankouskaya, A. & Sui, J. Dissociating biases towards the self and positive emotion. Q. J. Exp. Psychol. 2006 (70), 1–34 (2015).

Sui, J. & Humphreys, G. W. The interaction between self-bias and reward: Evidence for common and distinct processes. Q. J. Exp. Psychol. 68 , 1952–1964 (2015).

Download references

Acknowledgements

This work was supported by a Medical Research Council Fellowship (MR/P014097/1), a Christ Church Junior Research Fellowship and a Christ Church Research Centre Grant to P. L. L. The Wellcome Centre for Integrative Neuroimaging is supported by core funding from the Wellcome Trust (203139/Z/16/Z). The work was completed in partial fulfilment of a PhD funded through a DSTL studentship by the first author. We thank the late Professor Glyn Humphreys, who initially supervised the wider PhD project that this work formed a part of, for his helpful discussions and input into this research.

Author information

These authors contributed equally: Patricia L. Lockwood and Jie Sui.

Authors and Affiliations

Department of Experimental Psychology, University of Oxford, Oxford, OX2 6GG, UK

Florence E. Enock, Miles R. C. Hewstone & Patricia L. Lockwood

Department of Psychology, University of York, York, YO10 5DD, UK

Florence E. Enock

Wellcome Centre for Integrative Neuroimaging, University of Oxford, Oxford, UK

Patricia L. Lockwood

Centre for Human Brain Health, School of Psychology, University of Birmingham, Birmingham, UK

School of Psychology, University of Aberdeen, Aberdeen, AB24 3RT, UK

You can also search for this author in PubMed   Google Scholar

Contributions

F.E., M.H. and J.S. designed the studies, F.E. collected the data, F.E., P.L. and J.S. analysed the data and F.E., M.H., P.L. and J.S. wrote the paper. All authors approved the final version submitted for publication.

Corresponding authors

Correspondence to Florence E. Enock , Patricia L. Lockwood or Jie Sui .

Ethics declarations

Competing interests.

The authors have no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Enock, F.E., Hewstone, M.R.C., Lockwood, P.L. et al. Overlap in processing advantages for minimal ingroups and the self. Sci Rep 10 , 18933 (2020). https://doi.org/10.1038/s41598-020-76001-9

Download citation

Received : 09 February 2020

Accepted : 21 August 2020

Published : 03 November 2020

DOI : https://doi.org/10.1038/s41598-020-76001-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Overlap between mental representations of nationalities modulates perceptual matching.

• Wanghuan Lin
• Zhixin Zhang
• Chunliang Feng

Current Psychology (2023)

Levels of Self-representation and Their Sociocognitive Correlates in Late-Diagnosed Autistic Adults

• R. L. Moseley

Journal of Autism and Developmental Disorders (2022)

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

• Previous Article
• Next Article

The Social Identity Approach

Group identities as social realities, intergroup behavior, social reality and ingroup bias, normative influence and social projection, moral considerations, conclusions, statement of ethics, conflict of interest statement, funding sources, group identity and ingroup bias: the social identity approach.

• Split-Screen
• Article contents
• Figures & tables
• Supplementary Data
• Peer Review
• Open the PDF for in another window
• Get Permissions
• Cite Icon Cite
• Search Site

Maykel Verkuyten; Group Identity and Ingroup Bias: The Social Identity Approach. Human Development 14 December 2021; 65 (5-6): 311–324. https://doi.org/10.1159/000519089

Download citation file:

• Ris (Zotero)
• Reference Manager

This article discusses the social identity approach (social identity theory and self-categorization theory) for understanding children’s ingroup biases in attitudes and behaviors. It is argued that developmental research on ingroup bias will be enhanced by more fully considering the implications of this approach. These implications include (a) the conceptualization of group identity, (b) the importance of social reality and children’s epistemic motivation, (c) the role of processes of normative influence and social projection, and (d) the relevance of moral considerations. These four implications have not been fully considered in the developmental literature but indicate that the social identity approach offers the possibility for theoretically integrating and empirically examining various processes involved in children’s ingroup biases.

I watched what had been marvelous, cooperative, wonderful children, turn into nasty, vicious, discriminating, little third-graders in a space of fifteen minutes. I think I learned more from the superior children – from the children who were considered to be superior – than I did from the children who were considered inferior, because their personalities changed even more than the others did. Whether they are this, whether this is what they would like to be inside but society inhibits them, I don’t know, but for one day we removed their inhibitions, and they were ghastly. Jane Elliot

In 1968, one day after the assassination of Martin Luther King Jr., Jane Elliott, a primary schoolteacher in Iowa, decided to make the children in her elementary school class experience discrimination and racism. She carried out a famous two-day classroom intervention, which was recorded in the award-winning documentary The Eye of the Storm (Peters, 1970). All of her students were white and, to teach them something about race and racism, the documentary shows Elliott asking the class whether there is something that makes the children different from each other. One of the children suggested the color of their eyes. Elliott adopted the idea and divided the class into children with blue eyes and children with brown eyes. To make the distinction clearly visible, she had the two groups wear collars of a different color. This process of classification changed the situation from one in which there were individual pupils to a situation of two “racial” groups. The quote above describes what according to Elliott subsequently happened and why this might be.

