Social Psychology Quarterly 2015, Vol. 78(2) 127–150 Understanding the Ó American Sociological Association 2015 DOI: 10.1177/0190272514565252 Selection Bias: Social http://spq.sagepub.com Network Processes and the Effect of Prejudice on the Avoidance of Outgroup Friends

Tobias H. Stark1,2

Abstract Research has found that prejudiced people avoid friendships with members of ethnic out- groups. Results of this study suggest that this effect is mediated by a process. Longitudinal network analysis of a three-wave panel study of 12- to 13-year-olds (N = 453) found that more prejudiced majority group members formed fewer intergroup friendships than less prejudiced majority group members. This was caused indirectly by the preference to become friends of one’s friends’ friends (). More prejudiced majority members did not have a preference for actively avoiding minority group members. Rather, they had the tendency to avoid friends who already had minority group friends and thus could not be intro- duced to potential minority group friends. Instead they became friends with the majority group friends of their friends. This research shows how a social networks perspective can fur- ther our understanding of the processes underlying intergroup contact.

Keywords social networks, friendships, dynamic processes, intergroup contact, prejudice

Research has established that intergroup outgroup (Pettigrew et al. 2011). This contact and particularly intergroup selection bias leads to a vicious circle: friendships reduce prejudice toward other intergroup friendships reduce prejudice racial or ethnic groups (Davies et al. 2011; most effectively (Hodson 2011; Paluck Pettigrew and Tropp 2006). It also has and Green 2009), but prejudice prevents been demonstrated that prejudiced people the development of intergroup friend- avoid outgroup friends (Binder et al. ships (Hewstone and Swart 2011). 2009; Eller and Abrams 2003; Levin, van Laar, and Sidanius 2003; Sidanius et al. 1Utrecht University/ICS, the Netherlands 2008). Those who already have more pos- 2Stanford University, Stanford, CA, USA itive attitudes toward the outgroup form Corresponding Author: intergroup friendships, while those with Tobias Stark, ERCOMER, Utrecht University, less positive attitudes toward the out- Padualaan 14, CH Utrecht 3584, the Netherlands. group tend to avoid members of the Email: [email protected]

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 128 Social Psychology Quarterly 78(2)

Most previous research had an individ- The present research goes beyond the ualistic theoretical perspective and con- existing empirical and theoretical sidered friendships as being the result of approaches by testing whether such net- individual attitudes and preferences. work processes are involved in intergroup The development of intergroup friend- friendship dynamics. ships was examined as if friendships are independent from each other. However, Intergroup Contact and the Question friendships develop within larger social of Causality networks (Pettigrew et al. 2007). With whom we become friends is not only based Decades of research on the contact on individual characteristics (e.g., ethnic- hypothesis (Allport 1954) has established ity or prejudice) but is partly the result of evidence for a correlation between more who approaches us and to whom we are intergroup contact and less prejudice. introduced by others (Goodreau, Kitts, Yet, contact with outgroup members and Morris 2009; Wimmer and Lewis might reduce prejudice and less preju- 2010). Friendships form because people diced people might also engage more tend to return friendship invitations (rec- readily in intergroup contact. In recent iprocity) and we are likely to become years, longitudinal studies have started friends with the friends of our friends (tri- to examine these causal processes. All of adic closure; see Cartwright and Harary these studies found that intergroup con- 1956; Wasserman and Faust 1994). Ignor- tact, particularly in the form of inter- ing network processes such as reciprocity group friendships, reduces prejudice and and triadic closure in the formation of thus confirmed the main premises of the social networks, and thus the develop- contact hypothesis (e.g., Brown et al. ment of intergroup friendships, may let 2007; Feddes, Noack, and Rutland 2009). researchers wrongly attribute friendship Less consistent are the findings about formation to individual preferences (Wim- the selection bias, the reversed causal mer and Lewis 2010). process from prejudice to less contact In contrast to previous interpretations over time. Some studies using relatively of the selection bias (e.g., Binder et al. small samples (N \ 110) found that con- 2009; Pettigrew et al. 2011; Sidanius et tact reduces prejudice and that prejudice al. 2008), more prejudiced people may has no effect on contact (Brown et al. not actively avoid outgroup friends just 2007; Feddes et al. 2009; Vezzali, Giovan- as less prejudiced people may not actively nini, and Capozza 2010). In contrast, approach outgroup friends. Rather, more research with larger samples found prejudiced individuals may not receive effects in both directions. For instance, friendship invitations from outgroup among a sample of over 2,000 U.S. college members, which they could reciprocate. students, having outgroup friends or Alternatively, more prejudiced individu- roommates from other racial groups als may select themselves into network reduced prejudice, but prejudice also positions where they have fewer opportu- was associated with fewer friends from nities to establish intergroup friendships. other racial groups a year later (Levin et In contrast, less prejudiced people may al. 2003; Sidanius et al. 2008). Further- form intergroup friendships because more, in a sample of 465 minority group they receive friendship invitations from high school students in South Africa, outgroup members or because they select intergroup anxiety led to fewer cross- themselves into network clusters where group friendships (Swart et al. 2011). they are introduced to outgroup members. There was, however, no effect of another

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 129 measure of outgroup attitudes on subse- Hewstone 2011) or simply because these quent contact. Additionally, a study with ingroup members can introduce a person 1,655 school students in 3 European coun- to outgroup members. In contrast, indi- tries found that effects from prejudice to viduals might be discouraged from seek- contact were either as strong as the ing intergroup contact if none of their effects from contact to prejudice or even ingroup friends have contact with out- stronger (Binder et al. 2009). Since stud- group members. Furthermore, driven by ies with larger samples and thus more their attitudes toward the outgroup, statistical power have produced evidence some people may decide to join or avoid for the contact effect and the selection friendship that already involve bias, I expect: outgroup members, making it subse- quently easier or more difficult to estab- Hypothesis 1: Having more outgroup lish actual intergroup contact. friends reduces prejudice toward the Ignoring such social network processes outgroup (contact effect). may lead to inadequate theoretical con- Hypothesis 2: More prejudiced individu- clusions. For instance, friendships that als tend to select fewer outgroup mem- actually form within an ethnic group in bers as friends than less prejudiced response to friendship invitations (reci- individuals (selection bias). procity) or because people have a friend in common (triadic closure) can be A Social Network Perspective wrongly attributed to individual prefer- Social psychological research on the con- ence for ingroup friends (ethnic homo- tact hypothesis predominantly focuses phily) if such network processes are over- on individual attitudes and preferences: looked (Wimmer and Lewis 2010). Outgroup contact, particularly in the Likewise, the tendency of more prejudiced form of a friendship, affects an individu- individuals to avoid friends from other al’s attitude and an individual’s prejudi- ethnic groups may be weaker than previ- cial attitude is responsible for not having ously assumed. More prejudiced individu- outgroup friendships. Friendships, how- als may receive fewer friendship invita- ever, require at least two individuals: tions from outgroup members that the ability of one person to develop an they can reciprocate and they may not intergroup friendship depends on the have friends who can introduce them willingness of an outgroup member to to outgroup members. Although contact reciprocate this contact. research has seen important theore- A social network perspective allows tical and methodological improvements examining such dyadic processes and (Christ and Wagner 2013), cross-lagged also theoretical considerations that go models that have been used in research beyond two persons. Real-life contact on the selection bias have not adequately between two individuals does not take considered these network processes in the place in a social vacuum but in social set- formation of intergroup friendships (Steg- tings that involve other people (Pettigrew lich, Snijders, and Pearson 2010). 2008; Pettigrew et al. 2007). Ingroup members may have already established contact with outgroup members, and Reciprocity as Mediator this might facilitate the intergroup con- Figure 1 illustrates how the exact tact of an individual because existing same intergroup friendship can develop contact of ingroup members implies pro- through different processes. Imagine social ingroup norms (Dovidio, Eller, and there are two white individuals A and B,

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 130 Social Psychology Quarterly 78(2)

Figure 1. Three Different Processes That Could Lead to an Intergroup Friendship Note: Arrows indicate friendship nominations. of which person A has positive attitudes but does not send an invitation to person toward blacks whereas person B has neg- B because of B’s more negative outgroup ative attitudes. Panel I shows the devel- attitudes. Person A then forms the inter- opment of a friendship between individ- group friendship because he or she recipro- ual A and a black person C between cated the friendship invitation, whereas Time 1 and Time 2. If social network pro- individual B never received a friendship cesses are not considered, the more posi- invitation that could have been recipro- tive attitudes of A remain as the only cated. Accordingly, I expect that: explanation for the fact that there is a new friendship between A and C at Hypothesis 3: Reciprocation of friend- ships leads to fewer intergroup friend- Time 2 but not between B and C. ships for more prejudiced individuals Panel II of Figure 1 shows how the same than for less prejudiced individuals. friendship could have developed without being directly caused by A’s or B’s attitude What mechanism could bring more toward blacks. At Time 1, individual C prejudiced individuals into network posi- sends a friendship invitation to person A tions where reciprocation of friendships because of A’s positive outgroup attitudes leads to fewer intergroup friendships?

