Identity, Homophily and In-Group Bias *

Identity, Homophily and In-Group Bias∗

Sergio Currarini† Friederike Mengel‡,§

August 6, 2013

Abstract

Homophily refers to the widely documented fact that people tend to interact more oftenwith “similar” others. In this paper we propose a minimal design to study homophily inthe laboratory. Our design allows us to address questions about the sources of homophily(whether homophily is a matter of opportunities, stereotypes, beliefs more generally or pref-erences) which are hard to address in field studies. Our minimal design also allows us torelate homophily to another form of discriminatory behaviour known as in-group bias. Wefind evidence of homophily even in the absence of stereotypes, biased beliefs or differing op-portunities. We also find evidence for in-group biases, and that homophily and in-group biasare correlated phenomena. However, in-group biases either decrease or disappear altogetherwhen matching is made endogenous, i.e. whenever homophily is given a playing field, andthis is partly due to a shift in behaviour. Interestingly, this implies that letting agents incontrol of their matches may lead to higher segregation and lower discrimination. We alsofind that homophily is strongly correlated with risk aversion, and we use this fact to providean explanation for both the existence of homophily and the disappearance of in-group biasesunder endogenous matching.

Keywords: In-Group Bias, Homophily, Endogenous Matching, Experiments, Game Theory,Segregation, Discrimination.

JEL-classification: D03, D01, C91, C92, C7.

∗We thank Roland Benabou, Matt Jackson, Rachel Kranton, Yiqing Xing as well as seminar participants atExeter, Faro (SAET 2011), Koeln, Leicester, Maastricht, Malaga (EEA 2012), Paris (CTN 2012), Bristol andSciences Po for helpful comments. Friederike Mengel thanks the Dutch Science Foundation (NWO Veni grant451-11-020) for financial support.†University of Leicester, University of Venice and CMCC. The research leading to these results has received

funding from the Italian Ministry of Education, University and Research and the Italian Ministry of Environment,Land and Sea under the GEMINA project. Email: [email protected]‡Department of Economics, University of Essex, Wivenhoe Park, Colchester CO4 3SQ, United Kingdom,

e-mail: [email protected]§Department of Economics (AE1), Maastricht University, PO Box 616, 6200 MD Maastricht, The Netherlands

1

1 Introduction

Homophily refers to the empirical fact that people tend to interact with similar others, wheresimilarity usually refers to ethnicity, gender, religious views etc. The notion of homophily isamong the oldest in social sciences. It has been documented across many different dimensionsof similarity (e.g. ethnicity, gender, religious views etc...) and typologies of social ties (fromthe intimate relations of friendship and marriage to business collaborations and everyday inter-actions).1 By translating distance in individual characteristics into social distance, homophilyaffects socio-economic outcomes in various ways, limiting the diffusion of information (Jack-son and Lopez-Pintado (2011), Golub and Jackson (2011)), affecting the functioning of labourmarkets (Calvo-Armengol and Jackson (2004)) and marriage markets (Skopek et al. (2011)), de-termining who befriends who in social networks (Currarini et al. (2009) Currarini et al. (2010)),or who adopts health innovation and care (Centola (2011)). Homophily has also consequences onthe spread of social norms and the level of cooperation in a society (Hamilton (1964), Myersonet al. (1991) or Mengel (2008)) and, by fragmenting society, on the provision of public goods(Alesina et al. (1999)), on the level of infrastructure quality and literacy (Alesina et al. (2003))and on economic growth (Easterley and Levine (1997)).

Yet, many questions about the origin and the cognitive or emotional roots of homophilyremain unanswered. Recent research has stressed the distinction between two primary sourcesof homophily: preferences and opportunities (Moody (2001), Currarini et al. (2009), Currariniet al. (2010), Currarini and Vega-Redondo (2011)). While preferences refer to mental attitudesabout others, opportunities refer to biases that agents face in the process of forming social ties,such as, for instance, wealth constraints in neighborhood choice, academic tracking in schoolsor other organizational constraints. In empirical field studies it is difficult to disentangle thosetwo sources of homophily, since it is hard to control for matching opportunities. Assuming thatopportunities can be controlled for, once can further ask whether homophily results from beliefs(e.g. gender, racial or religious stereotypes) or is directly encoded in preferences.

In this paper we study homophily in the framework of a controlled laboratory experiment,which allows us to control for opportunities and to distinguish between preferences and beliefs.

In our experiment participants are randomly assigned to one of two groups, called the REDgroup and the BLUE group, with no further identity enhancing activity. We adopt, hence,what is called the ”near-minimal” group design paradigm (see Tajfel and Turner (1979) andthe literature cited below). Choosing such arbitrary assignment to minimal groups reduces thepossibility that stereotypes participants may have about others are triggered in the experiment.Everything participants know about the participant(s) they interact with in the experiment iswhether s/he is from the RED or BLUE group.

The minimal group design also allows us to introduce elements of discrimination in economictransactions along lines followed in previous studies of in-group bias (see for instance Chen andLi (2009)). Having such discriminatory elements in our experiment is of interest for variousreasons. First, in-group bias can, in principle, be a major driving force for homophily, sinceagents may strategically favor in-group matchings to avoid being discriminated against. Second,although both in-group bias and homophily involve some extent of discrimination, they describetwo clearly different things: homophily refers to a bias in who agents are matched with, whilein-group bias refers to discriminatory behaviour of agents given who they are matched with.All previous work on in-group bias has assumed exogenous random matching of agents, andhas therefore overlooked the interplay of in-group bias and homophily when matching is tosome extent endogenous. The present setting allows us to study this interplay, and to betterunderstand the complex empirical relationship between homophily and in-group bias.

Most of our results come from the analysis of two treatments, that differ in the adopted

1The observation that ”similarity begets friendship” dates back to Aristotle and Plato; see Mc Pherson et al.(2001) for a comprehensive survey of the sociological literature on homophily

2

matching institution: EXO, in which participants are matched according to a color-blind uniformrandom process, and ENDO, in which participants are allowed to affect their probability ofmatching with RED and BLUE types (see below for details). After matching has occurred, inboth EXO and ENDO participants play a series of 8 games designed to elicit their degree ofaltruism, positive reciprocity and negative reciprocity.2.

In treatment ENDO homophily is measured by the willingness of participants to be matchedwith opponents from the same RED/BLUE group. We elicit two measures of homophily. Firstwe ask for a pure expression of preference for a RED or BLUE match, which is non-incentivized.Second, we elicit participants’ willingness to pay (wtp) for an in-group (or out-group) match;Participants are then matched according to their wtp using an incentive-compatible mechanism,such that agents with a higher wtp for an in-group (out-group) match are more likely to bematched with in-group (out-group) opponents. Afterwards we let participants play the 8 games.We find evidence of pervasive homophily even in our minimal setting. About 45 percent ofparticipants in treatment ENDO indicate a strictly positive wtp for an in-group match and 70percent indicate a (weak) preference for an in-group match.

Homophily is clearly not a matter of opportunities in our experiment. Can we then explainhomophily as a strategic attempt to gain from preferential treatment in in-group meetings?Quite surprisingly, we find no evidence of in-group bias in the ENDO treatment. More precisely,average behavior in in-group matches does not statistically differ from average behavior in out-group matches. To the extent that participants’ expectations are correct, this suggests thathomophily does not stem from strategic anticipation of in-group bias. This intuition is confirmedby a comparison of levels of homophily in ENDO with two control treatments, one with no scopefor in-group bias and a second with substantial scope for in-group bias. We find no significantdifference with the first, and significantly lower levels of homophily in ENDO compared to thesecond.

Having found that homophily is not ”strategic” in the sense made clear above, we tried togain some insights on what is likely to be behind participants’ willingness to pay for in-groupmatches. One possible clue might lie in the evidence we gathered in our post-play questionnaireon risk attitudes, cognitive abilities, gender, age and nationality. In particular we found thathomophily is strongly correlated with a measure of risk aversion, but not with any of the othermeasures elicited. This finding echoes recent sociological theories, which interpret homophily asa way to reduce subjective uncertainty (Hogg (2000)). Shifting the focus on risk attitudes as theprimitive source of homophily requires some rationale for agents to perceive in-group interactionas less uncertain. One possible interpretation is that homophilous agents expect behavior inin-group matches to be more ”disciplined” (and therefore more predictable) in in-group matchescompared to out-group matches. This is indeed consistent with the behavioral patterns weobserve. In particular, negative reciprocity is higher in out-group matches (compared to in-group) in the EXO treatment, but not in the ENDO treatments. In fact, those participants thatare homophilous are even less negatively reciprocal in out-group compared to in-group matcheson average. Harsher punishments may result into weaker incentives to damage one’s opponent,and to the expectation of a more stable conduct which should attract risk averse participantsinto in-group matches in treatment ENDO.

The above patterns of behavior across treatments also unveil interesting insights on the effectof the matching institution for economic discrimination, and on the interplay of homophily andin-group bias. We find that homophily is indeed correlated with in-group bias and that onlyhomophilous agents in ENDO display in-group bias at all. At the aggregate level, however,we find that the in-group biases observed in EXO disappear when matching is endogenous(ENDO).3 Interestingly, we show that the self-selection of homophilous agents into in-group

2We use variants of some of the games described in Charness and Rabin (2002)3In EXO, participants are about 35 percent more likely to reward an in-group match for good behavior, but 40

percent less likely to punish an in-group match for misbehavior compared to out-group matches. These findings

3

matches cannot, alone, explain why in-group biases decrease. Some shift in behavior must havetherefore occurred in response to the change in the matching institution. As discussed above,we find that participants tend to be more negatively reciprocal (and less positively reciprocal)to in-group matches in ENDO compared to EXO.