Elliott’s exercise is not a scientific study, raises various ethical issues, and academic research does not support her interpretations (e.g., “change of personality”). However, throughout the paper I will use Elliott’s exercise and the documentary only for illustrative purposes as it is very helpful for explaining, in a concrete way, what is involved in group identities and how these affect relations between children. Theoretically, I will discuss the social identity approach which includes social identity theory and self-categorization theory (Reicher et al., 2010). The original minimal group experiments in which participants are randomly placed in ad hoc groups and that formed the basis of this approach were first conducted with Dutch early adolescents (Rabbie & Horowitz, 1969) and then British school children (Tajfel et al., 1971). The social identity approach, however, is not a developmental theory but a social psychological perspective. Yet, developmental social psychologists have provided developmental specifications to the social identity approach (e.g., Killen & Rutland, 2011; Nesdale, 2017; Verkuyten, 2016), and this approach has been increasingly used to account for children’s ethnic, racial, national, gender, novel and other group biases in different settings and contexts (e.g., Abrams et al., 2008; Dunham et al., 2011; Spielman, 2000). Much of this research has been conducted among majority group children (up to around 12 years) and in the context of the United States and Western Europe, and this limitation is reflected in the research that will be discussed.

I will first give a short overview of the main points of the social identity approach which is followed by a discussion of three interrelated aspects of the conceptualization of group identities and their public nature. Then I will elaborate on what is involved in intergroup behavior or children’s thinking, feeling and acting in terms of their group membership and in relation to other groups. This is followed by a discussion of ingroup bias and the striving for a positive group self in relation to reality constraints, processes of normative influence and social projection, and moral considerations. The overall aim is to demonstrate the broad usefulness of the social identity approach for understanding children’s ingroup biases, by discussing the nature of group identities and by focusing not only on the importance of establishing a positive group self but also on other processes, and epistemic motivation in particular (Piaget, 1929).

The social identity approach focuses on explaining attitudes and behavior as the outcome of the interaction between psychological processes with social, cultural and political circumstances. It is emphasized that the way in which psychological processes play out is dependent upon how the social world is structured. There are three features of this approach that are especially relevant for understanding children’s ingroup biases and that I will discuss and develop further in the various sections of the paper.

First, the social identity approach focuses on the processes involved in making group distinctions and the ways in which people define themselves and others as members of social groups. Group identity is simultaneously social and individual, public and private. It is considered a key construct for conceptualizing the relationship between the individual and society: identity is “the best device I know for bringing together ‘public issues’ and ‘private troubles’” (Jenkins, 1997, p. 3). Group identity involves public images (social distinctions) that inform private understandings (sense of self). What a racial, ethnic or other group identity means cannot be reduced to a person’s own perspective and idiosyncratic beliefs since it involves social reality in the form of shared meanings that are enshrined in, for example, rules, regulations, symbols, collective representations and cultural narratives (Verkuyten, 2018). This means that children have to learn what these public images are and gradually have to develop a sense of self based upon the images of the groups to which they belong.

Second, according to the social identity approach, group identity is what makes group behavior possible because it fundamentally changes and transforms people’s psychology and behavior. People think, feel and behave not only as individuals (“I”; e.g., personal self-esteem, personal interests) but also as group members with shared perceptions, understandings and goals (“we”; e.g., collective self-esteem, collective interests). With the act of defining oneself as a group member, the normative group understandings become self-relevant. There is a process of self-stereotyping in which one starts to understand oneself and others in terms of the norms, beliefs and values associated with that particular group identity.

Third, group identities provide a sense of ingroup belonging with the related tendency to seek positive group distinctiveness along valued dimensions of intergroup comparison. Because of their need for a positive group self, people tend to show ingroup bias. They try to enhance their sense of self-esteem by positively differentiating their ingroup from relevant outgroups on those dimensions that matter to their group. However, according to the social identity approach and the related social identity developmental theory (Nesdale, 2017), such a bias is by no means an automatic product of group distinctions. In addition to the strength of group identification and perceived outgroup threat (Nesdale, 2017), ingroup bias depends on social reality, prevailing social norms and moral considerations. Children do not only want to have a positive group self, but also want to develop an adequate understanding of social reality, are influenced by the norms of their group and try to do the morally right thing. Ingroup bias is multiply determined, and multiple explanations for bias in the real world, but also in minimal group experiments, co-exist (Spears & Otten, 2012). An explanation in terms of striving for a positive group self does not dismiss an explanation in terms of the epistemic desire to form an adequate understanding of social reality, a normative explanation, or a moral explanation (Abrams et al., 2008; Killen et al., 2016). Similar to adults, children’s intergroup behavior is likely to be driven by multiple motivations that interact with the tendency to make a positive distinction in favor of one’s own group.

In growing up, children increasingly become aware of the social distinctions that matter in society. They live in a pre-structured social world in which various differences between people are considered relevant, made meaningful in particular ways, and used to label individuals and interpret their behavior. The particular features and their meanings differ between societies, cultures and historical periods but distinctions based on, for example, age, gender, religion, language, race and ethnicity are common in most settings. From the beginning, infants are strongly interested to learn about how the social world is organized and made meaningful: what matters to society matters to them for knowing what the world is like and where they and others fit in and belong. For example, infants use language and accents as social markers and intuitively use native speakers as a particularly good source of culturally relevant information (Kinzler, 2021). Further, children as young as 4 make inferences about unknown groups only when generic information is communicated by a speaker who is considered to know a lot about these groups (Moty & Rhodes, 2021).

Group identities are about socially defined and recognized distinctions and designations. They imply intersubjective understandings that conceptually involve three interrelated components (Verkuyten, 2018): (a) social classification, or the sociostructural component, (b) specific behavioral and normative consequences and expectations bounded to the category, or the cultural component, and (c) judgements of an ontological nature , or the psychological component. Together, these three are the necessary and closely linked public aspects of what defines a group identity. The distinction between the three aspects is a conceptual, analytic one that allows us to ask specific questions and to examine in detail how the social world is organized and the ways in which people come to understand their social world. However, in everyday life the three aspects are closely interrelated and intermingled, and the importance of making an analytic distinction between the three is not always easy to see. The Jane Elliott example is useful here because it involves the formation of new group identities, based on eye color.