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 131

Figure 2. Network Processes May Mediate the Seemingly Direct Effect from Prejudice to Fewer Intergroup Friendships (path c to c’) Note: Arrows indicate causal paths.

According to the Information Search Thus, a more prejudiced person may Model (Vorauer 2006), people in inter- not actively avoid a certain individual of group contact situations often have another ethnic group as it was previously concerns about how they are evaluated suggested (e.g., Binder et al. 2009; Petti- by outgroup members. This is particu- grew et al. 2011; Sidanius et al. 2008). larly true for members of the lower- Instead, the more prejudiced person’s status minority group because they remarks about the other group in general attach higher importance to the opinions may fuel feelings of being rejected in the held by the majority group (Vorauer outgroup member. Vorauer and Sakamoto 2013). People who experience such evalu- (2006) have shown that people who think ative concerns pay close attention to the outgroup member’s expressions and that they are rejected by a member of behaviors in order to detect cues about another group respond with an actual how they are being perceived. Conse- rejection of that person as a friend. Thus, quently, minority group members are outgroup members may not offer friend- likely to take a majority group member’s ships to more prejudiced individuals comments more personally than they are (path a in Panel I of Figure 2). This leads actually intended (Vorauer 2013). to the following hypothesis:

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 132 Social Psychology Quarterly 78(2)

Hypothesis 4a: Outgroup members are which they are friends with their friends’ less likely to select more prejudiced friends, or foes with the foes of their individuals as friends than less preju- friends, or friends with the foes of their diced individuals. foes. Thus, individual A may have formed As a consequence, more prejudiced a friendship with C at Time 2 because of individuals cannot form intergroup triadic closure. In contrast, the more prej- friendships in response to friendship invi- udiced individual B was not a friend of tations from outgroup members (path b).1 person D at Time 1. Person B is then not Instead, reciprocation of friendship invi- introduced to the black person C and tations will only lead to new ingroup thus never experienced an unbalanced friendships for more prejudiced individu- state in the friendship network that could als. In this case, the selection bias may have been resolved by forming an inter- be better described as a being-selected group friendship. Based on this reason- bias. ing, I expect that:

Hypothesis 4b: More prejudiced indi- Hypothesis 5: Triadic closure leads to viduals receiving fewer friendship fewer intergroup friendships for more invitations from outgroup members prejudiced individuals than for less than less prejudiced individuals prejudiced individuals. mediates the selection bias. This can, however, only happen if more prejudiced individuals do not have friends in common with outgroup members. Triadic Closure as Mediator

Another way in which network processes Hypothesis 6a: More prejudiced individu- may indirectly cause and prevent inter- als are more likely than less preju- group friendships is shown in Panel III diced individuals to be in network of Figure 1. At Time 1, individual A may positions where their friends have already have a friendship with a third fewer outgroup friends. person D, who in turn is a friend of the The question then arises as to what black person C. Structural mechanism brings more prejudiced indi- states that people tend to close such open viduals in network positions where they friendship triangles to avoid an unbal- are less likely to be introduced to out- anced state of their friendship network, group members. According to the homo- in which person A is a friend of D but phily principle, people prefer to befriend not of C (Cartwright and Harary 1956; similar others (McPherson, Smith-Lovin, Davis 1963). This notion builds on bal- and Cook 2001; Wimmer and Lewis ance theory in which unbalanced states 2010). More prejudiced individuals may lead to cognitive dissonance (Heider draw conclusions about potential friends’ 1946, 1958). To avoid cognitive disso- outgroup attitudes from these potential nance, people strive for situations in friends willingness to engage with out- 1The reciprocity mechanism also implies that group members. More prejudiced individ- more prejudiced individuals would reciprocate uals may thus infer that others who a friendship invitation if they had been selected already have outgroup friends must by an outgroup member. This is in line with con- have positive outgroup attitudes. Because tact theory’s premise that more prejudiced indi- of the apparent dissimilarity in their atti- viduals are open to contact and even friendships with outgroup members (Brown and Hewstone tudes, they may then avoid those who 2005). already have outgroup friends (see path

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 133 a in Panel II of Figure 2). The same pre- Stochastic actor-based longitudinal diction follows from balance theory network analysis (Snijders, Van de Bunt, (Heider 1946, 1958). According to this and Steglich 2010; Steglich et al. 2010) theory, more prejudiced individuals do was used to test how lagged outgroup atti- not want friends who have outgroup tudes and preexisting friendships affected friends and are therefore less prejudiced the development of intergroup friendships. because this would lead to cognitive disso- This approach allows taking the interre- nance; they would like their new friends latedness of friendships in a network into but they would disagree on their attitudes account and reduces the risk of wrong con- toward the outgroup. clusions due to reversed causality (Schae- fer, Kornienko, and Fox 2011; Snijders Hypothesis 6b: More prejudiced individu- et al. 2010). als are more likely than less preju- diced individuals to actively avoid METHOD friends who have outgroup friends. Participants If people subsequently become friends Data were drawn from the secondary- with their friends’ friends, more preju- school module of the Arnhem School diced individuals can only form addi- Study (Stark and Flache 2012), a longitu- tional ingroup friendships, while less dinal study of students’ social networks prejudiced individuals can form inter- and interethnic attitudes in their first group friendships with the outgroup years of secondary education (age 12– friends of their new friends (path b in 13). The first wave took place in the sec- Figure 2). ond and third weeks after the start of sec- ondary education (September 2008). Hypothesis 7: More prejudiced individu- als’ avoidance of friends who have out- Wave 2 was conducted about three group friends mediates the selection months later (December 2008), and bias. Wave 3 took place about six additional months later (June 2009). In Arnhem, a midsized city in the Netherlands, 61 THE PRESENT STUDY (88 percent) of all first-year classes in sec- The present study makes use of sociomet- ondary schools took part in this study. ric data collected in the Netherlands to Response rates of the remaining students test whether the selection bias is due to were 95 percent at Wave 1, 93 percent at underlying network processes. Dutch Wave 2, and 88 percent at Wave 3. majority group students in twenty classes The present research focuses only on from eight middle schools nominated those Dutch majority group students their best friends among their classmates who had at least one Turkish or Moroccan and indicated their attitudes toward the classmate of each gender.2 A well-known ethnic minority groups of Turkish and preference for same-gender friends (e.g., Moroccan people (outgroup prejudice). Schaefer et al. 2011; Vermeij, Van Duijn, These groups represent the largest minor- and Baerveldt 2009) may otherwise dis- ity groups in the city in which the data tort the results. That is, if all Turkish or were collected. Supplementary analyses, 2 which are shown in the material associ- Most ethnic minority students were born in the Netherlands and will have the Dutch citizen- ated with this article online, found no ship. The label Turks or Moroccans hence indi- selection bias among ethnic minority cates an ancestral connection to Turkey or students. Morocco.