This leaves us with an interesting and somewhat provoking message. Even though morehomophily would probably lead to more discrimination (in-group bias) ”for any given matchinginstitution”, letting agents decide about their match (moving from exogenous to endogenousmatching) decreases aggregate discrimination. In particular, letting agents be in control of theirown economic relations has two opposite effects on social structure and economic outcomes:the degree of segregation (measured by the share of in-group interactions) will increase dueto homophily but, at the same time, social discrimination (measured by in-group bias) maydecrease due to a combination of self selection and changed individual behavior. Hence self-selected segregation need not necessarily be detrimental to the level of pro-social behaviour in asociety and need not increase discrimination. In this sense, the experimental evidence presentedin this paper provides new elements for the assessment of policies that affect social and economicsegregation.

To sum up we have found evidence for homophily even in a minimal design which controlsfor opportunities and even in the absence of stereotypes, otherwise biased beliefs or strategicincentives. Although we have found that homophily is correlated with in-group bias at theindividual level, their relation is not of a straightforward strategic nature, but rather of a morefundamental (preference driven) nature. In particular it seems that risk aversion may be a keydeterminant of homophily and the crucial link between the two phenomena. This correlationbetween risk aversion and homophily points towards a theory of homophily that merits, in ourview, further exploration.

Previous literature has documented homophily in several empirical field studies. See e.g.Currarini et al. (2009), Currarini et al. (2010), Centola (2011) or Mc Pherson et al. (2001) forsurvey of the sociological literature. As already mentioned above this literature is mostly focusedon documenting homophily as a biased matching pattern, but has found it difficult to identifythe sources of homophily. To our knowledge there are no experimental studies on homophilyyet.

There has been much more experimental work on in-group bias. A number of papers inpsychological research have found evidence in support of in-group favoritism under a ”minimalparadigm” design, where the assignment of agents to groups is made with no reference to previousinteraction, correlation of preferences or pre-formed identities (see the pioneering work by Tajfeland Turner (1979)). This has been confirmed in a recent study in experimental economics byChen and Li (2009), who have found in-group biases when agents are sorted in ad-hoc mannerinto two groups labeled with colours (maize and blue). Our treatment EXO is closely relatedto the study by Chen and Li (2009), who use a within-subject design and ask each participant(using the strategy method) to choose an action conditional on being matched within the groupor outside their group. In contrast to them we use a between-subject design and do not use thestrategy method. The advantage of Chen and Li (2009)’s design is that they can observe identitybiases within subjects. However a disadvantage of their design is that it could potentially beleading, since participants may feel that their answers should differ simply because they areasked to make a conditional decision. Our results from treatment EXO show, though, that theirfindings are robust to these changes in design. If anything, the in-group biases we find are slightlystronger. Other works in experimental economics have not found biases in a truly minimal design(see Charness et al. (2007) in the context of prisoner’s dilemma and battle of sexes, Eckel andGrossman (2005) in the context of team production and Chen and Chen (2012) in effort games.).

are in line with Chen and Li (2009), which we discuss below. We also find that whether a match is in-group ornot is more important than age, gender, the colour of one’s group, a measure of risk aversion and a measure ofcognitive reflection in determining behaviour.

4

Strong in-group biases have been found by Chen and Chen (2012), Charness et al. (2007) orChakravarty and Fonseca (2012) in designs that were not truly minimal. Ioannou and Rustichini(2012) find in-group biases only if group payoffs are displayed. There is also strong evidenceof the role of priming in reinforcing identity-based behavior and preferences (Charness et al.(2011), Benjamin et al. (2010) and Van Bavel et al. (2008)).

There have also been attempts to study in-group bias when the pattern of interaction isto some extent endogenous. Evidence of a preference for in-group matching is obtained, in apure allocation problem, by Foddy et al. (2009) who find that agents prefer to receive donationsfrom in-group members rather than from out-group members. They also show that this can beimputed to the expectation of better treatment inside the group. Our experimental evidenceonly partially confirms and points to a perceived reduction in uncertainty as a main motivefor homophily (see below). A number of papers have considered social dilemma games withan endogenous group structure (Coricelli et al. (2004), Keser and van Winden (2000), Grimmand Mengel (2009) and Ahn et al. (2009) among others). Since selection and exclusion - thatare indeed found to affect behaviour - are based on behavior rather than group membership oridentity in these studies, they are somewhat less related to our work.

Our paper is organized as follows. Section 2 describes our experimental design in full detail.Section 3 and 4 present the evidence for in-group bias and for homophily. Section 5 discussesthe possible determinants for these phenomena, and for the decrease in in-group bias underendogenous matching. Section 6 concludes.

2 Design

Our experiment was conducted at the BEE-Lab at Maastricht University between March-May2011. 258 participants participated in one of our 5 main treatments. Our basic treatments aretreatments ENDO, EXO and CONTROL.

Treatment ENDO In treatment ENDO participants were first randomly and exogenouslyallocated to the BLUE and RED group and informed about which group they belong to. Sub-sequently the experiment developed as follows: First participants were asked whether theypreferred to be matched with a member of the RED or BLUE group. Subsequently (on a newscreen) their willingness to pay (wtp) for their choice was elicited. (Details on the elicitationmechanism can be found below). They were then matched with a member of the RED or BLUEgroup (with probabilities depending on their wtp) and informed about the group of their match(RED or BLUE). Subsequently the two players in a given match were randomly allocated therole of Player A and B (with equal probabilities) and played 8 games with their match in fixedsequence.

First Mover (A) chooses Second Mover (B) choosesG1 no choice (400,400) or (750,375) AltruismG2 no choice (100,300) or (400,200) AltruismG3 (250,250) or let B choose (100,100) or (500,100) Negative ReciprocityG4 (50,650) or let B choose (0,100) or (100,100) Negative ReciprocityG5 (500,0) or let B choose (300,300) or (600,275) Positive Reciprocity (Inequ. increasing)G6 (250,0) or let B choose (100,100) or (250,50) Positive Reciprocity (Inequ. increasing)G7 (350,100) or let B choose (300,300) or (100,350) Positive Reciprocity (Inequ. decreasing)G8 (400,0) or let B choose (200,200) or (0,400) Positive Reciprocity (Inequ. decreasing)

Table 1: The 8 games. Payoffs are in format (πA, πB) where πi is the payoff of player i.

Table 1 describes the 8 games. They are variants of some of the games described in Charnessand Rabin (2002). In Games 1 and 2 there is no choice for Player A and Player B chooses betweentwo allocations. These games indicate how altruistically Player B behaves. In all other gamesPlayer A moves first and either ends the game by picking an allocation or lets Player B choose,

5

who then chooses between two allocations. Games 3 and 4 indicate how negatively reciprocalPlayer B behaves and Games 5-8 indicate how positively reciprocal Player B behaves. Wewill sometimes distinguish between the cases where positive reciprocity is inequality decreasing(Games 7 and 8) or increasing (Games 5 and 6). We can also distinguish whether altruism isinequality decreasing (Game 1) or not (Game 2) and whether negative reciprocity is inequalitydecreasing (Game 3) or increasing (Game 4).

Participants in the role of Player B were asked to make a conditional choice indicating whatthey would do if player A decided to let them choose (strategy method). This means we observedthe choice of each participant in the role of Player B irrespective of what Player A did actuallychoose. Hence even if there is little variation in player A’s behaviour or if player A’s behaviouris very different across treatments, we get a full set of observations from player B. Participantsdid not receive feedback about each other’s choices until all eight games had been played. Thisimplies that each participant can be treated as an independent observation in all the games.

Treatment EXO Our second treatment (EXO) coincides with ENDO except for the factthat participants were randomly and exogenously matched with either someone from the REDor BLUE group to play the 8 games. Comparing behaviour in EXO and ENDO enables us tounderstand the connection between homophily (which can manifest itself only under endogenousmatching) and in-group biases at the population level.

Treatment CONTROL In our third treatment (CONTROL) participants were randomlymatched to play the 8 games, but there were no RED or BLUE groups in this treatment.4

The control treatment allows us to see how behaviour is affected by the introduction of groups.Understanding how the creation of these artificial differences among our participants affectsbehaviour has relevant implications regarding the role of heterogeneity within a society.

Label Matching RED/BLUE groups Games Observations

T1 ENDO endogenous yes SocialPreferences 67T2 EXO exogenous yes SocialPreferences 73T3 CONTROL exogenous no SocialPreferences 44

T4 LOWB endogenous yes No Scope for Bias 38T5 COORD endogenous yes Coordination 36

Table 2: Main Treatments with Number of Independent Observations (Participants).

Those are our main treatments that we will use to understand homophily and in-group bias.Let us briefly define these two key notions, that we will discuss in more detail later.

Homophily As we mentioned before, by homophily we will refer to a preference for interactingwith agents from the same group. Most of the time, we will measure homophily by the willingnessto pay for an in-group match (and heterophily as the wtp for an out-group match). It is importantto note, though, that participants faced two subsequent screens. On the first, they express apure preference for RED or BLUE and only on the second screen they are asked for their wtp.This means that we also have a measure of homophily that is independent of the wtp and wewill use this measure as well for robustness checks. Given our design, more homophily (in thesense defined above) will of course also lead to a higher expected number of in-group matchesin the ENDO treatment.