Social Classification

First of all, there is the social classification of people into categories or groups. Children are predisposed to see the world in categories and attuned to cues that mark social difference. The social world is organized in many different ways, and these organizations are learned from an early age on and used to make distinctions between individuals. Group identity is a public reality – an assignment that divides people on the basis of particular criteria that can be used in different combinations including forms of intersectionality. Hence, people can be included or excluded on many different grounds and children learn that individuals can be categorized in different ways based on cues that are socially meaningful in a particular context.

Categorizing people, and hence differentiating between them, happens everywhere in society and for a variety of reasons. It occurs in everyday life and is a prerequisite for group functioning and the organization of society because people differ in their positions, capabilities, life circumstances, needs and desires. Making distinctions between groups of people is often functional and relevant, but also forms the basis for ingroup bias from a young age onwards (Bigler & Liben, 2007; Liberman et al., 2017). However, this does not have to imply outgroup negativity: making distinctions does not necessarily imply creating oppositions (Nesdale, 2017). Already infants show a preference for ingroup members because these members are considered similar to themselves, are often more familiar and can provide relevant group information, whereas outgroup dislike tends to develop around the age of 6 (Buttelmann & Böhm, 2014; Cameron et al., 2001), although it can develop earlier in deeply divided societies (Nasie et al., 2016).

Elliott used a visible classification that changed the classroom setting to a situation of two “racial” groups. She construed the relevant categories in a specific way, and she could have done so in another way. The tensions and conflicts that did arise in the classroom are not the inevitable consequences of creating a categorial distinction. For example, she could have used another, less visible, feature to make a distinction between groups of children (e.g., hobbies) and lower perceptual salience leads to less ingroup bias among children (e.g., Bigler et al., 1997). Or, she could have made a tripartite distinction between blue, brown and green eyes, or combine two categorization dimensions as a form of intersectionality (e.g., eye color and gender). The children’s understanding of themselves and each other might have been quite different and the intergroup processes are likely to be different in a “blue, brown, green” world compared to a binary “blue, brown” world. Similar to Elliott, the great majority of research on children’s ingroup biases uses a binary framework and, in doing so, tends to ignore more complex situations. Some experimental research among children that does use a third group has found only signs of ingroup bias within a competitive setting (e.g., Hartstone & Augoustinos, 1995; Spielman, 2000). Additionally, experimental work combining two categorization dimensions (e.g., gender and ethnicity) has found lower ingroup bias and more complex patterns of intergroup differentiation with, for example, intersecting categories (ingroup member on one dimension – e.g., gender – and an outgroup on the other dimension – e.g., ethnicity) being evaluated more positively than those who differ on both categorization dimensions (e.g., Brewer et al., 1987; Vanbeselaere, 1987). Following the social identity approach, one explanation for these findings is that a third group as well as the combination of categorization dimensions makes a binary distinction less salient and adequate which can influence responses to outgroup members.

Behavioral and Normative Meanings

A meaningful group identity does not only involve social classification, but also stereotypical expectations and behavioral consequences that provide shared content and meaning. Group identities are associated with particular (implicit and explicit) scripts , or socially recognized cultural meanings and norms of behavior that provide guidelines for how to think, feel and act in the social world. Even infants expect group members to behave in similar ways (Powell & Spelke, 2013), and children (4-years and older) prefer to learn from informants in consensus with one another (Chen et al., 2013) and use information about how a group is for making inference about how individual group members will and should be (Roberts et al., 2021).

Because eye color does not ordinarily have any categorical significance, this distinction can be used to make it meaningful in all sorts of ways. In her classroom, Elliott proceeded to define and treat the one group as if they were inferior people (and the other, superior) and the next day she subjected the other group to the same treatment. She did this by defining clear behavioral consequences. For example, the “inferior” children had to sit at the worst places in the back of the class, had to go last in line, were not allowed to use paper cups to drink from the water fountain, got no extra minutes at recess and were not allowed to go for seconds in the cafeteria. Further, she defined the two groups as deserving their inferior or superior treatment because of their alleged intrinsic differences (“The brown-eyed people are the better people in this room. They are cleaner and they are smarter, more civilized”). Thus, the categorization was made normative, and having blue or brown eyes was no longer an insignificant feature, but became consequential and meaningful (“We are intelligent and hence we look down upon them as stupid”).

However, Elliott could have made the same blue-brown distinction meaningful in another way with other behavioral implications. Rather than “superior” and “inferior,” she could have defined the distinction in terms of “being different” and “being complementary.” If she wanted to teach the children something about cooperation and the positive aspects of diversity, such a definition would be appropriate (“We are different and therefore can learn from them”). Although the same categorical distinction would be used (eye color), the consequences for the relationships within the class would be dramatically different and probably in line with the positive implications of multiculturalism for ingroup bias (Verkuyten & Thijs, 2013). Yet, in the case of “superior” and “inferior,” the situation can still develop in various directions depending on the particular meanings involved. “Superior” can also mean that you are expected to help the “inferior” children as much as possible and to be concerned about their ups and downs (“We are charitable and therefore help them”). Ingroup meanings can be prosocial and Elliott could have introduced such an understanding rather than define the group distinction in terms of unequal rights. Again, the consequences would have been different from those seen in the documentary: prosocial norms have been found to instigate more positive outgroup attitudes (e.g., Nesdale & Lawson, 2011) and to stimulate children’s helping behavior (e.g., Sierksma et al., 2014). The “superior” children can still feel valued and positive about themselves, but in this case because they belong to a group that help those in need of support. And the “inferior” children will feel accepted and cared for rather than excluded and rejected.