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 134 Social Psychology Quarterly 78(2)

Moroccan students in a class were Only positive traits were included females, one might observe Dutch boys because developmental research indi- without interethnic friendships. This cates that children older than seven could be mistaken for the selection bias years are less willing to discriminate when in fact it is due to the boys’ prefer- between social groups in terms of nega- ence for male friends. I excluded from tive dimensions, whereas they will do the analyses four classes with Turkish so in terms of positive traits (e.g., Rut- or Moroccan students of both genders land et al. 2007). All items were rela- because little change in the social net- tively normally distributed around the work or students’ outgroup attitudes led midpoint of the answer scale and the to convergence problems. Accordingly, interitem correlations were below .90 the final sample size was 453 students with a few exceptions. Descriptive statis- in 20 school classes. A replication of tics for the individual items are pre- results for 241 students in the 10 classes sented in Appendix A in the online sup- with at least two boys and two girls from plemental material. a Turkish or Moroccan background can To generate scales of prejudice, I be found in Appendix C in the online sup- reversed the directions of the four items plemental material. per ethnic group so that higher values indicated more negative outgroup atti- tudes. The items showed high internal Procedure consistencies for attitudes toward the Parents received an information letter Dutch, with Cronbach’s alphas of .95 or that offered them the opportunity to higher. Exploratory factor analysis refuse their child’s participation in the revealed that the eight items for atti- study. Students were informed that their tudes toward Turkish and Moroccan peo- answers would be treated confidentially ple loaded on one factor. A combined and that they were free to discontinue scale showed high internal consistencies participation at any time. Per school (alpha = .96 at all waves). Because of class, all students simultaneously com- this and because people tend to general- pleted the questionnaire online on sepa- ize from their attitudes toward one rate computers. A teacher read instruc- minority group to another (Schmid et tions to the students and supervised al. 2012), I analyzed majority group stu- completion of the questionnaires, which dents’ attitudes toward both minority took 30 minutes on average. groups together. Accordingly, the final measures of outgroup prejudice were an Variables additive index of attitudes toward Turks and Moroccans for all Dutch students Outgroup attitudes (prejudice). Stu- and an additive index of attitudes toward dents’ attitudes toward Turkish, Moroc- the Dutch for all non-Dutch students. can, and Dutch people were measured The values of these indexes were using four questions. Participants indi- rounded to the nearest integer because cated on seven-point scales how much this was a prerequisite of the statistical they agreed with the four propositions— model. Values ranged from 1 to 7 with that ‘‘all [ethnic groups] are’’ (a) honest, higher scores indicating more negative (b) friendly, (c) smart, and (d) helpful (1 outgroup attitudes (stronger prejudice). = totally disagree to 7 = totally agree) Descriptive statistics for the additive (Vervoort, Scholte, and Scheepers 2011). indexes can be found in Table 1.

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 135

Ethnicity. Students indicated the coun- .09 1

1.00 tries of birth of their parents. In the first 2 step, a participant was classified as

*** Dutch when both parents were born in

.29 the Netherlands. If at least one parent 1.00 2 was born outside the Netherlands, the student was assigned the ethnicity of that parent. In the second step, I tried .62 *** .58 *** 1.00 2 2 reducing misclassification of students with foreign-born parents who actually identified as being Dutch. All students .02 .04 .04 .10 * .41 *** .38 ***

1.00 were asked to answer the questions, ‘‘Do 2 2 you feel Dutch?’’ and, if they had a foreign-born parent, ‘‘Do you feel [ethnic- .01 .58 *** .39 *** .10 * .37 *** ity of parent]?’’ Answers were given on 1.00 2 2 five-point scales ranging from 1 = not at

Correlations all to 5 = very strongly. If students with 1 a foreign-born parent scored higher on .58 *** .10 .48 *** .33 *** .31 *** 1.00

2 the feeling-Dutch scale than on the other scale, they were recoded as being Dutch. ** ** *** 12 ** Thirty-nine students who scored equally .14 .13 .05 .23 .15 high on both scales could not be excluded 1.00 2 2 2 from the sample because the statistical model can only be applied to complete y networks. I assumed that students who .42 *** .11 * .17 *** .11 * .07 .09 .16 ** .09 1.00 2 2 identified equally strongly with an ethnic minority group as with the majority

123456789 group would also be perceived as a mem- .14 ** .19 *** ber of the minority group by their friends 2 2 from the majority group—at least some of the time. This assumption seems reason- able because 27 of the 39 students did not identify strongly as being Dutch. .55 .15 ** .23 .01 .07 .03 .10 .21 .54 .11 * These students probably did not consider themselves Dutch and were also not per- ceived as Dutch by their classmates. .001(two-tailed tests).

\ Accordingly, those students were

p assigned to the ethnic minority group of

*** their parents.

.01. There were 243 students in the Dutch

\ ethnic majority group, 103 Turkish or p

** Moroccan students, 93 students who belonged to another ethnic minority .05.

\ group, and 14 students who did not indi- p Means, Standard Deviations, and Correlations of All Variables (N = 453)

* cate their ethnicity. The percentage of

.10. Dutch students in each class varied Pairwise deletion of missing values.

\ between 13 and 86 (M = 55.35). Between p 3. Negative outgroup attitudes4. Wave Number 3 of outgroup friends Wave 3.97 1 1.43 .97 .35 *** 1.86 .11 * 5. Number of outgroup friends Wave 2 1.25 2.04 .09 12 1. Negative outgroup attitudes2. Wave Negative 1 outgroup attitudes Wave 2 3.80 1.36 3.96 1.00 1.40 .38 *** Variables Mean SD Table 1. Note: 1 6. Number of outgroup friends Wave 3 1.14 1.79 7. Ethnicity Dutch 8. Ethnicity Turkish/Moroccan 9. Ethnicity other 10. Gender (male) 7 and 58 percent of the students in the

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 136 Social Psychology Quarterly 78(2) classes were of Turkish or Moroccan ori- colleagues developed the stochastic gin (M = 23.55) and between 7 and 52 per- actor-based model (Snijders et al. 2010; cent of students were from other ethnic Steglich et al. 2010). Analysis can be groups (M = 21.20). The ethnic composi- done in the statistical software SIENA tion per class is presented in Appendix B 4.0 in R (Ripley, Snijders, and Preciado in the online supplemental material. 2011). The model can comprise two parts. The attitude function estimates effects of Friendship networks. To assess all friendship relationships within a school lagged covariates on attitude change. class, students were asked to indicate The network function allows for simulta- neous modeling of changes in the friend- whether they considered a particular class- ship network. matea‘‘bestfriend.’’Alistwiththenames of all classmates was displayed in the Attitude function. Several parameters online questionnaire, and the students can be included in the attitude function could check off the names of their best of the model to study the changes in stu- friends. Without limiting the number of dents’ attitudes and to determine on friends, information on the entire friend- what effects these changes may depend. ship network was obtained. On average, The current investigation focused on students nominated 3.91 classmates (17 main effects of lagged covariates, model- percent of all available classmates) as ing the effects that these covariates had best friends in Wave 1, 4.85 (21 percent) on attitude change. Among these, the in Wave 2, and 4.73 (21 percent) in Wave 3. effect of the number of outgroup friends Outgroup friends. The number of stu- represented the effect of intergroup con- dents’ outgroup friends was calculated by tact. A so-called ‘‘average similarity’’ counting the number of friends they nomi- effect tested for social influence among nated who belonged to the ethnic outgroup. friends. This effect is positive and signifi- Friends’ ethnicities were derived from their cant if students adjust their outgroup own answers to the ethnicity questions. attitude at a later wave to the attitudes Dutch students’ numbers of outgroup of their friends at the previous wave. friends were the numbers of Turkish and Network function. Only tie formation Moroccan classmates they nominated as (creation of a new friendship) but not tie best friends. For non-Dutch students, the maintenance was modeled in the analyses number of outgroup friends related to the because the theoretical models all reason number of Dutch classmates nominated. about forming new friendships with out- Control variables. Throughout all anal- group members. This was done with SIE- 3 yses, I controlled for students’ gender (1 NA’s so-called ‘‘creation function.’’ = boy,0=girl). A dummy variable of Both network processes and actor time (0 for the period between Wave 1 characteristics can be included in the and Wave 2 and 1 for the second period) model to simultaneously estimate their was included. This accounted for the effects on the creation of ties. In the expectation that more friendships would initial analyses, no network effects were be created at the beginning of the school included in order to estimate direct effects year than between the later two waves. 3The effects presented in this study replicated when tie formation and tie maintenance were Analytic Strategy modeled simultaneously (which is the default in SIENA), but they were not if only tie mainte- To study the coevolution of social net- nance was modeled with the so-called ‘‘endow- works and covariates, Snijders and ment function.’’

