Gross Religious Group Boundaries
We first present a descriptive picture of gross religious group boundaries in classroom networks, that is, the gross group segregation in positive and negative ties among youth of different religions. Gross group boundaries reflect the pattern we observe when not taking opportunities for intergroup contact into consideration, which make up students’ daily reality. Table 2 displays the percentages of in-class intra- and interreligious social ties per religious group. We compute these percentages based on nationally representative (weighted) data.
Table 2 Intra- and interreligious social ties in class We find, first, that Christian and non-religious youth are mainly friends with other Christian and non-religious classmates or with each other; they have very few Muslim friends. To illustrate, in Germany, 80% of the friends of Christian youth are Christian, 12% are non-religious, and only 5% are Muslim. In England, 59% of the friends of non-religious youth are non-religious, 35% are Christian, and only 2% are Muslim.
A second finding is that the friendship network of Muslims is much more diverse than that of Christian and non-religious youth. Muslim youth tend to befriend Muslims but they also have many friends that are non-religious and Christian. For example, in Sweden, 44% of Muslim youth have friends who are Muslim, 35% who are Christian, and 18% who are non-religious.
Third, we find that these two patterns are mirrored when examining negative ties, although the religious group boundaries appear slightly less strong. Thus, Christian and non-religious youth not only have most friends within their group or with each other, but they also tend to have negative ties most frequently in their group or with each other, rather than with Muslims. To illustrate, in the Netherlands, 62% of the negative in-class ties of non-religious youth are with non-religious peers; only 5% are with Muslims, which is just above the share of positive ties with them (3%). Likewise, we find that the negative ties of Muslims are more diverse than those of Christian and non-religious youth. However, Muslim youth have far fewer negative ties with Muslims than they have friendships with them. For example, in Germany, only 22% of the negative ties of Muslim youth are with other Muslims, while 44% of their friends are Muslim. Such a strong discrepancy between positive and negative ties is not observed among other groups. In Germany, for example, 44% of the friends of non-religious youth are non-religious and the same percentage of their negative ties is with non-religious peers.
The Role of Population Size
A key follow-up question is to what extent these gross religious group boundaries in positive and negative ties are shaped by mere opportunity. We look at two opportunity forces: population size (national) and school class composition (local). Table 2 presents the share of each religious group in the population. As can be seen, non-religious and Christian youth make up the largest population in each country, significantly more than the Muslim population. Consequently, their larger numbers may account for the more numerous intrareligious ties, as well as the frequent ties between youth from these two groups. This element may also account for the few ties Christian and non-religious groups have with Muslims.
However, group size in the population cannot fully account for gross religious group boundaries. A consistent finding is that the number of observed intrareligious ties exceeds the number of expected intrareligious ties based on population numbers. For example, whereas 70% of German youth are Christian, 80% of their friends are Christian. This discrepancy is largest among Muslims. In the Netherlands, only 5% of youth are Muslim, but among Muslims, 36% of their friends are Muslim. In summary, gross religious group boundaries are stronger than expected according to population size.
The Role of School Class Composition
Opportunities for contact are also dependent on the composition of classes. Children from different religions are sorted in school classes, which provide the immediate context for creating social ties and may thereby also account for gross religious boundaries. Table 3 presents results from regression models in which the share of respondents’ intrareligious ties is regressed on the share of intrareligious classmates. We find that this immediate context plays an important role: a one percent increase in co-religious classmates is associated with almost a one percent increase in co-religious positive and negative ties of ego. The percentage of intrareligious classmates accounts for 29–53% of the variance in the share of intrareligious positive and negative ties.
Table 3 Regression models predicting the percentage of intrareligious ties of ego in class In summary, these findings highlight that differences in group size and class composition are important drivers of gross religious boundaries. At the same time, the observed patterns cannot be fully explained by mere opportunity and significant unexplained variance remains.
Net Religious Boundaries
When the sizes of different religious groups in the class are taken into account, we obtain the net religious boundaries of students’ networks, which reveal more about their in- and out-group preferences. One way to capture net group boundaries is using odds ratios, which are insensitive to the distribution of the size of the different religious groups in the class (Moody 2001). To this end, for each religious group, we compute (a) the number of realized ties between members of the same affiliation, (b) the number of ties that are not realized between members of the same affiliation, (c) the number of realized ties between members of different affiliations, and (d) the number of ties that are not realized between members of different affiliations in each school class. The fraction of a/b to c/d corresponds to the odds of a positive (negative) tie between members of a same-religious dyad relative to the odds of a positive (negative) tie in an interreligious dyad. We combine these ratios across school classes in each country by way of a random-effects meta-analysis.
Figure 1 displays the mean odds ratios across classes with the 95% confidence intervals around those means. Results indicate that the odds of an intrareligious positive tie are higher than those of an interreligious positive tie for non-religious, Christian, and Muslim youth in all four countries. However, the tendency to create intrareligious friends is much stronger among Muslim youth than among non-religious and Christian youth. This (partly) explains why Muslim youth have many Muslim friends, despite their fewer numbers. When looking at negative ties, we see that the intrareligious tendencies are less strong for each religious group.