Experimental research tends to examine children’s ingroup bias using ad hoc minimal groups in which the division into two categories is made self-relevant and becomes all-important, the only available ground for making distinctions in favor of one’s ingroup. However, the social identity approach argues that attention to the contents of identity is needed for understanding how people in everyday life will respond toward others. Categorical distinctions trigger processes of category differentiation but the beliefs, norms and values associated with a particular identity determine the direction and nature of people’s reactions.

Ontological Nature

Labels for group identities are powerful not only because of the related behavioral expectations and consequences, but also because of the way in which the members of these categories are thought to be like as human beings. The ontological component implies that not all social classifications have to function as group identities in everyday life. For instance, in schools, teachers make use of all sorts of group distinctions for various educational purposes (e.g., children working on different tasks) but these distinctions do not necessarily indicate or develop into social identities that are considered to say something about the nature of groups of children. Social distinctions with specific behavioral consequences sometimes lack the third ontological component since in everyday life these distinctions do not function for defining and indicating what a person is like.

With group identities, knowing to which category someone belongs leads to a judgement about the kind of person they are. Group identities are expressed in the language of nouns that “cut slices” through our social environment (Allport, 1954; Rosenberg, 1979). Someone is a Turk, a Mexican or a Basque. In using such words, no quantifiable qualities are defined, as with adjectives. Descriptions of oneself or others as being religious, intelligent, shy or athletic do not necessarily indicate group identities because they can refer to personal beliefs and personal characteristics. But these same terms might also be used to indicate group identities when referring to membership of a religious group or the category of intellectuals, introverts or athletes. For the social identity approach, the critical issue is whether these terms are used to define people as part of a social group and therefore subject to the related behavioral expectations and beliefs about the nature of the people involved.

Nouns reflect separate social types, with typical characteristics and boundaries that suggest a more than superficial difference. Stating that one is Turkish or Jewish (adjectives) or that one is a Turk or a Jew (nouns) indicates a different degree of essentialist quality (Carnaghi et al., 2008; Gelman & Heyman, 1999). Already, preschoolers tend to use linguistic labels for kind-based reasoning and develop essentialist conceptions of people related to gender, and depending on the cultural context of race and ethnicity (e.g., Diesendruck et al., 2013; Waxman, 2010). They assume that people who are marked by the same label share an important category membership and are similar to each other. With nouns, one group of people is deemed to be like this, the other, like that. With that, they become real identities with real social and psychological consequences. The labelling and functional usage of novel groups has been found to contribute to the formation of ingroup bias among young children (Bigler et al., 1997; Billig & Tajfel, 1973).

Elliott turned the distinction into real group identities by starting to use the nouns “blue-eye” and “brown-eye,” and in class and on the playground the children themselves started to use these labels for describing and treating others. The labels became informative about the nature of the two groups of children. For example, when the class went to lunch, one “superior” child suggested that Elliott should inform the canteen staff about the rule that the “inferior” children may not be given extra portions and another child suggested that the teacher should keep her pointer on hand in case the “inferior” children were getting out of hand. And during class, the answers given by the “superior” children were interpreted in terms of intelligence and the ability to learn quickly, while the behavior of the “inferior” children was attributed to clumsiness and stupidity: when a blue-eyed student got an arithmetic problem wrong, a brown-eyed pupil said “Well, what do you expect from him, Mrs. Elliott, he’s a bluey!” Another example that shows that the categories had become real group identities that pupils used to define and understand themselves and others even in the absence of the teacher occurred after recess on the day when the brown-eyed children were considered inferior. One child was withdrawn and agitated because he had gotten into a fight and Elliott asked what had happened:

Ms. Elliott:What happened, John?

John:He called me names.

Ms. Elliott:What did he call you?

John:Brown-eye.

Pupils:They always call us that. “Hey brown-eye, comehere, brown-eye.”

Ms. Elliott:What’s wrong with being called a brown-eye?

John:It means we’re stupid and all that.

In this excerpt several pupils use the term “they” and in the last line John refers to “we.” Elliott had changed the classroom setting of individual pupils into an intergroup situation with two meaningful “racial” groups. The children started to think, feel and act in terms of “us and them” (“I feel that Ms. Elliott has taken our best friends away from us”). In the opening quote, Elliott talks about a change in personality within the space of 15 min, but from the social identity approach there is a (temporary and situational) change from personal self-understanding (“I”) to collective self-understanding (“we”).

The children started to label themselves in terms of the category to which they belong (“brown-eye”) and distinguished themselves from the category to which they did not belong (“blue-eye”). Research has demonstrated that the process of self-categorization and the ability to talk and think in terms of “us and them” develops early in life. Whereas emotional identification with ingroup members seems to start to develop around 6 years of age, the ability to socially categorize oneself and others develops earlier (Ruble et al., 2004; Sani & Bennett, 2004). Children quickly learn to which category they belong, and this is a sufficient precondition for ingroup bias (Bennett et al., 1998; Dunham, 2018).

Furthermore, children will start to think, feel and behave in agreement with the way in which the categories are socially meaningful. In Elliott’s classroom, the one group started to feel superior and the other group inferior, and the children started to act in terms of how the groups were understood. When you are one of the better people and better means having more rights and privileges, then it makes good sense to behave in such a way. In other words, one can argue that within the logic of the situation as defined, the children behaved rationally, reality oriented. However, the children also started to use the labels in name calling, and became arrogant, bossy and otherwise unpleasant to their “inferior” classmates for feeling positive about themselves. This suggests that not only the way in which group identities are made real is important, but also that self-favoring considerations mattered. Furthermore, the name calling and unpleasant behavior was considered mean and wrong by some children (“what he did was mean”) which indicates that moral considerations were also relevant.