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 137 of outgroup attitudes without taking the The stochastic actor-based model potential mediation of such processes requires data on complete social net- into account. I included only an outdegree works. This means that all students in effect, which represents the log-odds for the classes were represented in the net- creating a tie and models the overall den- works. Data for students who entered sity of the network. In the next steps of the class in a later wave or left the class the analysis, I included the two network before the end of the school year were processes, reciprocity and triadic closure, treated as structurally missing. Missing that may potentially mediate the effect values for individual attributes and net- of prejudice on friendship selection. Reci- work ties for students who did not answer procity corresponds to a dummy variable although they were part of the classes in that indicates for each potential friend a given wave were imputed and treated whether he or she already sent out as noninformative in the estimation pro- a friendship nomination. Triadic closure cess (Huisman and Steglich 2008). To indicates the number of friends that are estimate effects in 20 classes, I applied already shared with each potential friend. the multigroup option of SIENA (Ripley Mathematical formulas for all effects are et al. 2011). This estimated the effects in given in Ripley et al. (2011). each class and combined the results Effects of actor characteristics such as under the assumption that the parameter students’ gender or outgroup attitudes values were the same. can be added to the network function to SIENA automatically centers effects account for the fact that actors with cer- for actor characteristics, but not network tain characteristics are more likely than effects, before they are included in the others to select friends, are more likely analyses (Ripley et al. 2011). This makes to be selected as friends, or are more it difficult to compare estimated coeffi- likely to choose each other. For instance, cients, which are the log-odds for forming gender differences in friendship selection a friendship or adopting more negative that were found repeatedly in school net- outgroup attitudes. I present odds ratios works (e.g., Schaefer et al. 2011; Vermeij (OR) wherever these seem to add insight et al. 2009) were modeled with the follow- beyond the direction and significance ing effects: boy ego (are boys more likely level of a parameter. to nominate friends than girls?), boy alter (are boys more likely than girls to be nom- RESULTS inated as friends?), and same gender (are students more likely to nominate same- Tests for selective attrition showed very sex classmates as friends than classmates few differences between respondents in of the other sex?). the twenty classes included in the sample Interactions between such actor char- and those in the four classes that were acteristics were used to test for a selection excluded. There were more interethnic bias. A three-way interaction between friendships in excluded classes at Wave the variables Dutch ego, negative atti- 2, t(473) = 2.34, p = .02, and the outgroup tudes ego,andTurkish/Moroccan alter attitudes were significantly more positive modeled the lower likelihood for Dutch in the same wave, t(485) = 2.02, p = .04. students (Dutch ego) with more negative All other characteristics were not statisti- outgroup attitudes (negative attitudes cally different. ego) to select Turkish or Moroccan Table 1 presents the means, standard friends (Turkish/Moroccan alter) in the deviations, and correlations of all varia- next wave. bles in the study. Students’ negative

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 138 Social Psychology Quarterly 78(2) outgroup attitudes were highly correlated (not shown), indicating that the insignifi- between all waves, as were the numbers cant intergroup contact effect was the of outgroup friends. Dutch students had same for all ethnic groups. There was, significantly more negative attitudes however, a statistically significant indica- toward the ethnic outgroup of Turkish/ tion of social influence. Students adjusted Moroccan people (M = 3.98) than had stu- their attitudes toward the ethnic outgroup dents from other ethnic groups toward in a later wave to the attitudes of their the Dutch (M = 3.57) at Wave 1, t(402) = friends (est = 5.24, SE = .89, p \ .001). 2.34, p = .002. This attitude difference The predicted change in the friendship was no longer significant at Wave 2, networks over time is shown in the lower t(404) = 1.49, p = .14, or Wave 3, t(387) = part of Model 1 in Table 2. The outdegree 1.86, p = .06. effect represents the overall tendency to form ties. The negative coefficient indi- cates that the networks were sparse Intergroup Contact and Ethnic with students selecting far less than half Homophily of their classmates as friends (est = Results of the stochastic actor-based anal- 22.53, SE = .10, p \ .001). The time ysis predicting attitude change and dummy variable’s positive effect (est = change in the friendship network over .31, SE = .09, p = .001) indicates that stu- time are presented in Model 1 of Table dents were forming ties on a more explor- 2. This model excluded network processes atory basis between Wave 1 and Wave 2. and thus closely resembles a classical Later in the school year, when students cross-lagged approach, in which depen- knew their classmates better, friendship dencies between friendships in a network selection was more strongly driven by are ignored. The first seven parameters in the factors represented in the model the table present results for the attitude such as ethnicity. Boys were significantly function in which change of negative out- more likely than girls to select friends group attitudes between the three consec- over time (boys ego: est = .42, SE = .10, utive waves is modeled. The marginally p \ .001), but they were significantly significant linear shape parameter shows less likely to be chosen as friends (boys that students’ attitudes on average devel- alter: est = 2.36, SE = .07, p \ .001). oped toward higher values. The signifi- Not surprisingly, there was a strong ten- cant quadratic shape effect indicates dency to select friends of the same gender a positive feedback effect (estimate = .06, (est = 1.63, SE = .06, p \ .001). There was SE = .03, p = .04). Students with very neg- also evidence that students with negative ative or very positive attitudes developed outgroup attitudes formed more friend- even more extreme attitudes. ships than students with less negative There was no support for the contact outgroup attitudes (negative attitudes hypothesis in the subsample under study. ego: est = .23, SE = .05, p \ .001). Having more outgroup friends was not Results for the ethnicity variables indi- significantly related to the development cated that Dutch and Turkish/Moroccan of less negative attitudes toward the out- students tended to avoid interethnic group (est = 2.02, SE = .02, p = .24). More- friendships. Dutch students tended to over, none of the control variables were nominate more friends from their own significant predictors of attitude change. ethnic group than they should have by There were also no significant interac- chance alone (same Dutch: est = .41, SE tions between the ethnicity variables = .07, p \ .001). Also, students from the and students’ number of outgroup friends Turkish or Moroccan minority group had

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Table 2. Selection Bias and Underlying Mechanisms: Stochastic Actor-Based Multigroup Analyses of the Co-Evolution of Twenty Classroom Friendship Networks and Students’ Outgroup Attitudes (N = 453)

Model 1 Model 2 Model 3 Model 4 Est SE Est SE Est SE Est SE

Downloaded from Attitude function Data tendency: linear shape .061 .03 .061 .03 .051 .03 .061 .03 Data tendency: quadratic shape .06* .03 .06* .03 .051 .03 .051 .03 Number of outgroup friends 2.02 .02 2.02 .02 2.02 .02 2.02 .02 spq.sagepub.com Gender (boys) .01 .07 .01 .06 .02 .06 .02 .06 Ethnicity Dutcha .07 .09 .07 .09 .07 .09 .07 .10 Ethnicity Turk/Moroccana .16 .10 .16 .10 .16 .10 .16 .10 Social influence (average similarity) 5.24*** .89 5.24*** .89 5.11*** .84 5.35*** .93 atASA-American SociologicalAssociation onJune4,2015 Network function Network effects Outdegree 22.53*** .10 22.55*** .10 22.81*** .11 23.58*** .12 Reciprocity — — 1.72*** .11 1.00*** .12 Triadic closure — — — .38*** .02 Time: first period .31*** .09 .32*** .09 .33*** .08 .53*** .08 Gender Boys ego .42*** .10 .40*** .10 .47*** .10 .31** .09 Boys alter 2.36*** .07 2.36*** .06 2.39*** .07 2.41*** .07 Same gender 1.63*** .06 1.63*** .06 1.28*** .10 .96*** .07 Negative outgroup attitude Negative attitude ego .23*** .05 .17** .06 .17** .06 .13** .05 Negative attitude alter 2.002 .03 2.004 .03 2.004 .03 2.0004 .04 Attitude similarity .09 .30 .12 .30 .16 .32 .30 .33 Ethnicity Dutch Dutch ego 2.64*** .13 2.59*** .14 2.62*** .14 2.43*** .12 Dutch alter .161 .09 .20* .10 .30** .10 .30** .10 Same Dutch .41*** .07 .48*** .09 .43*** .09 .40*** .10