In-group Bias By in-group bias we will refer to differences in behavior across in- and out-group matches, which we will measure by the behavior of player B in our eight games.

4Sample Instructions for treatments ENDO, EXO and CONTROL can be found in the Appendix.

6

Additional Treatments We conducted two additional treatments to understand the reasonsbehind homophily. Both LOWB and COORD coincide with ENDO, but the 8 games weredifferent in each case. In LOWB games were such that there is no scope whatsoever for in-groupbiases. In other words in LOWB there are no strategic reasons to be homophilous and hencewe consider the amount of homophily observed in this treatment a lower bound. In COORD,however, games were such that being from the same group could potentially help to resolvecoordination problems. Hence we expected there to be more homophily in COORD comparedto ENDO compared to LOWB. Understanding where ENDO lies in the range between LOWBand COORD can help us understand to which extent homophily is strategic. Sample Gamesfrom these treatments can be found in Appendix G. Table 2 summarizes the treatment structuretogether with the number of (independent) observations per treatment.5

Mechanism to elicit wtp We use the following mechanism to elicit the wtp of participantsin treatments ENDO, LOWB and COORD. Participants are endowed with 500 ECU at thebeginning of the experiment. This endowment was not given to them in connection with theelicitation of their wtp. To elicit their wtp they are asked to state a number between 0 and 100that indicates how much they would be willing to pay to be matched with their preferred group.We then draw a random number between 0 and 100 from a uniform distribution. If the randomlydrawn number exceeded the number stated by the participant they were matched randomly.Otherwise they were matched with their preferred group and an amount corresponding to theirnumber was deducted from their endowment. Since we had a finite number of participants inthe experiment, there was a small chance that this mechanism is infeasible. In this case (whichdidn’t happen) we would have matched participants randomly and not deducted anything fromtheir endowment. Participants were fully informed about all these details.

Minimal Group Design The design we used to induce in- and out- groups is called theminimal group paradigm in social psychology (See e.g. Tajfel and Turner (1979)). Accordingto the minimal group design groups are created using trivial and sometimes meaningless tasks.Summarizing 15 years of sociological research Tajfel and Turner (1986) conclude that “thetrivial, ad-hoc intergroup categorization leads to in-group favoritism and discrimination againstthe out-group”. Chen and Li (2009) have compared ad-hoc categorization according to twocolours (maize and blue) with categorization according to expressed preferences for paintings bydifferent artists in an economic experiment. They found that both procedures lead to significantin-group biases, while there are no significant effects of the procedure on either size or directionof in-group biases. In our study we used colours RED and BLUE as group labels to avoidhierarchical labels (such as group 1 and 2 or group A and B) and to avoid that labels arecorrelated with things we can’t control for (such as stereotypes regarding gender, race etc..).

Many studies have shown that expectations about positive and negative reciprocity varygreatly between different cultures and across genders and there are interaction effects betweenthe two as well. See e.g. Gaechter and Herrmann (2009) or Bohnet et al. (2010). Hence to besure that the wtp for in-group matches captures homophily and not e.g. differing expectationsacross dimensions such as gender, culture etc.., it is important to use neutral groups. If we usednon-neutral groups (such as e.g. gender), then we couldn’t be sure that what we call homophilydoes not simply reflect a gender stereotype. Of course in reality gender- and other stereotypesmight well create homophily. In this study, however, we want to focus on deeply rooted culturallylearned preferences for in-group matches, which are activated even for meaningless labels suchas RED and BLUE. In a sense we will measure the component of homophily that is common

554 additional participants participated in two more treatments LABEL and FIXED. The design and someresults of treatment LABEL can be found in the Appendix. The results of treatment FIXED will be used foranother study and are available upon request. Other than the treatments reported we did not run any additionalsessions or treatments, nor did we conduct pilot studies.

7

to all dimensions (race, gender, social status etc.) in which homophily will manifest itself inreal life. In the instructions we also alternated between “RED and BLUE” and “BLUE andRED” to avoid creating a hierarchy between the groups. We also test whether the color affectsbehavior per se and find that generally it doesn’t (see Appendix B).

The abstract minimal group environment is certainly less rich than the natural environmentin the field. However, as we have outlined in Section 1, most existing studies in this literaturesuffer from the opposite problem, namely that their environment is “too rich” to disentangledifferent motives or sources for homophily and in-group biases. We will see below that our verystylized setting will help us to identify clear links between these phenomena.

Other Details We used the experimental software z-tree by Fischbacher (2007) and the re-cruitment system Orzee by Greiner (2004). Participants were paid the sum of payoffs obtainedin all games (with an exchange rate 1 Euro=500 ECU), a show up fee of 2 Euros as well as (intreatments ENDO, LOWB and COORD) whatever remained from their initial endowment of500 ECU. In some of the sessions a participant in the role of Player A was matched to two partic-ipants in the role of player B to determine their earnings. This was done because of low show-up.In this case participants in the role of Player A only received feedback and payments from one oftheir matches which was randomly selected. At the end of the experiment participants filled in aquestionnaire where we elicited their risk aversion and a measure of cognitive ability and askedfor their gender, age and nationality. Each session lasted between 30min (CONTROL) - 70min(ENDO) and participants earned on average 13,40 with a minimum of 5,70 and a maximum of24,60 Euros.

3 Homophily

We start by assessing how prevalent homophily is in our experiment. To these ends, we definethree types of agents, based on their declared willingness to pay and on the type of preferredmatch. Strictly homophilous agents are those with a strictly positive willingness to pay for anin-group match; Strictly heterophilous agents have a strictly positive willingness to pay for anout-group match; and Neutral agents have a willingness to pay of zero.

Figure 1 shows the share of these types in the population. The left-most bar in figure 1(a)measures the share of agents that declared to prefer an in-group match, including those thatafterwards declared a willingness to pay of zero. The three right bars in Figure 1(a) show thepercentages of strictly homophilous, neutral and strictly heterophilous participants in ENDO.

There are about 45% of strictly homophilous agents, about the same proportion of neutralagents, and only few strictly heterophilous agents that make up for about 10% of the population.Figure 1(b) shows the average willingness to pay (wtp) for in-group and out-group matches,respectively. The average wtp is computed as the average of all agents who stated to preferan in-group or out-group match, hence it includes neutral agents. This average willingness topay is significantly larger than zero for in-group, but not for out-group matches, suggestingthat homophily is not only more prevalent but also associated with more intense preferencesover matchings. Among stricly homophilous agents the average wtp for an in-group match is29.03 (t-test, p < 0.0001) and among strictly heterophilous agents it is 23.06. The fact that theaverage wtp for an in-group match among all agents (including neutral and heterophilous agentswhich are assigned values of 0) is only 13.82 (out of a max of 100) and 29.03 among strictlyhomophilous agents, is a reflection of both the substantial share of neutral agents (with wtp= 0)and a first indication that homophily seems to be non-strategic and to work at a relativelysubconscious level in the experiment with minimal groups. The entire distribution of the wtp intreatment ENDO can be found in Figure 7 in the Appendix. Note also that the distribution ofthe wtp for an in-group match is not truncated at zero, since agents can pay up to 100 for anout-group match, i.e. to avoid an in-group match.

8

(a) Matching Preference (b) Willingness to Pay

Figure 1: More than 50 percent of participants prefer an in-group match (p < 0.001 accordingto a binomial test). The average willingness to pay for an in-group match (among all agents) ishigher than that for an out-group match, which does not significantly differ from zero. (t-test,p < 0.0001) (∗∗∗1%,∗∗5%,∗10%).

Having assessed the presence of substantial homophily in our experimental setting, our aimis to use our controlled environment to better understand what the sources of homophily insocial behaviour are. One natural conjecture is that agents prefer to match with similar othersbecause they strategically anticipate potential favourable in-group biases in the interaction tofollow, i.e. higher payoffs in in-group matches compared to out-group matches. This conjectureviews homophily as driven by in-group bias, through agents’ anticipation of potential gains.To test this conjecture, we compare our main treatment ENDO to treatments LOWB andCOORD. In LOWB, strategic interaction is such that players cannot favour or discriminateothers on grounds of group membership. In other words, in LOWB there is no scope for in-group bias. In COORD, instead, there are clear incentives for in-group matching (sample gamesfrom treatments LOWB and COORD can be found in Appendix G). Where ENDO stands withrespect to these treatments helps us understand to what extent homophily in ENDO is drivenby expected game payoffs. Comparing ENDO and LOWB we record almost no difference inaverage willingnesses to pay, which is even slightly higher in LOWB, and exactly the samepercentage of strictly homophilous agents (see Figure 2). In contrast, behaviour in treatmentCOORD significantly differs from that in both ENDO and LOWB, both in terms of the averagewtp and in terms of the percentage of strictly homophile agents (Mann-Whitney, p < 0.0019and p < 0.0581, respectively). This evidence suggests that while agents do respond for whenthey face explicit incentives to be homophilous (as in COORD), the homophily we observe inENDO is not strategic in that it is not driven by the anticipation of more favorable treatmentin in-groups (or in-group bias).

Result 1: Homophily in a minimal design

• There are about 45% of strictly homophilous agents, 45% neutral agents, and 10%strictly heterophilous agents in the ENDO treatment.

• Homophily is not a matter of opportunities and is not driven by the anticipation ofin-group biases in the social preference games.