One of the central contributions of the social identity approach is that it tries to establish a link between group identities and social reality. Social categories and group understandings are taken to reflect the intersubjective and normative nature of the world people find themselves in. Group stereotypes, for example, are not considered misperceptions but rather conceptualized and empirically found to be sensitive to the intergroup realities in the social world (Oakes et al., 1994). People are organized and understood in terms of meaningful social categories because this is how they are organized and considered in the real world. The children in Elliott’s classroom found themselves in a context where pupils are organized by “race” and where thinking in terms of the group differences between “brown-eyes” and “blue-eyes” fits the social context.

The social identity approach conceptualizes ingroup bias as the interactive result of human motivations and social realities. Children have a strong desire to acquire knowledge and are curious about the meaningful aspects that constitute the world they try to make sense of (Piaget, 1929). An adequate understanding of social reality, including social groups (Hirschfeld, 1996), is a crucial aspect of the process of growing up, and defining the nature of groups as being factual has been found to play a role in early adolescents’ discussions about outgroup friendships (Verkuyten & Steenhuis, 2005). This also means, for example, that knowledgeable authorities such as parents and teachers exert an important influence on the child’s understanding and beliefs about social groups.

In Elliott’s classroom, she was the authority but her definition of the two group identities was not simply accepted by the children. At first, some children were hesitant and resistant towards her statement that brown-eyed people are better than blue-eyed people. The children questioned the notion that the ontological nature of these two groups of children differs. They readily understood that one can make such a distinction and let it have specific social consequences for functional reasons, but not that you can use it to tell something about the types of children involved (“we are like this, and they are like that”). In other words, for some children the ontological component was missing and therefore they did not consider eye color a real social identity. Elliot, however, tried to present her definition as reality-based and accurate, for example, by referring to the eye color of George Washington, by presenting evidence from everyday life, and by stating that melanin was linked to the higher intelligence and learning ability of brown-eyed people (“Brown-eyed people have more of that chemical in their eyes, so brown-eyed people are better than those with blue eyes”). In doing so, she explicitly stressed several times that it is a fact that brown-eyed people are better than blue-eyed people. This provided the social identity the required “reality depth” so that it became more than a game or a distinction made for educational purposes.

The social identity approach argues that people strive to have a positive sense of their group self and therefore are motivated to evaluate their own group positively: establishing favorable distinctiveness of one’s own group vis-à-vis other groups (ingroup bias) contributes to a positive self-identity. Research among children has shown that such a bias does indeed positively and causally affect self-feelings (e.g., Verkuyten, 2007). In Elliott’s classroom, the “superior” children felt superior and positive about themselves (“I felt that I was better than them, happy”), which was shown in their emotional and facial expressions and the fact that their grades on simple tests were better, and they completed mathematical and reading tasks that had seemed outside their ability before (Ambady et al., 2001).

However, social reality can constrain children’s preference for a positive group identity. This is illustrated in the fact that children are well aware of social status differences and that low-status group children are less likely to show ingroup bias on status-related dimensions (Baron, 2015). Furthermore, perceived competence rests on realities of power, status and resources, making ingroup bias on this dimension less malleable than on the dimension of warmth/friendliness (Yzerbyt, 2018). Research has found ingroup bias among infants and young children (e.g., Dunham et al., 2011; Hetherington et al., 2014) but there are also reality constraints that children take into account, such as group differences in wealth, opportunities and performances. In one study, preschool and elementary school-age children received bogus scores for an egg-and-spoon race that ostensibly placed them in a fast or a slow team (Yee & Brown, 1992). The children were also told that members of the fast and slow groups were selected for membership in a particular group on the basis of their performance in the race. Both groups were found to show ingroup bias. However, there was a difference in performance ratings. The ingroup team was rated as faster than the other team by children placed on the fast team, whereas children on the slow team rated their ingroup as slower than the other team. Hence, these performance ratings reflected reality rather than positive group identity concerns that, probably, guided the affective ratings. However, in making self-evaluations and intergroup evaluations 5- and 6-year-olds disregarded factual information more easily than 9- and 10-year-olds (Ruble et al., 1976; Yee & Brown, 1992).

Reality and children’s desire for adequate knowledge can constrain their tendency to make favorable ingroup comparisons on specific dimensions of comparison. Another example is an experimental study among three age groups (6-, 8- and 10-year-olds) that examined trait attribution effects of reality constraints on eye color differences and national group differences (Verkuyten & De Wolf, 2007). For all three age groups, it was found that children take information about social reality into account when giving intergroup ratings, and when children were not fully convinced about the reality claims that were made in the experiment context thus leaving more room for interpretation, the youngest children were found to show the clearest pattern of ingroup bias. These findings indicate that the striving for a positive group identity is constrained by social reality, as is proposed by the social identity approach. The tendency to feel good about oneself by making a positive distinction in favor of one’s own group is bounded by the desire to develop an adequate understanding of the social world one finds oneself in. Thus, in addition to the striving for a positive group self, the social identity approach emphasizes the importance of group level social reality and people’s epistemic motivation for understanding ingroup bias. This latter motivation is also considered important for processes of normative influence.

Although the children in Elliott’s classroom appeared to display ingroup bias in the absence of the teacher (e.g., at the playground), the bias might also be about meeting teacher’s expectations and Elliott’s related praise for work done in class. Children try to make sense of group differences and follow normative expectations and social rules from authorities and those that they develop in interaction with each other. Already, preschoolers recognize, create, follow and enforce social norms (Schmidt & Rakoczy, 2018), and peer groups develop shared understandings of the social world as well as group norms.