139 (continued) 140 Downloaded from

Table 2. (continued)

spq.sagepub.com Model 1 Model 2 Model 3 Model 4 Est SE Est SE Est SE Est SE

Ethnicity Turks/Moroccans atASA-American SociologicalAssociation onJune4,2015 Turks/Moroccans ego Reference Reference Reference Reference Turks/Moroccans alter .191 .11 .261 .14 .23 .15 .18 .15 Same Turks/Moroccans .53*** .08 .60*** .11 .52*** .12 .47*** .12 Ethnicity other Ethnicity other ego 2.301 .16 2.24 .19 2.24 .19 2.15 .16 Ethnicity other alter Reference Reference Reference Reference Same other ethnicity 2.36*** .09 2.43*** .11 2.37*** .11 2.35*** .09 Selection bias of Dutch Dutch ego 3 negative attitude ego — 2.48*** .12 2.47*** .11 2.35*** .09 Dutch ego 3 Turks/Moroccans alter — .27 .26 .35 .29 .33 .29 Negative attitude ego 3 Turks/Moroccans alter — 2.20* .09 2.191 .10 2.201 .11 Dutch ego 3 negative attitude ego 3 Turks/ Moroccans alter — 2.41* .17 2.36* .18 2.23 .21

Note: Unstandardized effects. aReference category is ethnicity other. 1p \ .10.*p \ .05. **p \ .01. ***p \ .001(two-tailed tests). Selection Bias in Social Networks 141 a preference for friends from their own fewer intergroup friendships (path c in ethnic group (same Turks/Moroccans: est Figure 2). = .53, SE = .08, p \ .001). This was differ- ent for students who belonged to any of the other ethnic minority groups. Taking Network Processes the opportunity structure into account, There was no evidence that the network these students were less likely to form process reciprocity could underlie the friendships with students who were also selection bias. Reciprocity, which was from other minority groups than they added in Model 3 of Table 2, was a strong should have by chance alone (same other and highly significant predictor of friend- ethnicity: est = 2.36, SE = .09, p \ .001). ship formation (est = 1.72, SE = .11, p \ .001). Thus, students had a strong ten- dency to reciprocate friendship nomina- Selection Bias tions at a later wave. Adding reciprocity Just like in earlier research, there was to the model reduced the coefficient of evidence for a selection bias according to the three-way interaction, indicating the which negative outgroup attitudes pre- selection bias slightly, but it remained vented the formation of intergroup friend- significant (est = 2.36, SE = .18, p = ships among the Dutch majority group. .049). This refutes Hypothesis 3. This was tested with a three-way interac- In contrast, there was no indication of a selection bias once triadic closure was tion that was added in Model 2 along with taken into account (Model 4 of Table 2). the conditional two-way interactions. The positive and highly significant triadic This three-way interaction was negative closure effect indicates that students had and significant (est = 2.41, SE = .17, p = a strong tendency to nominate the friends .02), indicating that Dutch students of their friends’ friends (est = .38, SE = (Dutch ego) with more negative attitudes .02, p \ .001). Including this network pro- (negative attitudes ego) were less likely cess in the model cut the coefficient for to select Turkish or Moroccan classmates the three-way interaction indicating the (Turkish/Moroccan alter) as friends than selection bias almost in half and rendered Dutch students with more positive atti- it insignificant (est = 2.23, SE = .21, p = tudes. The odds for Dutch students with .27). This supports Hypothesis 5, suggest- unfavorable outgroup attitudes (M 1 1 ing that negative outgroup attitudes do SD) to form friendships with arbitrary not directly cause a rejection of outgroup Turkish or Moroccan classmates was friends. Triadic closure seems to play only .03.4 In contrast, the odds for Dutch a role in it, but the exact underlying students with favorable outgroup atti- mechanism was not yet revealed by these tudes (M 2 1 SD) to do so was .15, which analyses. means that unfavorable outgroup atti- tudes made intergroup friendships more than 5 times less likely (OR = 5.15). This Mediation through Network supports Hypothesis 2, indicating a direct Processes effect from negative outgroup attitudes to To test the mediation models, I followed as closely as possible the method proposed 4 To calculate predicted values, each variable by Baron and Kenny (1986). This method has to take some value. Odds ratio calculations are based on a Dutch boy selecting a Turkish or has been criticized, not because it might Moroccan boy with average outgroup attitudes detect mediation processes that are not between Waves 1 and 2. real but because it lacks statistical power

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 142 Social Psychology Quarterly 78(2) and may overlook existing mediation pro- where their friends tended to have fewer cesses (Zhao, Lunch, and Chen 2010). outgroup friends. A three-way interaction Unfortunately, a bootstrap test for indi- of Dutch ego 3 negative attitude ego 3 rect effects (Preacher and Hayes 2004) Turkish/Moroccan alter at distance 2 does not exist for social network data. was negative and statistically significant The first mediation model proposes (est = 21.02, SE = .51, p = .047, Model 2 that negative outgroup attitudes lead to of Table 3). The last variable indicates fewer friendship invitations from out- whether students are more likely to nom- group members (path a in Figure 2, I). inate a friend, the more Turkish or Moroc- Accordingly, I tested whether negative can friends this potential friend already outgroup attitudes reduced the likelihood has. This model does not include the basic of Dutch students to be nominated as network effects reciprocity and triadic clo- a friend by a Turkish or Moroccan class- sure. As such, the coefficient of the three- mate. This translated into an interaction way interaction cannot be interpreted as between negative attitudes alter, Dutch a behavioral tendency to avoid classmates alter, and Turkish/Moroccan ego.5 with outgroup friends. Rather, it shows Outgroup attitudes did not affect a descriptive pattern in the data. The Dutch majority group students’ likelihood more prejudiced Dutch students were, of being nominated as a friend by Turkish the less likely they were to have friends or Moroccan minority group classmates. who already had Turkish or Moroccan Model 1 in Table 3 shows a negative but friends. This is in line with Hypothesis insignificant coefficient for this three- 6a. To illustrate the strength of this rela- way interaction (est = 2.10, SE = .08, tionship, I calculated predicted values for p = .23). This leads to a rejection of the likelihood of a hypothetical Dutch stu- Hypothesis 4a and is in line with the find- dent to be friends with another Dutch stu- ing that reciprocity does not underlie the dent who had the same number of Dutch selection bias. Reciprocating friendship and of Turkish or Moroccan friends. If invitations was not related to the forma- the hypothetical student had positive atti- tion of interethnic friendships because tudes (M 2 1 SD), he was twice as likely compared to less prejudiced students, to be a friend of the other Dutch student more prejudiced students were not less than when he had negative attitudes (M likely to receive friendship invitations 1 1 SD, OR = 2.13). from outgroup members. Thus, the medi- The data did not clearly indicate that ation model building on reciprocity (as more prejudiced students deliberately proposed in Hypothesis 4b) does not selected themselves into network posi- underlie the selection bias. tions where their friends had fewer out- In line with the second mediation group friends (path a in Figure 2, II). model (Figure 2, II), more prejudiced When the basic network processes reci- Dutch students were in network positions procity and transitivity were added to the model, the three-way interaction 5The stochastic actor-based model does not between Dutch ego 3 negative attitude allow three-way interactions that include two ego 3 Turkish/Moroccan alter at distance alter variables. I solved this problem by first com- 2 was only marginally significant (est = puting a new variable by multiplying negative 2.83, SE = .49, p = .094, Model 3 of Table attitude and Dutch. The three-way interaction 3). This suggests that there was a ten- presented in Model 1 of Table 3 is thus actually a two-way interaction between Turkish/Moroccan dency to avoid friends with outgroup ego and the alter effect of this newly generated friends. However, the descriptive pattern variable. in the data was also partially due to the

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Table 3. Test of the Mediation Models: Stochastic Actor-Based Multigroup Analyses of the Co-Evolution of Twenty Classroom Friendship Networks and Students’ Outgoup Attitudes (N = 453)