We will discuss possible alternative explanations for homophily in ENDO in Section 5. Someof our discussion there will be based on the relation between in-group bias and homophily, andon how in-group bias and discrimination are affected by the matching institution. These topicsare covered in detail in the next section.

9

Figure 2: Comparison of wtp across treatments ENDO, LOWB and COORD. In all treatmentsonly the average wtp for in-group matches is significantly different from zero (t-test, p < 0.0001).The average wtp for an in-group match is statistically indistinguishable in ENDO and LOWB,and significantly higher in COORD compared to either (Mann-Whitney, p < 0.0019).

4 In-Group Bias and the Role of the Matching Institution

We start by giving a descriptive overview of behaviour in the 8 social preferences games, thendefine our measure of in-group bias, and present our evidence for the EXO and ENDO treat-ments.

EXO ENDO EXO ENDOA A B B

G1 − − 0.30 0.29− − (0.37, 0.23) (0.31, 0.27)

G2 − − 0.15 0.28− − (0.23, 0.06) (0.26, 0.30)

G3 0.49 0.26 0.37 0.35(0.53, 0.44) (0.24, 0.29) (0.18, 0.61) (0.43, 0.26)

G4 0.61 0.61 0.35 0.18(0.70, 0.50) (0.59, 0.69) (0.18, 0.55) (0.23, 0.15)

G5 0.34 0.34 0.32 0.38(0.36, 0.32) (0.62, 0.08) (0.50, 0.12) (0.44, 0.30)

G6 0.25 0.37 0.23 0.20(0.30, 0.19) (0.62, 0.15) (0.37, 0.08) (0.29, 0.18)

G7 0.16 0.30 0.65 0.65(0.24, 0.07) (0.54, 0.08) (0.69, 0.62) (0.59, 0.70)

G8 0.05 0.15 0.82 0.72(0.08, 0.00) (0.24, 0.08) (0.91, 0.72) (0.85, 0.67)

Table 3: Overview of Behaviour Treatments EXO and ENDO. Percentages in A columnnsindicate the frequency with which player A moves “In”, i.e. lets Player B choose. Percentages inB columns indicate the average frequency with which Player B chooses the altruistic (G1-G2),negatively reciprocal (G3-G4) or positively reciprocal (G5-G8) options, respectively. Numbersin brackets indicate the average frequency in in- and out-group matches.

A first look at Table 3 suggests that there are not many differences in average behaviour ofPlayer B between ENDO and EXO. However, there seem to be consistently higher differencesbetween in- and out-group behaviour in treatment EXO compared to treatment ENDO. Forplayer A we find differences also in average levels: Player A’s seem to be somewhat more willingto let Player B choose in ENDO (compared to EXO) in positive reciprocity games and lesswilling in negative reciprocity games.

We define in-group bias as the difference in behaviour of player B between in-group andout-group matches. We only focus on Player B in measuring in-group bias since in-/out-group

10

differences in the behaviour of Player A could be due not only to a genuine tendency to dis-criminate between groups (which the notion of in-group bias aims to capture), but also to the(strategic) anticipation of differential behaviour by Player B in in/out-group matches. PlayerB, in contrast, moves last and has therefore no such strategic reasons, and any behaviour dif-ferences for player B are clean evidence of in-group biases. Since our design does not rely onthe strategy method, all participants play the games either in an in-group situation or in anout-group situation. One advantage is that our design is not leading, since it avoids conditionalchoices of the type ”if I was matched within the ”RED” group I would do this, otherwise I woulddo that”. Such contingent choices might be seen as suggestive, i.e. participants might expectthat they should choose differently just because a difference is made by the experimenter.6 Thedownside of our design is that we can’t directly observe in-group bias at the individual agentlevel. So, in measuring in-group bias we will compare the behaviour of different participants inin-group and in out-group matches.

4.1 In-Group Bias with Exogenous Matching

We start with treatment EXO, where agents have no way of affecting their matching probabilities.Table 4 shows the results of simple logit regressions where a binary variable, indicating whethera participant in the role of player B displayed altruistic, negatively or positively reciprocalbehavior is regressed on another binary variable that indicates whether behaviour took place inan in-group match (variable “in-group”). We clustered standard errors by individual, since weobserved each participant in two games for each category. Remember the categories are altruism(games 1,2), negative reciprocity (3,4), positive reciprocity (inequality increasing case 5,6) andpositive reciprocity (inequality decreasing 7,8).

We find that there is significantly less negative reciprocity in in-group compared to out-groupmatches (p < 0.002, see G3-G4 in Table 4) and more positive reciprocity in Games 5-6 (wherepositive reciprocity is inequality increasing) (p < 0.011, see Table 4). In particular there is a34 percent increase in positive reciprocity and a 40 percent decrease in negative reciprocity inin-group matches according to the marginal effects of the logit regression below. (See also figure3). Those results are consistent with Chen and Li (2009) who, with a within subject design, finda 19 percent increase in positive reciprocity and a 13 percent decrease in negative reciprocity inin-group compared to out-group matches. If anything, the in-group biases we find are slightlystronger. This is consistent with evidence from e.g. political science, where it has been foundthat people tend to suppress discriminatory tendencies when directly confronted with multipleoptions (Kuklinski et al. (1997)).

EXO G1-G2 G3-G4 G5-G6 G7-G8

constant −1.8245∗∗∗ 0.3364∗∗∗ −2.3978∗∗∗ −0.6931∗

in-group 0.9555 −1.8405∗∗∗ 2.1234∗∗ −0.6649

groups 40 40 40 40observations 80 80 80 80

Table 4: In-group Bias in EXO. Logit Regression with standard errors clustered by id. 40individuals in the role of Player B, 80 observations. 53 percent of all matches were in-groupmatches. ∗∗∗1%,∗∗5%,∗10% significance.

Table 5 shows the same regression except that we included additional variables from thequestionnaire as well as a dummy “RED” that takes the value 1 for participants from the REDgroup to see whether colour matters per se. The questionnaire variables are age, a gender

6Of course the opposite might also occur and participants might suppress discriminatory tendencies whendirectly confronted with conditional choices.

11

EXO G1-G2 G3-G4 G5-G6 G7-G8

constant −8.1378∗∗ 1.9952∗∗ −9.7469∗∗ 8.9449∗

in-group 1.8166∗∗ −1.5726∗∗∗ 2.1414∗∗∗ 1.0084age 0.0697 −0.1177 0.2888∗ −0.3150gender −1.1008 0.5008 −0.0190 −1.7060∗

RED 2.0637∗ −0.5904 0.2084 −0.4652risk aversion 0.5911∗∗ 0.2217 −0.0388 0.2656cognitive reflection 0.9022∗∗ 0.0548 0.5933∗ −0.5270


Table 5: In-group Bias in EXO. Logit Regression with additional variables (Standard errorsclustered by id). 39 individuals (groups) in the role of Player B, 78 observations. One individualdropped who preferred not to answer gender/age question in questionnaire. 53 percent of allmatches were in-group matches. ∗∗∗1%,∗∗5%,∗10% significance.

dummy (1=female), a measure of risk aversion and a measure of cognitive reflection. All thevariables are described in detail in Appendix D.

The main message is that in-group biases identified in Games 3-4 and 5-6 remain significantwhen these variables are introduced and the coefficients are of about the same size. In factstatistical significance is even higher for positive reciprocity when controlling for these variables(G5-G6). Interestingly a new in-group bias appears for the category Altruism (Games 1-2) whencontrolling for these variables. Participants are more altruistic towards in-group members. Inaddition, it seems that more risk averse, more cognitively reflected people and people from theRED group are more altruistic. However none of these variables can robustly explain behaviorin any of the other game categories. The single most important variable is the in-group dummy.

4.2 In-Group Bias with Endogenous Matching

We next measure in-group bias in treatment ENDO with endogenous matching, where agentscan affect the probability of an in-group (or out-group) match at a monetary cost. Our mainresult here is that when agents choose who to match with, the aggregate in-group bias eitherdiminishes or totally vanishes in statistical terms. This is illustrated in Figure 3, depictingin-group bias in both treatments EXO and ENDO, for the different pairs of games.

One important issue in comparing the EXO and ENDO treatments is that in the latter,whether or not a person ends up in an in-group match is not exogenous to behavior. In particular,the same characteristic that leads an agent to show differential behavior across in-group andout-group matches may affect that agent’s preference over who to be matched with.7 We willcome back to this question in detail below. To control for the willingness to pay, and thereforeisolate the effect of a match’s type on behavior, we have included the variable ”wtp × in-group”in some of the logit regressions reported in Table 6, which measures an agent’s willingness to payin order to be assigned to an in-group match (his/her degree of ”homophily”); The right columnfor each of the pairs of games shows coefficients where the wtp variable is omitted. We findthat, irrespective of whether we control for wtp × in-group or not, the coefficients on in-groupare not significant. The exception is a marginally significant positive bias for games 5 and 6(positive reciprocity), which disappears once we control for wtp. Still, because of these potentialendogeneity issues, the results in Table 6 should be read with care. As additional evidence,we also present the correlation coefficients from a Spearman correlation test (Table 8), wherewe correlate behaviour with the binary variable “in-group”. This table illustrates that biases(correlation coefficients) are uniformly larger in EXO compared to ENDO (in terms of their

7As we will show below, this is indeed the case. ”Homophilous” agents do behave differently from others.