An increasing number of studies argue and demonstrate that social norms matter for children’s intergroup relations. For example, the social identity development model (Nesdale, 2017) and the developmental subjective ingroup dynamics model (Abrams et al., 2008) state that children’s ingroup biases depend on the norms of their ingroup. Empirically, children’s attitudes toward ethnic, racial and other groups have been found to be affected by experimentally induced (e.g., Nesdale & Dalton, 2011) as well as perceived peer group norms (e.g., Brenick & Romano, 2016). For instance, children use their understanding of group loyalty norms as a basis for evaluating peers (Abrams et al., 2008; Rutland et al., 2015).

Normative influence is often conceptualized as complying with the group, based on children’s motivation to gain or maintain social approval and avoid disapproval. Normative influence is considered externally motivated in its focus on social acceptance and belonging which makes children respond publicly without ingroup bias and outgroup negativity (Rutland, 2004). However, there is also informational influence which is considered “true” influence because it leads to the internalized acceptance of messages. This relates, for example, to the research on the formative role that parents and teachers play in the development of children’s outgroup attitudes (e.g., Degner & Dalege, 2013; Geerlings et al., 2019).

In the social identity approach, group influence is less about normative compliance but rather reflects an internal process, akin to informational influence (Turner, 1991). Group members tend to trust their ingroup, which as a refence group is considered to provide valid and relevant information about the world. Children who self-categorize as a group member rely on ingroup members for their epistemic goals and social reality testing, and not only for gaining social approval. Infants and children prefer to learn from ingroup, compared to outgroup, members who provide relevant cultural information (Begus et al., 2016) and are considered more trustworthy (Chen et al., 2013). Furthermore, with the act of defining oneself as a group member, the shared group understandings become self-relevant. You start to understand yourself and others in terms of the norms, beliefs and values associated with that particular group identity. A process of self-stereotyping occurs whereby the child comes to think about him/herself in terms of the attributes, characteristics and behavioral expectations that are commonly associated with what it means to be, for example, a “blue-eyed” or “brown-eyed” person.

Processes of self-stereotyping imply that the self extends beyond the individual person to an inclusive social unit (“I am like my group”). With age, children develop an increasing social understanding of group differences and what characterizes various groups, including their ingroup. This is important because the process of self-stereotyping requires such an understanding that makes the assimilation of the self to typical ingroup attributes and characteristics possible. This understanding depends, in turn, on children’s cognitive capabilities, their social experiences and the information provided by the social surrounding. In a Piagetian perspective, young children tend to assume that other people see, hear and feel the same as the child does. From middle childhood on, children become less self-centered and increasingly interested in group differences, develop perspective-taking abilities and have more experiences with groups in various situations. As a result, they become more sensitive to group norms and develop abstract understandings of intergroup differences (Karcher & Fischer, 2004). Research on gender categories, for example, has shown that already in middle childhood, children are able to gender self-stereotype (Sani & Bennett, 2004). Further, experimental research among preadolescents (10–12 years) has found evidence for self-stereotyping differences depending on the particular cultural identity that is salient (Verkuyten & Pouliasi, 2002).

In Elliott’s classroom, the nature of the two groups was made explicit, which made it relatively easy for the children to think about themselves in the related group terms. However, such a process of self-stereotyping is more difficult when children are faced with groups that lack clarity in their identity content (Van Veelen et al., 2016). For example, moving to an unfamiliar setting or to a context with high cultural diversity and rapid cultural changes often implies that groups are diverse and also ambiguously defined. In such contexts there is no clear group information readily and consensually available to assimilate the self to, which means that self-stereotyping is more difficult.

However, because children expect others of their ingroup to be similar to themselves, a process of social projection is likely in these situations. With social projection, personal attributes are projected onto the ingroup (“the group is like me”), and it is assumed that others have similar views and beliefs as the self (Robbins & Krueger, 2005). From a developmental perspective, this process is more likely among younger (4–6 years old) children who tend to differentiate less between themselves and others and have less advanced perspective-taking abilities (Abrams, 2011). However, social projection is not simply a matter of children’s focus on themselves or limited cognitive abilities but occurs at all ages and especially when the information about one’s ingroup and its members is rather unclear or hard to believe (Van Veelen et al., 2016). For example, older children (10 years onwards) do not want to come across as prejudiced. They tend to avoid acknowledging race even when it is a relevant category (Apfelbaum et al., 2008) and inhibit the expression of ingroup bias (Olson & Dunham, 2010). This makes it rather difficult to know what peers really think about particular outgroups. As a result, social projection is more likely whereby children assume that ingroup peers have similar outgroup attitudes as themselves.

In a study on attitudes toward various racial groups (White, Black, Chinese), it was found that children’s (8–11 years old) own attitudes were generally unrelated to the actual attitudes of their friends, but strongly associated with the expected friends’ attitudes (Aboud & Doyle, 1996). Thus, children seemed to assume that their friends share their racial attitudes without this being the case. In another survey study, it was examined whether children (9–12 years old) have the tendency to project their own ethnic ingroup favoring attitude on their classroom peers (Thijs & Verkuyten, 2016). It was found that children’s perception of the attitudes of their classmates was partly accurate but also the result of children’s own unique attitudes (not shared with their classmates). This tendency of social projection was found among ethnic minority and majority students alike and across age groups. However, the tendency was stronger for children who had a stronger sense of classroom belonging.

In a further study, a longitudinal design was used to demonstrate that children’s own unique ethnic attitudes predict their perception of the attitudes of their classmates half a year later (Thijs & Zee, 2019). This social projection effect was found to be equally strong for 7- to 9-year-olds and for 10- to 12-year-olds, and among ethnic majority and minority children. Additionally, there was also, although a smaller, social influence effect in that children’s perception of the classroom attitudes predicted their own attitude over time. Interestingly, this study found no evidence for social projection in relation to attitudes toward the ethnic majority group. Rather, children were quite accurate in their perception of what their classmates think about majority group children. This suggests that classmates are less likely to refrain from expressing their attitudes toward the secure high-status group and, therefore, that there is less uncertainty about the classroom belief about this group and thus a lower need for social projection (Van Veelen et al., 2016).