Model 1 Model 2 Model 3 Model 4

Downloaded from Est SE Est SE Est SE Est SE

Attitude function Included Included Included Included Network function Network effects spq.sagepub.com Outdegree 22.47*** .10 22.48*** .10 23.53*** .11 23.58*** .12 Reciprocity — — .99*** .13 .98*** .13 Triadic closure — — .38*** .02 .38*** .02 Time and gender Included Included Included Included atASA-American SociologicalAssociation onJune4,2015 Negative outgroup attitude Negative attitude ego .23*** .06 .13* .06 .101 .05 .101 .05 Negative attitude alter .005 .04 2.005 .03 .0003 .04 2.003 .04 Attitude similarity .09 .30 .13 .31 .30 .32 .31 .32 Ethnicity Dutch Dutch ego 2.41** .14 2.53*** .14 2.39** .12 2.45** .13 Dutch alter .11 .16 .14 .10 .26* .11 .31** .11 Same Dutch .27** .10 .23** .08 .29** .09 .37*** .10 Ethnicity Turks/Moroccans Turks/Moroccans ego .13 .17 Reference Reference Reference Turks/Moroccans alter Reference .07 .12 .04 .13 .17 .16 Same Turks/Moroccans .40*** .10 .45*** .08 .36*** .10 .47*** .12 Ethnicity other Ethnicity other ego Reference 2.08 .18 2.05 .16 2.14 .17 Ethnicity other alter 2.07 .12 Reference Reference Reference Same other ethnicity 2.211 .11 2.30** .09 2.27** .10 2.36** .12 (continued) 143 144

Downloaded from Table 3. (continued)

Model 1 Model 2 Model 3 Model 4 spq.sagepub.com Est SE Est SE Est SE Est SE

Test of reciprocity as mediator Dutch alter 3 negative attitude alter 2.02 .04 — — —

atASA-American SociologicalAssociation onJune4,2015 Dutch alter 3 Turkish/Moroccan ego 2.17 .39 — — — Turkish/Moroccan ego 3 negative attitude alter .02 .08 — — — Dutch alter 3 Turkish/ Moroccan ego 3 negative attitude alter 2.10 .08 — — — Test of triadic closure as mediator Turkish/Moroccan alter at distance 2 — 2.13 .35 .01 .32 2.01 .32 Dutch ego 3 negative attitude ego — 2.51*** .12 2.40*** .10 2.40*** .10 Dutch ego 3 Turkish/Moroccan alter at distance 2 — 22.11*** .52 2.36 .51 2.46 .51 Turkish/Moroccan alter at distance 2 3 negative attitude ego — 2.36 .27 .521 .27 2.45 .32 Dutch ego 3 negative attitude ego 3 Turkish/Moroccan alter at distance 2 — 21.02* .51 2.831 .49 2.74 .61 Selection bias of Dutch Dutch ego 3 Turks/Moroccans alter — — — .39 .30 Negative attitude ego 3 Turks/Moroccans alter — — — 2.11 .13 Dutch ego 3 negative attitude ego 3 Turks/ Moroccans alter — — — 2.10 .24

Note: Unstandardized effects. The complete table can be found in Appendix E in the online supplemental material. 1p \ .10.*p \ .05. **p \ .01. ***p \ .001(two-tailed tests). Selection Bias in Social Networks 145 basic network processes. Triadic closure than less prejudiced students to have (becoming friends with the friends of friends who already had friends from eth- a friend) most likely reinforced students’ nic minority groups. This was partially preference for ingroup friends (ethnic due to more prejudiced students’ prefer- homophily) and more prejudiced students’ ence to not befriend classmates with out- slight tendency to avoid friends who group friends and partially due to their already had Turkish or Moroccan friends. tendency to become friends with their Thus, there was only partial support for friends’ friends (triadic closure). The Hypothesis 6b. direct effect of prejudice on fewer inter- The test of the full mediation model group friendships (the selection bias) dis- remained inconclusive. Following the appeared once students’ preference for Baron and Kenny (1986) method, I tested triadic closure was included in the mod- how the coefficient of path c in Figure 2, II els. This suggests that more prejudiced changed once the process underlying the students did not actively avoid outgroup potential mediators were included in the members. Rather, they became friends model. Model 4 in Table 3 thus includes with their friends’ friends who happened the three-way interaction representing to be from the ingroup, not the outgroup. the selection bias, the three-way interac- There was no evidence for an alterna- tion representing the mediator, and tive model, which suggested that reciproc- triadic closure as the underlying mecha- ity mediates the selection bias. More prej- nism. Triadic closure remained signifi- udiced individuals might be rejected by cant in this model (est = .38, SE = .02, outgroup members and may thus not p \ .001) whereas the coefficient of receive friendship invitations from the three-way interaction for the selection outgroup. However, the empirical results bias was close to zero (est = 2.10, SE = did not support this model. Students .24, p = .68). This would be in line with from the ethnic minority groups were complete mediation according to Hypoth- not less likely to select more prejudiced esis 7. However, the coefficient of the majority group students as friends. Reci- mediator also was insignificant in this procity could thus not underlie the selec- model (est = 2.74, SE = .61, p = .22). tion bias. This was most likely due to high multicol- linearity between the estimates given Mediation Instead of a Direct Effect? seven interaction effects that were par- tially based on the same main effects. The findings indicate that outgroup atti- Based on these findings, it cannot be con- tudes play a role in the development of cluded with certainty that the proposed intergroup friendships, but the underly- mediation model in Figure 2, II is the ulti- ing process looks different from what is mate cause of the selection bias. commonly assumed. Earlier longitudinal studies that found evidence for the selec- tion bias explained it as a psychological DISCUSSION preference of more prejudiced individuals The present study found that a seemingly to avoid outgroup members and of less direct effect of majority group students’ prejudiced people to engage in intergroup negative outgroup attitudes on the avoid- contact (e.g., Binder et al. 2009; Levin et ance of outgroup friends (the selection al. 2003; Sidanius et al. 2008). However, bias) was actually caused by the network intergroup friendships do not develop process triadic closure. More prejudiced because students with positive attitudes majority group students were less likely seek out friends from other groups. It is

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 146 Social Psychology Quarterly 78(2) not so much individual preferences but remained inconclusive. As expected, the rather network processes that are respon- coefficient of the selection bias was insig- sible for the link between attitudes and nificant and close to zero when triadic clo- intergroup contact. Students with posi- sure and students’ preference for avoiding tive attitudes are more likely to be in net- friends who already had outgroup friends work positions where they have friends were added to the model. However, this who already have outgroup friends. Sub- latter coefficient was also insignificant sequently, new and old friends are intro- in the full model, most likely due to high duced to each other and intergroup multicollinearity. friendships form as a consequence of tri- adic closure. In contrast, more prejudiced Extended and Direct Contact students’ friends do not form bridges between the ethnic ingroup and the out- The results of this study may seem partic- group and therefore cannot introduce ularly concerning in light of the extended more prejudiced individuals to outgroup contact hypothesis (Wright et al. 1997). members. Instead, triadic closure only Research has found that merely knowing leads to more ingroup friends for more that ingroup friends have positive con- prejudiced students. tact with outgroup members reduces More prejudiced students ended up in prejudice (see Dovidio et al. 2011). The network positions with fewer friends present study suggests that more preju- who already had contact with the out- diced individuals do not experience group partially due to network processes extended contact because they have (triadic closure) and partially due to their a tendency to avoid friends who have out- own decision. This tendency of more prej- group friends. udiced students to avoid others who However, this study focused on small already have outgroup friends could be social settings in which existing friend- a consequence of the homophily principle ships between a potential friend and out- (McPherson et al. 2001). Students in group members could be easily observed. schools often befriend those who have More prejudiced students could infer similar attitudes or opinions on what are that classmates with outgroup friends considered important issues (Stark and must have positive outgroup attitudes Flache 2012). Outgroup attitudes may be and reject them because of their dissimi- such an important issue for those who lar attitudes. These processes might look have negative attitudes. More prejudiced different if extended contact does not students may infer from existing inter- take place within a small social setting. group friendships that those with out- New friendships may be formed within group friends must have positive out- one social setting and people may be group attitudes and thus avoid them as unaware of the outgroup friends that friends. Alternatively, students may not their new friends already have from select friends who already have outgroup another setting. Once they learn about friends because they want to avoid cogni- these outgroup friends they are faced tive dissonance; they would like their new with cognitive dissonance: they like their friends but disagree on their attitudes new friend but disagree about their atti- toward the outgroup (see balance theory, tudes toward the outgroup. Balance the- Heider 1946, 1958). ory (Heider 1946, 1958) states that people Even though there was evidence for can solve this dissonance by either dis- each step in the proposed mediation solving their friendship or reassessing model, a test of the complete model their attitudes toward the outgroup. The