12

Figure 3: In-Group Biases. The graph shows the difference in the percentages of B-playersdisplaying “altruism” (Games 1-2), “negative reciprocity” (Games 3-4), “positive reciprocity”(inequality increasing (Games 5-6)) or “positive reciprocity” (inequality decreasing (Games 7-8)) between in-group matches and out-group matches. On the left hand side is ENDO andon the right hand side is EXO. ∗∗∗ 1%,∗∗ 5%,∗ 10% significantly different from zero (minimumsignificance level across Tables 4, 5, 6 , 7 or 8).

absolute value) and are only statistically significant in EXO, with the exception of a marginallysignificant coefficient for Games 5 and 6 (positive reciprocity) in ENDO.

G1-G2 G1-G2 G3-G4 G3-G4

constant −0.9555∗∗∗ −1.3332∗∗∗ −1.3633∗∗∗ −1.3633∗∗

in-group −0.2615 −0.0430 0.9167 0.8933wtp x in-group 0.0094 −0.0011


G5-G6 G5-G6 G7-G8 G7-G8

constant −1.3633∗∗∗ −1.3633∗∗∗ 0.7777∗∗ 0.7777∗∗

in-group 0.9375 1.2091∗ 0.6442 0.0332wtp x in-group 0.0134∗∗ −0.0239


Table 6: Logit Regression with standard errors clustered by id. In-group Bias in ENDO. 40individuals (groups) in the role of Player B, 80 observations. 62 percent of all matches werein-group matches. ∗∗∗ 1%,∗∗ 5%,∗ 10%.

Table 7 shows the results of the same regression as the base regressions in Table 6, but againwe have included some variables from the questionnaire. In this regression as well we do notfind any significant in-group biases for treatment ENDO. The marginally significant coefficienton in-group in Games 5-6 disappears here and instead in-group appears as marginally significantin Games 3-4. Taken together there is no evidence for robustly significant in-group biases in

13

ENDO G1-G2 G3-G4 G5-G6 G7-G8

constant −2.3640∗∗ −1.1748∗∗ −0.5385∗∗ −0.1517in-group −0.1902 1.4295∗ 0.6632 0.3072age 0.1409 0.0446 −0.0152 0.0347gender −0.9131 0.4831 −0.1231 0.1231RED 0.2853 −0.7944 0.9275 −0.5058risk aversion −0.3137 −0.2068 −0.1492 0.1716cognitive reflection −0.3051 −0.1064 −0.2423 −0.0878


Table 7: In-group Bias in ENDO. Logit Regression with additional variables (Standard errorsclustered by id). 40 individuals (groups) in the role of Player B, 80 observations. 62 percent ofall matches were in-group matches. ∗∗∗1%,∗∗5%,∗10% significance.

ENDO EXO

Altruism (Games 1-2) −0.0314 0.2728∗

Neg. Reciprocity (Games 3-4) 0.2363 −0.4551∗∗∗

Pos. Reciprocity (I) (Games 5-6) 0.3103∗ 0.5392∗∗∗

Pos. Reciprocity (d) (Games 7-8) 0.0205 0.1870

Table 8: Spearman correlation coefficients. Correlation between the frequency of altruistic,negatively reciprocal etc.. behavior and in-group match dummy. 40 individuals in the role ofPlayer B, 80 observations. ∗∗∗ 1%,∗∗ 5%,∗ 10%.

ENDO.

4.3 Self Selection and Shifts in Behaviour

Having measured in-group bias as a difference in behaviour across diverse types of matches, weneed to be careful in interpreting the reduction of in-group bias from EXO to ENDO in termsof a change in individual behaviour. This is because the change in the matching institutioncomes with a reassortment of players’ types across in- and out-group matches; homophilous(heterophilous) agents tend to interact in in-group matches more (less) frequently in ENDOthan in EXO. But can this ”self-selection” of types alone explain the observed change in in-group bias? Or can we conclude that some shift in individual behaviour has happened as aresult of the change in the matching institution?

It must be first noted that while self selection may potentially drive the result, the way inwhich it operates is not obvious. For instance, it is clear that if homophilous agents displayedmore reciprocity in both in- and out-group matches compared to non-homophilous agents, thenwe would necessarily have an increase in in-group biases as more homophilous agents assemblein in-groups in ENDO compared to EXO. If, however, homophilous agents are less reciprocal inall matches compared to others, we would have a decrease in the bias from EXO to ENDO etc.Hence, ex ante, it is not clear in which direction self-selection would push in-group biases.

To rule out self-selection as the unique cause of the decline in in-group bias we have to answertwo questions: (i) are there possible realizations of the matching process that can produce theobserved bias in EXO when holding contingent behavior in ENDO for all three types constant(and placing a 95 percent confidence interval on the distribution of types)? and (ii) is contingentbehavior in in-group and out-group matches significantly different between ENDO and EXO?We answer these questions separately for the case of negative and positive reciprocity. Note thatby doing so we are testing a weaker hypothesis, since it is the same realization of the matchingprocess that has to produce both the bias in negative and positive reciprocity. We start with

14

(a) Negative Reciprocity (b) Positive Reciprocity

Figure 4: Levels of reciprocity in ENDO (by type of matching preference) and in EXO (rounded).Remember that 45 percent of agents are classified as homophilous and neutral, respectively and10 percent as heterophilous.

the case of negative reciprocity illustrated in Figure 4(a).

Negative Reciprocity For all types, negative reciprocity is weakly higher in in-group matchescompared to out-group matches in ENDO. At the same time, in the EXO treatment we observedsignificantly less negative reciprocity at the aggregate level in in-group compared to out-groupmatches . Furthermore, in ENDO all types are more negatively reciprocal in in-group matchescompared to average behavior in in-group matches in EXO (t-test for equality of means p =0.1015, one-sided: p = 0.0508). And all types are less negatively reciprocal in out-group matchesin ENDO compared to average behavior in out-group matches in EXO (t-test for equality ofmeans p < 0.0001). Hence, even allowing for any distribution of types, there is no realization ofthe matching process that could have produced this outcome without a shift in behavior.

Positive Reciprocity Here we see that all types are more positively reciprocal in out-groupmatches in ENDO compared to EXO (t-test for equality of means, p = 0.0537).8 This impliesagain that no realization of the matching process in EXO may have generated the observedaverage level of reciprocity, clearly demonstrating that self-selection alone cannot explain thechanges in aggregate behaviour we observe by moving from exogenous to endogenous matching.

We can now summarise the main insights we have obtained in this section about the role ofthe matching institution for in-group biases in socio-economic interaction.

Result 2: In-group bias and the matching institution

• Aggregate in-group biases vanish or decrease in the transition from exogenous to en-dogenous matching.

• Self selection alone cannot explain the change in in-group bias from exogenous toendogenous matching, and a shift in behaviour due to the change in the matchinginstitution must have occurred.

4.4 Homophily, Discrimination and the Matching Institution

The above results suggest that the variance in behaviour at society level between in-group andout-group matches should decrease as a result of endogenous matching. One way of interpreting

8We can also conduct this test separately for all types and find significance levels of p = 0.0929 (strictlyhomophilous) and p = 0.0656 (neutral) for the one-sided test, but no significance for heterophilous agents due totheir small number.

15

this result is to say that by letting social actors be in greater control of their interaction patternswe will decrease discrimination, as behavioural differences across types of matches vanish. So,while on the one hand endogenizing matches has the effect of increasing the degree of segregationof social contacts, on the other hand it may mitigate the degree of discrimination in socialinteraction through the joint effect of self selection and of the shift in behaviour describe above.

A related but conceptually different issue is whether increases in homophily would affectthe degree of discrimination for a given fixed matching institution and, if so, in what direction.Figure 5 shows in-group biases separately for the three different types of matching preferencesin negative and positive reciprocity games (3-4 and 5-6, respectively). In the positive reciprocitygames, we record a significant and positive in-group bias only for homophilous agents, while nobias is present on average for neutral and heterophilous agents. No significant bias is presentin negative reciprocity games. This evidence is consistent with the results of the regressionsreported in Table 6, where the only significant interaction effect between the degree of homophily(wtp) and bias was found for the case of positive reciprocity. The type of bias we find in thepositive reciprocity games is of the type one would naturally expect: agents that prefer to matchwith similar agents, are also prone to reward similar others in return for kind actions; by thesame logic agents that prefer to match with dissimilar others would be expected to reward similaragents less. We find evidence for both intuitions, although only in the case of homophilous agentsthe effect is statistically significant. Further evidence can be found from a Spearman correlationtest reported in Table 9.

Figure 5: In treatment ENDO in-group biases are only found when restricting to homophilousagents. ∗∗∗ 1%,∗∗ 5%,∗ 10%.

Homophilous WTP=0 Heterophilous

Neg. Reciprocity (Games 3-4) 0.3158 0.0558 0.3043Pos. Reciprocity (Games 5-6) 0.5190∗∗ 0.1484 -0.6667

Number of Observations (Share) 18(45%) 17(43%) 5(12%)

Table 9: Spearman correlation coefficients. Correlation between the frequency of altruistic,negatively reciprocal etc.. behavior and in-group match dummy. 40 individuals in the role ofPlayer B, 80 observations. ∗∗∗ 1%,∗∗ 5%,∗ 10%

Result 3 Under endogenous matching, in-group biases can be found only among homophilousagents, who are more positively reciprocal towards the in-group than towards the out-group.