The findings of these studies indicate that children are concerned with what their ingroup members think, especially when they have a relatively strong sense of group belonging. However, this does not imply a one-way process of social influence but might involve a process of social projection to ingroups in situations in which children do not fully know, misperceive or misunderstand the views of their ingroup members. This is more likely in some situations than others. For example, and similar to Elliott’s exercise, in deeply divided societies, such as Israel and Northern Ireland, or societies with strong ethnic group boundaries, such as Malaysia and Mauritius, children learn from very early on about the relevant group distinctions, making self-stereotyping more likely and processes of social projection less likely (Connolly et al., 2009; Nasie et al., 2016; Ng Tseung Wong & Verkuyten, 2015).

In contrast, social projection seems more likely in societies and settings in which there are relatively open and overlapping group boundaries and various intersecting identities exist. However, in these situations peer groups can also develop group norms. Those who identify together do not automatically know each other’s views or immediately agree with each other. However, group identity is an important basis for mutual influence and developing ingroup consensus (Turner, 1991). The social identity approach argues that there is a process of consensualization in which ingroup members expect and strive to agree. For example, when people think about themselves as belonging to the same ethnic or racial category, there is an expectation of agreement and a motivation to reach consensus on the meanings and implications of the ethnic or racial identity (Haslam et al., 1999). When learning new information, young children look to ingroup, rather than outgroup, members (Begus et al., 2016), and they seek and endorse consensual information from members of their own group, rather than a different group (Chen et al., 2013). A common identity means that one is similar to others in one way or another and belong together, and this leads group members to seek agreement and try to create normative consensus. This means that it is important to examine the ways in which children collectively and interactively develop group understandings and create and enforce social norms (e.g., Connolly, 2001; Van Ausdale & Feagan, 2002). According to the social identity approach, ingroup bias is a shared product of social processes of influence and communication.

Children want to feel good about their group membership, and they can do so by making a positive distinction in favor of their own group. However, this is not unrestricted because there is also the need to develop an adequate understanding of social reality and there are relevant group norms. Furthermore, there are moral considerations that can constrain the tendency to favor one’s ingroup (Killen et al., 2016).

In the documentary about Elliott’s exercise, there is talk about being mean and about fairness. For example, the name calling and teasing that occurred are called by some children as being mean and wrong. At the end of the documentary, Elliott talks about the unfairness of discrimination: “We are treating people a certain way because they are different from the rest of us. Is that fair?,” to which the children reply “no,” after which she continues with “nothing fair about it. We did not say this was going to be a fair day, did we (‘No’), and it isn’t. It is a horrid day.”

According to sociocognitive domain theory (Turiel, 2002), children’s social reasoning can reflect psychological, conventional and moral considerations. From early childhood on, children apply these forms of reasoning to understand social behavior. Moral considerations relate predominantly to issues of well-being and fairness and are considered objective, obligatory, inalterable and relatively independent from authority and peer influences (Skitka et al., 2021). Moral criteria are general and apply to ingroup and outgroup behavior, and moral information curbs children’s ingroup bias because they feel internally obligated to act on it. Moral principles like fairness and equality are often privileged in children’s reasoning and can trump what group stereotypes and group norms prescribe. Children (4 years onwards), for example, tend to give priority to fairness over stereotypic gender expectations in gender relations (Killen et al., 2001), use moral reasoning to evaluate gender and ethnic exclusion (Møller & Tenenbaum, 2011), demonstrate less ingroup bias in morality-based judgments compared to convention-based judgments (e.g., Abrams et al., 2008), and sometimes hardly show any ingroup bias in rejections of unfair allocations (Gonzalez et al., 2020, but see Fehr et al., 2008) and in explaining a perpetrator’s victimization behavior (Verkuyten, 2003). Further, in settings marred by religious conflict (India), older children judge it wrong to harm religious others (Srinivasan et al., 2019), and children prioritize cooperative helping norms when these conflict with group identity concerns (Gonzalez-Gadea et al., 2020).

Furthermore, research demonstrates that children as young as 7 years understand the difference between having external and internal reasons for outgroup attitudes (Hughes et al., 2016; Jargon & Thijs, 2020). External reasons entail the desire to conform to social norms and avoid social disapproval, while internal reasons stem from moral beliefs about equality, harm and fairness. Internal reasons tend to be more strongly and consistently associated with outgroup acceptance than external reasons (Hughes et al., 2016). For example, research among older children (9–12 years) in elementary schools has found that egalitarian and equality messages can stimulate self-endorsed reasons for accepting outgroups, whereas explicit prosocial norms lead to concerns about social sanctions and compliance (Jargon & Thijs, 2020). Furthermore, trait-like empathetic (Miklikowska, 2012) as well as situationally induced empathic understanding (Sierksma et al., 2015) can overcome children’s ingroup biases. Empathic concern implies that one cares for the well-being of others and experiences parallel emotions, which is related to more positive outgroup attitudes (Nesdale et al., 2005; Van Bommel et al., 2021). In one study, children were found to empathize with asylum seekers who had to flee their country and leave their friends behind, arguing that “a refugee can really use my friendship,” and “it is very sad and so, and then you want to be friends with them” (Verkuyten & Steenhuis, 2005). Children can empathize with others and are intrinsically motivated to help others from a very young age onwards (e.g., Fehr et al., 2008; Schmidt & Sommerville, 2011; Warneken & Tomasello, 2006).