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 147 decision to remain friends and reduce The present study only focused on stu- prejudice may be one of the mechanisms dents’ friends among their classmates behind the extended contact effect (Mun- although friendship networks outside of niksma et al. 2013). the classroom may be just as important. There was no support for the tradi- For example, minority members with neg- tional contact hypothesis. Having more ative attitudes toward the majority group friendships with outgroup members did may retreat into social networks of their not reduce prejudice. This was most likely own ethnic or religious community out- because of too little statistical power in side of school (Maliepaard and Phalet the small sample under study. Contact 2012). As a result, these individuals may had the expected effect in additional anal- not form ties with majority group mem- yses with larger samples in which I tested bers or with others who already have for the selection bias among the minority friends from the majority within the inte- group (Appendix D in the online supple- grated context of a school class. mental material). CONCLUSIONS Limitations Psychologists are increasingly interested Social network analyses are rarely repre- in social network processes and argue sentative for an entire society (Wasser- for incorporating these processes in the- man and Faust 1994), and the present ory formation and empirical research study relied on data from schools in one (see Westaby, Pfaff, and Redding 2014; city in the Netherlands. Moreover, due Wo¨lfer, Faber, and Hewstone 2015). The to convergence problems, 4 of 24 classes present paper suggests that this is partic- could not be considered in the analyses. ularly beneficial for research on inter- It is thus unclear how well these group contact. Intergroup friendships do results generalize to other settings. Future not only depend on one individual’s atti- research should test whether in samples tudes and preferences but concern differ- with older or more prejudiced respondents ent individuals. Moreover, friendships outgroup attitudes more strongly affect often develop within social contexts that people’s willingness to have intergroup involve other people (Pettigrew 2008; Pet- contact than in this student sample. tigrew et al. 2007). A social network per- The test of the reciprocity model spective offers a new theoretical view on revealed that minority group students the processes underlying intergroup con- were not less likely to choose Dutch tact. For instance, research has found majority peers with negative outgroup that contact effects are mediated by per- attitudes as friends than they were to ceived ingroup norms (Dovidio et al. select those with positive attitudes. This 2011). Social network research, in con- finding may be restricted to settings in trast, has demonstrated that outgroup which negative attitudes do not reflect attitudes are influenced by direct network very strong prejudice. It could also be that contacts (Van Zalk et al. 2013). A combi- students find it difficult to identify their nation of both approaches may further classmates’ actual attitudes. Minority our understanding of which outgroup or group students may not have chosen more ingroup members exert most influence prejudiced friends if they had been aware on people’s outgroup attitudes. The net- of these attitudes. Future research that work perspective may also lead to new assesses perceived attitudes of potential hypotheses about the type of network con- friends should test these two alternatives. figurations (e.g., high or low clustering) in

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 148 Social Psychology Quarterly 78(2) which intergroup contact is less or more and the European Commission (FP7-PEOPLE- consequential for outgroup attitudes. In 2011-IOF, Grant Agreement Number 299939). fact, a social network perspective allows us to systematically consider the level of REFERENCES social interactions and thereby to go Allport, Gordon W. 1954. The Nature of Preju- beyond the more individualistic approach dice. Cambridge: Addison-Wesley. that characterizes most of the existing Baron, Reuben M. and David A. Kenny. 1986. work on the contact hypothesis. ‘‘The Moderator-Mediator Variable Distinc- Next to the theoretical relevance, this tion in Social Psychological Research: Con- research has implications for practi- ceptual, Strategic and Statistical Consider- ations.’’ Journal of Personality and Social tioners who want to promote intergroup Psychology 51:1173–82. friendships among more prejudiced Binder, Jens, Hanna Zagefka, Rupert Brown, majority group members. Earlier findings Friederich Funke, Thomas Kessler, Amelie of a selection bias suggested that more Mummendey, Annemie Maquil, Stephanie prejudiced people’s attitudes have to be Demoulin, and Jacques-Philippe Leyens. 2009. ‘‘Does Contact Reduce Prejudice or improved before contact or even friend- Does Prejudice Reduce Contact? A Longitu- ships with minority group members can dinal Test of the Contact Hypothesis among be established. I argued that this leads Majority and Minority Groups in Three to a vicious circle as contact proved to be European Countries.’’ Journal of Personal- ity and Social Psychology 96:843–56. one of the most effective means to reduce Brown, Rupert, Anja Eller, Sarah Leeds, and prejudice (Hodson 2011; Paluck and Kim Stace. 2007. ‘‘Intergroup Contact and Green 2009). The results of the present Intergroup Attitudes: A Longitudinal Study.’’ research indicate that there may not be European Journal of Social Psychology such a vicious circle after all. More preju- 37:692–703. Brown, Rupert and Miles Hewstone. 2005. ‘‘An diced majority group members tend to be Integrative Theory of Intergroup Contact.’’ in network positions where they are less Advances in Experimental Social Psychol- likely to meet minority group members ogy 37:255–343. and social network processes translate Cartwright, Dorwin and . 1956. these positions into fewer intergroup ‘‘Structural Balance—a Generalization of Heider Theory.’’ Psychological Review 63: friendships. Intervention programs may 277–93. be able to bypass these network processes Christ, Oliver and Ulrich Wagner. 2013. through directly establishing contact ‘‘Methodological Issues in the Study of between members of different groups. Intergroup Contact: Towards a New Wave of Research.’’ Pp. 233–61 in Advances in Intergroup Contact, edited by G. Hodson ACKNOWLEDGMENTS and M. Hewstone. London: Psychology Press. The author thanks Maykel Verkuyten, Jon Kros- Davies, Kristin, Linda R. Tropp, Arthur Aron, nick, Andreas Flache, and Miles Hewstone for Thomas F. Pettigrew, and Stephen C. their helpful comments on earlier versions of Wright. 2011. ‘‘Cross-Group Friendships this manuscript. and Intergroup Attitudes: A Meta-Analytic Review.’’ Personality and Social Psychology FUNDING Review 15:332–51. Davis, James A. 1963. ‘‘Structural Balance, The author(s) disclosed receipt of the following Mechanical Solidarity, and Interpersonal financial support for the research, authorship, Relations.’’ American Journal of and/or publication of this article: This work was 68:444–62. supported by the Arnhem Housing Corporations Dovidio, John F., Anja Eller, and Miles Hew- (GOW), the Institute for Integration and Social stone. 2011. ‘‘Improving Intergroup Rela- Efficacy (ISW) at the University of Groningen, tions through Direct, Extended and Other

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 Selection Bias in Social Networks 149