16

5 On the Sources of Homophily: a Discussion

Within our minimal design, homophily is certainly not a matter of opportunities or stereotypes.Agents meet according to a matching process which is in all respects anonymous except for groupmembership (RED or BLUE) and the way in which agents affect probabilities according to theirdeclared wtp; likewise, the way in which agents are assigned to the BLUE and RED group isindependent of any individual characteristic of participants. In Section 3 we have also seen thatin our minimal design homophily is not a driven by the strategic reaction to the anticipationof in-group biases. In this section we develop and discuss some (evidence-based) conjectures onwhat the roots of homophily might be, and relate them to the results of the previous section.

We start by reporting the results from the post experimental questionnaire in which weelicited risk aversion of participants. As a measure of risk aversion we use a variable thatcounts in how many lottery choice questions in the questionnaire the participants preferred asure outcome to a non-degenerate lottery. (The questions can be found in the Appendix). Thevariable ranges from 0 (least risk averse) to 7 (most risk averse). Table 10 shows the distributionsof the variable risk aversion in treatments ENDO and EXO, which are remarkably similar.

ENDO EXO

0 (least risk averse) 0.10 0.101 0.07 0.082 0.29 0.213 0.31 0.344 0.18 0.175 0.02 0.056 0 07 (most risk averse) 0 0.01

Table 10: Distribution of the variable Risk aversion.

Participants with a higher willingness to pay for an in-group match are more risk averse(Spearman test ρ = 0.3586∗∗∗). Figure 6 illustrates the predicted values of a linear regression,where we explain the wtp for an in-group match via the risk aversion variable, as well as datapoints in risk aversion - wtp space. The figure omits strictly heterophilous agents and displaysfitted lines from two regressions - one without and one with a square term. Clearly, the wtp foran in-group match seems to increase with an agent’s risk aversion.

We also conducted a cognitive reflection test in the questionnaire and found no significantcorrelation between cognitive reflection and a binary measure of homophily nor between cognitivereflection and willingness to pay. Table 11 shows OLS regressions, where we regress the wtp forhomophilous agents on all our questionnaire variables. It can be seen that the only variable thatis significant is the degree of risk aversion of the participant. We run three different regressionswhere we control for different sets of questionnaire variables. The estimated coefficient on ourrisk aversion variable is pretty stable across all these.

Result 4 Homophily is positively correlated with risk aversion.

This evidence points to a possible psychological source of homophily. Suppose agents per-ceived the behavior of similar agents to be more predictable (for reasons to be discussed below).Those who are more risk averse would end up investing more in relationships with similar others(by declaring a higher wtp), in which they expect ex-ante less strategic uncertainty. Interestingly,an interpretation based on risk aversion as the primitive force behind homophily is consistentwith the recent sociological theory that explains group identification as a reflection of agents’desire to decrease their perceived uncertainty over the outcome of social interaction (Hogg and

17

Figure 6: Fitted values from a linear OLS regression of the wtp for an in-group match onRisk-Aversion.

(1) (2) (3)

constant 9.7662 0.4308 8.0820age 0.1219gender −9.0224risk aversion 7.8638∗ 9.6014∗∗ 8.8207∗∗

cognitive reflection 2.9185 4.4167

observations(groups) 32 32 32

Table 11: Homophily (wtp) . OLS Regression of wtp on questionnaire variables for 32 ho-mophilous participants in ENDO. ∗∗∗1%,∗∗5%,∗10% significance.

18

Abrams (1993), Hogg (2000)). But why should agents believe to better predict behavior inin-group matches?

One possible explanation is that agents expect in-group interaction to be regulated by behav-iors that stabilize outcomes on some implicit norm. Such behaviors may take the form of harshpunishments of bad behavior from in-group opponents, possibly grounded in the psychologicaldistress of having invested resources (wtp) in vain, or in a reaction to unexpected ill treatmentfrom other agents from the same group. It is not uncommon in reality to see in-group mem-bers being punished more harshly than out-group members for a given behaviour. Through selfselection, homophilous agents tend to cluster in in-group matches, forming disjoint close-knitcommunities within cultural groups. Such communities tend to be characterized by strong socialnorms, which are enforced by means of selective punishments and rewards. Such effects seem tounderly social phenomena such as, for instance, the punishing of acting white behavior whichrefers to a person’s perceived betrayal of identity.9

Figure 3 shows that, indeed, there might be more negative reciprocity in in-group matchescompared to out-group matches in ENDO, even though the effect is not significant. If we focusonly on game 3, though, (see Table 1), we do detect a marginally significant effect. Negativereciprocity is higher in game 3 in in-group compared to out-group matches (54% of type B agentsnegatively reciprocate in in-group matches in ENDO,as opposed to 26% in out-group matches;p = 0.086, Mann-Whitney ranksum test). Consistently with more negative reciprocity, player Aopts out, i.e. does not let player B choose, more often in in-group matches (56 percent of thetime) compared to out-group matches (46 percent) in ENDO (see Appendix A). This differenceis not significant, however. Note that wtp’s are declared prior to knowing which type of playerone will be (either A or B); therefore, more negative reciprocity leads to a better outcome fortype B if expected as a norm by type A, and to a worse outcome for type A, who is forced tobehave “nicely”, and, overall, to less variable payoffs ex-ante.

5.1 The Effect of the Introduction of Groups

Before we conclude we have a brief look at our CONTROL treatment where agents played thesame 8 games as in EXO and ENDO, but where no mention was made of different groups.Understanding how behaviour in EXO differs from CONTROL can help understand how theintroduction artificial group differences or identities affects behavior.

EXO in-group EXO out-group CONTROL

G1-G2 0.30 0.13 0.22G3-G4 0.18 0.58 0.26G5-G6 0.43 0.08 0.30G7-G8 0.79 0.66 0.74

Table 12: Comparison of Behavior EXO and CONTROL.

Table 12 shows the result of this comparison. In all four categories of games behavior inCONTROL lies in between average behaviour in in- and out-groups in EXO. This shows that- at least in our setting - the introduction of different identities does not lead to substantiallydifferent behaviour on average. An exception is the case of negative reciprocity (games 3-4),where behaviour in the CONTROL treatment is much closer to in-group behavior in EXOand significantly different from out-group behavior in EXO. Hence the introduction of artificialidentities leads to a loss in welfare (i) due to an increase in negative reciprocity as well as (ii)due to discrimination (different behaviour in in- and out-group matches) in our case. This istrue only, as our previous results have shown, under exogenous matching conditions.

9It is reasonable to assume, of course, that such heuristics are learned in real groups outside the laboratoryand triggered in the unfamiliar situation encountered in the laboratory.

19

6 Conclusions

We proposed a minimal design to study homophily in the laboratory and found evidence forhomophily in such a design even in the absence of stereotypes, differing opportunities or thestrategic anticipation of preferential treatment by others. We further studied the connectionbetween homophily and in-group bias. We found that they are tightly linked: Giving homophilya playing field (by allowing for endogenous matching) significantly reduces in-group biases.Furthermore in-group biases are only found among agents that are also homophilous. We alsofound that risk aversion is strongly correlated with homophily.

These results should be of interest to any social scientist studying discrimination, segregationand the like. They also have implications for a number of important policy dimensions, especiallyfor situations where matching is an issue. Those include matching workers into teams, childrento schools or social workers or field agents to different neighborhoods. Our results show thatallowing some degree of choice in who to work with, can reduce in-group biases and hencediscrimination at the work place. In applications, of course, other factor that were blended outin this experiment, need to be considered as well, such as information asymmetries betweendifferent groups or complementarities in skills etc.

Future research, both in the lab and in the field, is needed to understand the sources ofhomophily and the reasons why in-group biases decrease with homophily at work through en-dogenous matching. One direction for this research could lie in the theory brought forwardin this article which is based on two assumptions: In-group interactions among homophilousagents obey behavioral norms and deviations from these norms are harshly punished (hence thereduction in in-group bias). Secondly, these norms (seem to) make behavior in in-group matchesmore predictable, hence providing a motive for risk averse agents to be homophilous.

References

Ahn, T., R. Isaac, and T. C. Salmon (2009). Coming and going: Experiments on endogenousgroup sizes for excludable public goods. Journal of Public Economics 93(1-2), 336–351.

Alesina, A., R. Baqir, and W. Easterley (1999). Public goods and ethnic divisions. QuarterlyJournal of Economics 114(4), 1243–1284.

Alesina, A., A. Devleeschauwer, W. Easterley, S. Kurlat, and R. Wacziarg (2003). Fractional-ization. Journal of Economic Growth 8(2), 155–194.

Benjamin, D. J., J. Choi, and A. Strickland (2010). Social identity and preferences. AmericanEconomic Review 100(4), 1913–1928.

Bohnet, I., B. Herrmann, and R. Zeckhauser (2010). Trust and the reference point for trustwor-thiness in gulf and western countries. Quarterly Journal of Economics 125(2), 811–828.

Calvo-Armengol, A. and M. Jackson (2004). The effects of social networks on employment andinequality. American Economic Review 94(3), 426–454.

Centola, D. (2011). An experimental study of homophily in the adoption of health behavior.Science 334 (6060), 1269–1272.

Chakravarty, S. and M. A. Fonseca (2012). The effect of social fragmentation on public goodprovision: an experimental study.

Charness, G., R. Cobo-Reyes, and N. Jimenez (2011). Efficiency, team building, and identity ina public-goods game.

20

Charness, G. and M. Rabin (2002). Understanding social preferences with simple tests. AmericanEconomic Review 117(3), 817–869.