Some research using the minimal group paradigm has observed that children’s choices tend to represent a balancing or coordination between the goals of ingroup bias and considerations of fairness (Branthwaite et al., 1979; Tajfel, 1970; see also Turiel & Gingo, 2017), that strategies of fairness and equality are more common in children’s punishment behavior (McAuliffe & Dunham, 2017) and in negative outcomes and allocating sanctions between ingroups and outgroups (Rutland et al., 2007), and that 4- and 5-year-olds’ ingroup preference is significantly attenuated in the presence of an antisocial ingroup member (Hetherington et al., 2014). These findings indicate that developing a positive group self is not the only thing that matters for how children relate to other groups, not even in minimal group settings.

The importance of moral concerns is compatible with the social identity approach because the content of group identity, for example, can emphasize moral values (“we are fair-minded, charitable, honest”; Ellemers, 2017) and because the findings mentioned might be interpreted in terms of the level of self-categorization. Being asked to administer negative outcomes is unusual and rather socially inappropriate and therefore might lead in a minimal group setting to a recategorization into “us” (the two groups) versus “them” (i.e., the experimenters) which reduces the categorical difference between the two groups and the related ingroup bias (Gardham & Brown, 2001).

The social identity approach argues that people can think about themselves at different levels of abstraction, such as an ethnic group member, a national and a human being. All these levels have their own specific identity contents, and the moral value of egalitarianism and fairness is prevalent within superordinate identities like shared nationality and common humanity. Whereas children can favor their ethnic or racial ingroup, there is, at the superordinate level of society and humanity, the moral principle to treat all individuals and groups equally, fairly and without harm. These principles go beyond the specific intergroup situation by foregrounding a superordinate identity with the related moral values that proscribe ingroup bias and encourage intergroup fairness (Iacoviello & Spears, 2018). For example, children have been found to explain that it is unfair and discriminatory to not want to be friends with an outgroup peer since “we are all just humans” and “they are normal children or people, just like us” (Verkuyten & Steenhuis, 2005).

According to the social identity approach, group identity underlies ingroup bias. It is because children are able to conceive of themselves and others as group members that they are able to differentiate their ingroup from outgroups. Children want to understand their social world and define themselves in terms of the social realities that are meaningful in particular settings, cultures or historical periods. Their own group provides them with relevant information about how to understand the social world and their own place in it, and gives them a sense of belonging. Children have a strong epistemic motivation, want to belong, and are concerned with maintaining or developing a positive sense of self. Furthermore, striving for a positive group self and the resulting tendency to favor one’s ingroup does not tell us much about how this is achieved. There are many ways in which one’s group can stand out positively, such as by being competitive or, rather, by being cooperative and supportive. The direction for how to act is provided by the way in which the particular group distinction is made meaningful. When, in a particular situation, children see themselves as members of a group, they will think, feel and act in terms of the norms, values and beliefs that are considered to characterize the group, and these can also be prosocial and moral.

There are various possible reasons for ingroup bias and the particular forms that it takes. This makes it important to try to examine the development of the different processes and how these can simultaneously affect children’s attitudes and behavior. Obviously, the striving for a positive group self is important but ingroup bias is likely to be multiply determined and also dependent on epistemic motivations, social influences and moral considerations. The “superior” children in Elliott’s exercise did not only feel positive and happy about themselves, but they also started to think and act in terms of the logic of the situation as defined and raised moral concerns about the negative intergroup behavior that occurred. The social identity approach does not only focus on ways to establish a positive identity but provides a framework for thinking about the role of other motivations and considerations that also develop early in life. Much of children’s social life involves balancing multiple considerations, and this includes questions about the nature of group distinctions and the implications that these have for oneself and others. Further, developmental research is required in order to fully understand how this balancing process develops and plays out in children’s ingroup biases in different sociocultural contexts.

Although much of the research on ingroup biases is conducted among majority group children in Western societies, the social identity approach emphasizes that the way in which psychological processes play out is dependent upon the local and broader societal context. It is important to know that children strive for a positive group identity, but that tells us little about how this is achieved, and which group distinctions are considered relevant, what they mean and how they feature in children’s lives. Intergroup biases can take different forms in different local contexts (e.g., schools, neighborhoods), and children’s understanding and evaluation of group differences are likely to be different in, for example, strongly racialized and unequal societies (United States, South Africa) than in more multicultural and egalitarian nations (e.g., Singapore, New Zealand). The social identity approach has much more to offer than the well-known prediction that children tend to show ingroup bias because of their need for a positive group self. The approach offers a theoretical framework that is broader and richer in focusing on cognitive and motivational processes and how these operate and work out within the local and broader social world in which children grow up.

No ethical approval was required for the preparation of this paper, as no human or animal subjects were used.

The author has no conflicts of interest to declare.

While working on this paper, the author was supported by a European Research Council Advanced Grant under the European Union’s Horizon 2020 research and innovation program (grant No. 740788).

Email alerts

Citing articles via, suggested reading.

• Online ISSN 1423-0054
• Print ISSN 0018-716X

INFORMATION

• Contact & Support
• Information & Downloads
• Rights & Permissions
• Terms & Conditions
• Catalogue & Pricing
• Policies & Information
• People & Organization
• Stay Up-to-Date
• Regional Offices
• Community Voice

SERVICES FOR

• Researchers
• Healthcare Professionals
• Patients & Supporters
• Health Sciences Industry
• Medical Societies
• Agents & Booksellers

Karger International

• S. Karger AG
• P.O Box, CH-4009 Basel (Switzerland)
• Allschwilerstrasse 10, CH-4055 Basel
• Tel: +41 61 306 11 11
• Fax: +41 61 306 12 34
• Contact: Front Office
• Experience Blog
• Privacy Policy
• Terms of Use

This Feature Is Available To Subscribers Only

Sign In or Create an Account