Forms of Indirect Contact.’’ Group Pro- Paluck, Elizabeth L. and Donald P. Green. cesses & Intergroup Relations 14:147–60. 2009. ‘‘Prejudice Reduction: What Works? Eller, Anja and Dominic Abrams. 2003. ‘‘‘Grin- A Review and Assessment of Research and gos’ in Mexico: Cross-Sectional and Longi- Practice.’’ Annual Review of Psychology tudinal Effects of Language School-Pro- 60:339–67. moted Contact on Intergroup Bias.’’ Group Pettigrew, Thomas F. 2008. ‘‘Future Directions Processes & Intergroup Relations 6:55–75. for Intergroup Contact Theory and Feddes, Allard R., Peter Noack, and Adam Research.’’ International Journal of Inter- Rutland. 2009. ‘‘Direct and Extended cultural Relations 32:187–99. Friendship Effects on Minority and Major- Pettigrew, Thomas F., O. Christ, U. Wagner, ity Children’s Interethnic Attitudes: A Lon- and J. Stellmacher. 2007. ‘‘Direct and Indi- gitudinal Study.’’ Child Development rect Intergroup Contact Effects on Preju- 80:377–90. dice: A Normative Interpretation.’’ Interna- Goodreau, Steven M., James A. Kitts, and tional Journal of Intercultural Relations Martina Morris. 2009. ‘‘Birds of a Feather, 31:411–25. or Friend of a Friend? Using Statistical Net- Pettigrew, Thomas F. and Linda R. Tropp. work Analysis to Investigate Adolescent 2006. ‘‘A Meta-Analytic Test of Intergroup Social Networks.’’ Demography 45:103–25. Contact Theory.’’ Journal of Personality Heider, Fritz. 1946. ‘‘Attitudes and Cognitive and Social Psychology 90:751–83. Organization.’’ Journal of Psychology Pettigrew, Thomas F., Linda R. Tropp, Ulrich 21:107–12. Wagner, and Oliver Christ. 2011. ‘‘Recent Heider, Fritz. 1958. The Psychology of Inter- Advances in Intergroup Contact Theory.’’ personal Relations. New York: Wiley. International Journal of Intercultural Rela- Hewstone, Miles and Hermann Swart. 2011. tions 35:271–80. ‘‘Fifty-Odd Years of Inter-Group Contact: Preacher, Kristopher J. and Andrew F. Hayes. From Hypothesis to Integrated Theory.’’ 2004. ‘‘SPSS and SAS Procedures for Esti- British Journal of Social Psychology mating Indirect Effects in Simple Media- 50:374–86. tion Models.’’ Behavior Research Methods, Hodson, Gordon. 2011. ‘‘Do Ideologically Intol- Instruments, and Computers 36:717–31. erant People Benefit from Intergroup Con- Ripley, Ruth M., Tom A. B. Snijders, and Pau- tact?’’ Current Directions in Psychological lina Preciado. 2011. Manual for Siena Ver- Science 20:154–59. sion 4.0 (Version August 27, 2013). Oxford: Huisman, Mark and Christian E.G. Steglich. University of Oxford, Department of Statis- 2008. ‘‘Treatment of Non-Response in Lon- tics; Nuffield College. http://www.stats.ox gitudinal Network Studies.’’ Social Net- .ac.uk/siena/. works 30:297–308. Rutland, Adam, Rupert J. Brown, Lindsey Levin, Shana, Colette van Laar, and Jim Sida- Cameron, Anni Ahmavaara, Katie Arnold, nius. 2003. ‘‘The Effects of Ingroup and and Jenni Samson. 2007. ‘‘Development of Outgroup Friendships on Ethnic Attitudes the Positive-Negative Asymmetry Effect: in College: A Longitudinal Study.’’ Group In-Group Exclusion Norm as a Mediator of Processes & Intergroup Relations 6:76–92. Children’s Evaluations on Negative Attrib- Maliepaard, Mieke and Karen Phalet. 2012. utes.’’ European Journal of Social Psychol- ‘‘Religious Identity Expression among ogy 37:171–90. Dutch Muslims: The Role of Minority and Schaefer, David R., Olga Kornienko, and Majority Group Contact.’’ Social Psychology Andrew M. Fox. 2011. ‘‘Misery Does Not Quarterly 75:131–48. Love Company: Network Selection Mecha- McPherson, Miller, Lynn Smith-Lovin, and nisms and Depression Homophily.’’ Ameri- James M. Cook. 2001. ‘‘Birds of a Feather: can Sociological Review 76:764–85. Homophily in Social Networks.’’ Annual Schmid, Katharina, Miles Hewstone, Beate Review of Sociology 27:415–44. Ku¨pper, Andreas Zick, and Ulrich Wagner. Munniksma, Anke, Tobias H. Stark, Maykel 2012. ‘‘Secondary Transfer Effects of Inter- Verkuyten, Andreas Flache, and D. Rene group Contact: A Cross-National Compari- Veenstra. 2013. ‘‘Extended Intergroup son in Europe.’’ Social Psychology Quar- Friendships within Social Settings: The terly 75:28–51. Moderating Role of Initial Outgroup Atti- Sidanius, Jim, Shana Levin, Colette Van Laar, tudes.’’ Group Processes & Intergroup Rela- and David O. Sears. 2008. The Diversity tions 16:752–70. Challenge: Social Identity and Intergroup

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015 150 Social Psychology Quarterly 78(2)

Relations on the College Campus. New G. Hodson and M. Hewstone. London: Psy- York: Russell Sage Foundation. chology Press. Snijders, Tom A. B., Gerhard G. Van de Bunt, Vorauer, Jacquie D. and Yumiko Sakamoto. and Christian E. G. Steglich. 2010. ‘‘Intro- 2006. ‘‘I Thought We Could Be Friends, duction to Stochastic Actor-Based Models but . . . Systematic Miscommunication and for Network Dynamics.’’ Social Networks Defensive Distancing as Obstacles to 32:44–60. Cross-Group Friendship Formation.’’ Psy- Stark, Tobias H. and Andreas Flache. 2012. chological Science 17:326–31. ‘‘The Double Edge of Common Interest: Wasserman, Stanley and Katherine Faust. Ethnic Segregation as an Unintended 1994. : Methods Byproduct of Opinion Homophily.’’ Sociol- and Applications. Cambridge: University ogy of Education 85:179–99. Press. Steglich, Christian E. G., Tom A. B. Snijders, Westaby, J. D., D. L. Pfaff, and N. Redding. and Michael Pearson. 2010. ‘‘Dynamic Net- 2014. ‘‘Psychology and Social Networks works and Behavior: Separating Selection a Dynamic Network Theory Perspective.’’ from Influence.’’ Sociological Methodology American Psychologist 69:269–84. 40:329–93. Wimmer, Andreas and Kevin Lewis. 2010. Swart, Hermann, Miles Hewstone, Oliver ‘‘Below and Beyond Racial Homophily. Erg Christ, and Alberto Voci. 2011. ‘‘Affective Models of a Friendship Network Docu- Mediators of Intergroup Contact: A Three- mented on Facebook.’’ American Journal Wave Longitudinal Study in South Africa.’’ of Sociology 116:583–642. Journal of Personality and Social Psychol- Wright, Stephen C., Arthur Aron, Tracy ogy 101:1221–38. McLaughlin-Volpe, and Stacy A. Ropp. Van Zalk, Maarten H. W., Margaret Kerr, 1997. ‘‘The Extended Contact Effect: Nejra Van Zalk, and Hakan Stattin. 2013. Knowledge of Cross-Group Friendships ‘‘Xenophobia and Tolerance toward Immi- and Prejudice.’’ Journal of Personality and grants in Adolescence: Cross-Influence Pro- Social Psychology 73:73–90. cesses within Friendships.’’ Journal of Wo¨lfer, Ralf, Nadira S. Faber, and Miles Hew- Abnormal Child Psychology 41:627–39. stone. 2015. ‘‘Social Network Analysis in Vermeij, Lotte, Marijtje A. J. Van Duijn, and the Science of Groups: Cross-Sectional and Chris Baerveldt. 2009. ‘‘Ethnic Segregation Longitudinal Applications for Studying in Context: Social Discrimination among Intra- and Intergroup Behavior.’’ Group Native Dutch Pupils and Their Ethnic Dynamics: Theory, Research, and Practice Minority Classmates.’’ Social Networks 19:45–61. 31:230–39. Zhao, Xinshu, John G. Lunch Jr., and Qimei Vervoort, Miranda H. M., Ron H. J. Scholte, Chen. 2010. ‘‘Reconsidering Baron and and Peer L. H. Scheepers. 2011. ‘‘Ethnic Kenny: Myths and Truths about Mediation Composition of School Classes, Majority- Analysis.’’ Journal of Consumer Research Minority Friendships, and Adolescents’ 37:197–206. Intergroup Attitudes in the Netherlands.’’ Journal of Adolescence 34:257–67. BIO Vezzali, Loris, Dino Giovannini, and Dora Capozza. 2010. ‘‘Longitudinal Effects of Tobias Stark is a Marie Curie Interna- Contact on Intergroup Relations: The Role of Majority and Minority Group Member- tional Outgoing Fellow at the European ship and Intergroup Emotions.’’ Journal of Research Centre on Migration and Ethnic Community & Applied Social Psychology Relations (ERCOMER) at Utrecht Uni- 20:462–79. versity in the Netherlands and he was Vorauer, Jacquie D. 2006. ‘‘An Information a visiting scholar at the Department of Search Model of Evaluative Concerns in Intergroup Interaction.’’ Psychological Communication at Stanford University Review 113:862–86. between 2012 and 2014. His main Vorauer, Jacquie D. 2013. ‘‘Getting Past the research interest is in the measurement Self: Understanding and Removing Evalua- of racial attitudes as well as in the link tive Concerns as an Obstacle to Positive between such attitudes, interethnic rela- Intergroup Contact Effects.’’ Pp. 23–48 in Advances in Intergroup Contact, edited by tions, and social networks.

Downloaded from spq.sagepub.com at ASA - American Sociological Association on June 4, 2015