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

Chen, R. and Y. Chen (2012). The potential of social identity for equilibrium selection. AmericanEconomic Review forthcoming.

Chen, Y. and S. X. Li (2009). Group identity and social preferences. American EconomicReview 99:1, 431–457.

Coricelli, G., D. Fehr, and G. Fellner (2004). Partner selection in public goods experiments.Journal of Conflict Resolution 48(3), 356–378.

Currarini, S., M. Jackson, and P. Pin (2009). An economic model of friendship: Homophily,minorities and segregation. Econometrica 77(4), 1003–1045.

Currarini, S., M. Jackson, and P. Pin (2010). Identifying the roles of choice and chance innetwork formation: Racial biases in high school friendships. PNAS 107, 4857–4861.

Currarini, S. and F. Vega-Redondo (2011). A simple model of homophily in social networks.mimeo.

Easterley, W. and R. Levine (1997). Africa’s growth tragedy: policies and ethnic divisions.Quarterly Journal of Economics 112(4), 1203–1250.

Eckel, C. and P. Grossman (2005). Managing diversity by creating team identity. Journal ofEconomic Behavior and Organization 3, 371–392.

Fischbacher, U. (2007). z-tree: Zurich toolbox for ready-made economic experiments. Experi-mental Economics 10(2), 171–178.

Foddy, M., M. Platow, and T. Yamagishi (2009). Group-based trust in strangers. PsychologicalScience 20(4), 419–422.

Gaechter, S. and B. Herrmann (2009). Reciprocity, culture and human cooperation: previousinsights and a new cross-cultural experiment. Philosophical Transactions of the Royal SocietyB - Biological Sciences 364, 791–806.

Golub, B. and M. Jackson (2011). How homophily affects the speed of learning and best responsedynamics.

Greiner, B. (2004). An online recruitment system for economic experiments. In K. Kremer andV. Macho (Eds.), Forshcung und wissenschaftliches Rechnen. GDWG Bericht 63.

Grimm, V. and F. Mengel (2009). Cooperation in viscous populations - experimental evidence.Games and Economic Behavior 66(1), 202–220.

Hamilton, W. (1964). The genetic evolution of social behaviour. Journal of Theoretical Biology 7,1–52.

Hogg, M. (2000). Subjective uncertainty reduction through self-categorization: a motivationaltheory of social identity processes and group phenomena. European Review of Social Psychol-ogy 11, 22355.

Ioannou, C. and A. Rustichini (2012). Group outcomes and reciprocity.

21

Jackson, M. and D. Lopez-Pintado (2011). Diffusion and contagion in networks with heteroge-neous agents and homophily.

Keser, C. and F. van Winden (2000). Conditional cooperation and voluntary contribution topublic goods. Scandinavian Journal of Economics 102(1), 23–29.

Kuklinski, J. H., P. M. Sniderman, K. Knight, T. Piazza, P. Tetlock, G. Lawrence, andB. Mellers. (1997). The politics of affirmative action. American Journal of Political Sci-ence 41, 402–419.

Mc Pherson, M., L. Smith-Lovin, and J. Cook (2001). Birds of a feather: Homophily in socialnetworks. Annual Review of Sociology 27, 415–444.

Mengel, F. (2008). Matching structure and the cultural transmission of social norms. Journalof Economic Behavior and Organization 67, 608–623.

Moody, J. (2001). Race, school integration and friendship segregation in america. AmericanJournal of Sociology 107(3), 679–716.

Myerson, R., R. Pollock, and J. Swinkels (1991). Viscous population equilibria. Games andEconomic Behavior .

Skopek, J., F. Schulz, and H. Blossfeld (2011). Who contacts whom? educational homophily inonline mate selection. European Sociological Review 27, 180–195.

Tajfel, H. and J. Turner (1986). The social identity theory of inter-group behavior. Chicago:Nelson Hall.

Tajfel, H. and J. C. Turner (1979). An integrative theory of intergroup conflict. In The socialpsychology of integroup relations. ed. W.G. Austin and S. Worchel Monterey, CA: Brooks/Cole.

Van Bavel, J., D. Packer, and W. Cunningham (2008). The neural substrates of in-group bias.Psychological Science 19, (11).

22

A Appendix: Behavior Player A

Player A’s decision to let Player B decide or not can be motivated by (i) concerns on thedistribution of payoffs (such as efficiency, equality etc.) and (ii) by anticipation of Player B’sbehaviour. If a participant in the role of Player A lets Player B choose in Game 3, then wecan reasonably interpret this to mean that s/he does not expect B to be negatively reciprocal.If s/he lets Player B choose in Games 7-8, then we can interpret this to mean that s/he doesexpect Player B to be positively reciprocal. In Games 4,5 or 6 a case could be made that PlayerA may let B choose because this will be inequality reducing. This makes player A’s behaviourmore difficult to interpret in terms of in-group bias, which is why we have focused on Player B’sdecision for the main part of the paper.

Games 3 and 4 - Negative Reciprocity There are no significant differences between in-and out-group matches in terms of the behavior of Player A in ENDO, where 56 percent and46 percent of them, respectively, decide not to let player B make a choice (the difference ismarginally significant for game 3 alone). There is also no significant difference between in- andout-group matches in EXO, where 38 percent and 53 percent respectively decide not to let playerB make a choice. The difference between ENDO and EXO is marginally significant for in-groupmatches and insignificant for out-group matches.

Games 5 and 6 - Positive Reciprocity In ENDO there are strong and significant differencesbetween in-group and out-group matches. In out-group matches only about 10 percent of agentslet Player B choose, while in in-group matches 62 percent let Player B choose (p = 0.0010). Intreatment EXO there are no significant differences between in- and out-group matches with 33percent and 25 percent respectively letting Player B choose.

Games 7 and 8 - Positive Reciprocity In ENDO only 7 percent of agents let Player Bchoose in out-group matches, while 38 percent let Player B choose in in-group matches. Thedifference is significant (p = 0.0273). In treatment EXO 3 percent let Player B choose in out-group matches, while 14 percent let Player B choose in in-group matches. This difference is notsignificant.

B Appendix: Effect of Group Labels

We first ask whether people have a higher willingness to pay to play with a RED rather than theBLUE group. The average willingness to pay to play with the RED group is 20.41 in ENDO andthe average w.t.p to play with BLUE is 19.77. The two are not significantly different accordingto a Mann-Whitney test. We then ask whether behavior is affected by being in a RED or BLUEgroup per se.

red blue

Altruism 0.23 0.15

Neg. Rec. (Games 3-4) 0.32 0.32

Pos. Rec. (Games 5-6) 0.35 0.16

Pos. Rec. (Games 7-8) 0.67 0.76

Table 13: Colours don’t matter per se. ENDO

There are no significant differences between blue and red players in either of the treatmentsand for any of the games with the exception of Altruism in treatment EXO. The colour of thegroup does not seem to matter per se.

23

red blue

Altruism 0.23 0.12

Neg. Rec. (Games 3-4) 0.31 0.43

Pos. Rec. (Games 5-6) 0.29 0.28

Pos. Rec. (Games 7-8) 0.71 0.73

Table 14: Colours don’t matter per se. EXO

C Appendix: Distribution of wtp

Figure 7 shows the distribution of the variable wtp. It peaks at zero, reflecting the relativelyhigh percentage of neutral agents. There are few heterophilous agents, whose wtp ranges from0 to 50 and many more homophilous agents whose wtp ranges from 0 to 100.

Figure 7: The distribution of willingness to pay in treatment ENDO. Negative Numbers indicatewtp for an out-group match and positive numbers wtp for an in-group match. Bin width is 10.

D Appendix: Questionnaire

We first report the questions asked in the questionnaire and then describe the variables derivedfrom it.

D.1 Questions

• General Questions with multiple options to tick.

– What is your nationality?

– What is your gender?

– What is your age?

– What is your field of study?

24

• Cognitive Reflection Test

– A bat and a ball cost 1.10 Euros in total. The bat costs 1 Euro more than the ball.How much does the ball cost?

– If it takes 5 machines 5 minutes to make 5 widgets, how long would it take 100machines to make 100 widgets?

– In a lake, there is a patch of lily pads. Every day, the patch doubles in size. If ittakes 48 days for the patch to cover the entire lake, how long would it take for thepatch to cover half of the lake?

• Risk Aversion

– If you had the choice between throwing a coin and receiving 100 Euros if heads comesup and 0 Euros if tails come up or a deal where you get 10 Euros for sure which wouldyou prefer?







D.2 Overview Variables

interpretation range EXO mean EXO range ENDO mean ENDOrisk aversion 7 most risk averse, 0 least [0,7] 2.63 [0,5] 2.48gender 1= female , 0=male [0,1] 0.55 [0,1 0.65age [18,25] 21.71 [18,29] 21.66cognitive reflection number of correct answers [0,3] 1.71 [0,3] 1.40

Table 15: Questionnaire Variables used in Regressions Section 3.

E Appendix: Additional Treatment LABEL

Treatment LABEL coincides with ENDO, but we labeled the actions in the 8 games by eitherRED or BLUE to see whether such labels can induce different norms of behaviour in the twogroups and whether such norms could possibly crowd out existing norms of behaviour. Ourevidence shows that this was not the case.

25

red blue

prefer in-group 0.65 0.65

wtp in-group 14.4 12.41

wtp out-group 0.75 6.58

Table 16: Percentage of red and blue participants preferring an in-group match, as well asaverage willingness to pay for a in-group or out-group match in Period 1.

None of the differences between red and blue participants are significant (Mann-Whitney,p > 0.7267). However there is a higher willingness to pay for a in-group than for a out-groupmatch across all participants (Mann-Whitney, all periods, p = 0.0092).

red blue

Altruism 0.20 0.30

Neg. Recip. 0.25 0.30

Pos. Recip. 0.30 0.37

Ant. PosR 0.15 0.18

Ant. PosR 0.15 0.18

Table 17: Are labels followed? Altruistic and Reciprocal Behavior by Red and Blue participantsas well as Anticipation of Positive and Negative Reciprocity (Player A) if the match is blue/red.

The table shows that labels are not followed. Red participants are not more altruistic orreciprocal than blue participants. (None of the differences are significant).

F Sample Instructions

F.1 ENDO

Welcome and thanks for participating in this experiment. Please read these instructions care-fully. They are identical for all the participants with whom you will interact during this exper-iment. If you have any questions please raise your hand. One of the experimenters will cometo you and answer your questions. From now on communication with other participants is notallowed. If you do not conform to these rules we are sorry to have to exclude you from theexperiment. Please do also switch off your mobile phone at this moment.

For your participation you will receive 2 Euros. During the experiment you can earn more.How much depends on your behavior and the behavior of the other participants. During theexperiment we will use ECU (Experimental Currency Units) and at the end we will pay youin Euros according to the exchange rate 1 Euro =500 ECU. All your decisions will be treatedconfidentially.

These instructions are exactly the same for all participants in the room.

THE EXPERIMENT

Phase 1: The RED and BLUE groups

At the beginning of the experiment each participant receives an endowment of 500 ECU.All participants in this room are randomly assigned to one of two groups, the BLUE group

and the RED group. Once the software starts you will be informed about which group youbelong to.

Phase 2: Your choice of who to interact with

26

In Phase 3 of the Experiment that we describe below you will interact in a number of games.Before that - in Phase 2 - you can state a preference for whom to interact with. In particularyou can first state a preference for whether you would like to interact with someone from theRED or BLUE group.

Afterwards you will see a new screen. If you have previously stated that you would like tointeract with someone from a particular group (either BLUE or RED), then you will have thepossibility to increase the chances that you will actually be matched with someone from thatgroup on the third screen.

You can increase your chances to be matched with someone from your desired group asfollows:

First you will indicate a number between 0 and 100 that indicates how much you would bewilling to pay to be matched with your preferred group that you indicated on Screen 2. We willdraw a random number between 0 and 100 (where all numbers have the same probability). If ournumber is smaller than your number, you will be matched for sure with your preferred group.In this case an amount corresponding to your number will be deducted from your endowment.If our number is larger than your number you will be matched randomly with a member ofeither group. Example: Here is an example of how the procedure works: Assume you indicatethe number 21 and we draw randomly the number 46. Then since 46 > 21 you will be matchedrandomly with someone from either the RED or the BLUE group. However if e.g. you indicate83 and we draw randomly the number 16, then since 83 > 16 you will be matched with thegroup you prefer. However, in this case your endowment will be reduced by 83 (the number thatyou stated).

Hence the higher the number you state, the more likely it is that you will be matched withyour desired group. On the other hand your endowment is also reduced by more if you state ahigher number. There is a small probability that it is not feasible to match everyone with theirdesired group (e.g. if everyone states 100). If this is the case you will be randomly matched andyour endowment will not be reduced.

Phase 3: The Games

In Phase 3 you will play the following 8 games together with an interaction partner. Whetheryour interaction partner will be from the BLUE or RED group will depend on your choices inPhase 2. Before Phase 3 begins you will be informed about which group your interaction partnerbelongs to. Similarly your interaction partner will be informed about which group you belongto.

Here is a list of all 8 games you will play. You will play each game exactly once. Before thegames begin either you or your interaction partner will be randomly determined to be PlayerA. If you are not Player A then you will be Player B in all 8 games. Whether you or yourinteraction partner will be Player A is fully independent of your or your interaction partner’sgroup.

1. Game 1: Player B chooses between two outcomes. (400,400) meaning that each playergets 400 or (750,375) meaning that Player B gets 375 and Player A 750.

2. Game 2: Player B chooses between two outcomes. (100,300) meaning that player A gets100 and Player B 300 or (400,200) meaning that Player A gets 400 and Player B 200.

3. Game 3: Player A chooses between two options. (250,250) meaning that each player gets250 or s/he lets player B choose. If s/he lets B choose then B can choose between (100,100)meaning that each player gets 100 and (500,100) meaning that player A gets 500 and playerB gets 100.

4. Game 4: Player A chooses between two options. (50,650) meaning that player A gets 50and player B 650 or s/he lets player B choose. If s/he lets B choose then B can choose

27

between (0,100) meaning that A gets nothing and B gets 100 and (100,100) meaning thateach player gets 100.

5. Game 5: Player A chooses between two options. (500,0) meaning that player A gets 500and player B gets 0 or s/he lets player B choose. If s/he lets B choose then B can choosebetween (300,300) meaning that each player gets 300 and (600,275) meaning that playerA gets 600 and player B gets 275.

6. Game 6: Player A chooses between two options. (250,0) meaning that player A gets 250and player B gets 0 or s/he lets player B choose. If s/he lets B choose then B can choosebetween (100,100) meaning that each player gets 100 and (250,50) meaning that player Agets 250 and player B gets 50.

7. Game 7: Player A chooses between two options. (350,100) meaning that player A gets350 and player B gets 100 or s/he lets player B choose. If s/he lets B choose then B canchoose between (300,300) meaning that each player gets 300 and (100,350) meaning thatplayer A gets 100 and player B gets 350.


It is important to note that you won’t get any feedback about what your interaction partnerchose until the last game has been played.

Phase 4: A post-experimental Questionnaire

At the end of the experiment there will be a short questionnaire to be filled in which willtake about 5 minutes.

Summary

1. First you will be informed about which group (RED or BLUE) you belong to.

2. You can state a preference for being matched with someone of the BLUE or RED groupand you can increase your chances of being matched with someone of that group via theprocedure explained above.

3. You will play the 8 games described above with your interaction partner. Whether yourinteraction partner is from the RED or BLUE group will depend on your choices in Step2.

4. There will be a short questionnaire at the end.

Enjoy the Experiment!

F.2 EXO


28



THE EXPERIMENT

Phase 1: The RED and BLUE groups

All participants in this room are randomly assigned to one of two groups, the BLUE groupand the RED group. Once the software starts you will be informed about which group youbelong to.

Phase 2: Your match

In Phase 2 of the Experiment you will be randomly matched with an interaction partnerfrom either the RED or BLUE group to play the games we will describe below.

Phase 3: The Games

Before the start of Phase 3 you will be informed about whether your interaction partner isfrom the BLUE or RED group. Similarly your interaction partner will be informed about whichgroup you belong to.

Here is a list of all 8 games you will play. You will play each game exactly once. Before thegames begin either you or your interaction partner will be randomly determined to be PlayerA. If you are not Player A then you will be Player B in all 8 games. Whether you or yourinteraction partner will be Player A is fully independent of your or your interaction partner’sgroup.




4. Game 4: Player A chooses between two options. (50,650) meaning that player A gets 50and player B 650 or s/he lets player B choose. If s/he lets B choose then B can choosebetween (0,100) meaning that A gets nothing and B gets 100 and (100,100) meaning thateach player gets 100.



29




Phase 4: A post-experimental Questionnaire


Summary

1. First you will be informed about which group (RED or BLUE) you belong to.

2. You will be matched with someone of the BLUE or RED group.

3. You will play the 8 games described above with your interaction partner. Whether yourinteraction partner is from the RED or BLUE group will be determined in Step 2.

4. There will be a short questionnaire at the end.


F.3 CONTROL




THE EXPERIMENT

In the experiment you will play 8 games with another randomly selected participant. Hereis a list of all 8 games you will play. You will play each game exactly once. Before the gamesbegin either you or your interaction partner will be randomly determined to be Player A. If youare not Player A then you will be Player B in all 8 games. Whether you or your interactionpartner will be Player A is fully independent of your or your interaction partner’s group.


30



4. Game 4: Player A chooses between two options. (50,650) meaning that player A gets 50and player B 650 or s/he lets player B choose. If s/he lets B choose then B can choosebetween (0,100) meaning that A gets nothing and B gets 100 and (100,100) meaning thateach player gets 100.






A post-experimental Questionnaire



31

G Sample Games LOWB and COORD

Tables 18 and 19 show sample games from treatments LOWB and COORD.

X Y

X 0,0 200,200

Y 200,200 0,0

Table 18: Game 1 in treatment LOWB

In treatment LOWB, players played a series of pure (anti-) coordination games where playershave to choose different actions to reach the good payoffs. There is no possibility for a playerto favor an in-group opponent here by choosing either X or Y.

BLUE RED

RED 0,0 200,200

BLUE 200,200 0,0

Table 19: Game 1 in treatment COORD

In treatment COORD players played the same games as in LOWB. However, actions hereare labeled by group labels RED or BLUE, in such a way that - if players choose the action thatcorresponds to their group label - players of the same group will manage to coordinate whileplayers of different groups won’t.

32

Identity, Homophily and In-Group Bias *

Documents

Transcript of Identity, Homophily and In-Group Bias *