Testing Hypotheses on Religious Homophily
While Fig. 1 provides a good picture of religious group boundaries in youth networks independent of religious group sizes in the class, it does not take into account other ego, dyadic, and class characteristics. In the next step, we consider other characteristics in studying patterns that drive net religious group boundaries. We test the hypotheses using Additive and Multiplicative Random Effects (AME) modelsFootnote 7 (see “Appendix 2” for sensitivity analysis employing exponential random graph models as an alternative method).
In network analyses, one has to account for interdependencies between observations (e.g., structural dependencies of ties in networks) when estimating the effects of covariates. AME models achieve this through a set of additive and multiplicative effects. Specifically, the additive part of the AME model accounts for variation in the number of nominations made and received by ego and alter as well as the reciprocity of the nominations. The multiplicative part of the model captures the most established third-order dependence patterns (e.g., transitive triads and stochastic equivalence) (Minhas et al. 2016a; b).Footnote 8 Third-order dependencies are of particular relevance given that previous research has shown they substantially contribute to tie formation mechanisms; disregarding these dependencies, therefore, may result in the overestimation of homophily coefficients (Wimmer and Lewis 2010).Footnote 9
We follow a two-stage procedure in estimating the homophily coefficients, which is a common method in the analysis of multiple small networks. In the first stage, each class network is analyzed separately; in the second one, the parameter estimates obtained in the first stage are summarized across classes using meta-analysis (Lubbers 2013; Smith et al. 2016; Snijders and Baerveldt 2003).
The dependent variables are ‘positive ties’ and ‘negative ties’; both are measured at the dyadic level. We construct two models for predicting each dependent variable. In models 1a and 1b, we include four dyadic covariates. The first variable is ‘same religious affiliation’, which takes on the value 1 if ego and alter have the same religious affiliation and 0 otherwise. The second variable is “difference in religiosity”, which is constructed by calculating the absolute difference between scores of ego and alter on the religiosity scale. Lastly, we include two more homophily measures, namely, ‘different ethnicity’ and ‘different gender’Footnote 10 as control variables (see Table 7 in “Appendix 1” for a detailed description of variables used in the analysis). Aside from dyadic covariates, the models contain eight nodal covariates: ego’s and alter’s percentage of co-religious classmates; ego’s and alter’s percentage of co-ethnic classmates; ego’s and alter’s percentage of same-gender classmates (to account for relevant meeting opportunities in class); ego’s and alter’s religiosity (as a proxy for meeting opportunities in terms of religiosity and to avoid misspecification of the models).
To study the salience of specific religious boundaries, in Models 2a and 2b (Tables 4, 5) we replace the dummy ‘same religious affiliation’ with six dummies that capture all possible combinations in which ego’s religious affiliation is not the same as alter’s affiliation, comparing them to dyads with the same religious affiliation. Otherwise, the models are identical to models 1a and 1b. In all the models, the same control variables are used.
Table 4 Additive and multiplicative random effects models predicting the log odds of positive ties among classmates (standard errors in parentheses) Table 5 Additive and multiplicative random effects models predicting the log odds of negative ties among classmates (standard errors in parentheses) In the second stage, we run a univariate random effects meta-analysis to pool the effects over the classes.Footnote 11 In a random effects meta-analysis, the weights assigned to separate class findings depend on both within-study variance and estimated between-study variance. These weights ensure that classes with large standard errors contribute less to the average effect size than classes with small standard errors. In addition, these weights take into account that additional variation across study findings might occur due to differences in the way studies are carried out (Veroniki et al. 2015). We present the effects of primary interest (i.e., the dyadic homophily measures) in the log-odds units in Tables 4 and 5. The full models, including the effects of nodal covariates, can be found in “Appendix 1” (see Tables 9, 10).
The results in Table 4, Model 1a, reveal that in Germany, the Netherlands, and Sweden same-religious positive ties are more likely than cross-religious positive ties, even after taking into account ethnic and gender homophily as well as other control variables. These findings lend support to Hypothesis 1, suggesting that classmates with different religious affiliations are less likely to have positive ties with each other than classmates who have the same religious affiliation.
Model 2a in Table 4 shows that compared to the case in which ego and alter have the same religious affiliation, the odds of a positive tie are 34%-54% smaller for a Muslim/Christian dyad and 40%-53% smaller for a Muslim/non-religious dyad in all the countries. Importantly, in none of the countries, the likelihood of a positive tie differs between when ego and alter are of the same affiliation and when ego(alter) is non-religious and alter(ego) is Christian. Additional analysis (Table 8 in “Appendix 1”) further reveals that the odds of a friendship tie between a Muslim and a non-Muslim youth are 34%-37% smaller (p < 0.001) in Germany and Sweden and 26% smaller (p < 0.05) in the Netherlands compared to other ego-alter combinations. In England, the corresponding odds for a friendship tie in a Muslim/non-Muslim dyad are rather similar to the ones in the Netherlands but fall short of statistical significance (p = 0.067). Overall, these findings provide support for Hypothesis 3a, suggesting that Muslims stand out as a group and the boundaries between them and other religious affiliations are the strongest in terms of friendships.
Turning to negative ties, Model 1b in Table 5 shows that the discrepancy in religious affiliation does not relate to classmates’ reluctance to sit together. This is in contrast to Hypothesis 2, which predicted heterophobia among youth based on religious affiliation. Examining all possible combinations of dyads based on religious affiliation (Model 2b in Table 5), we see that the likelihood of a negative tie between a Christian and a Muslim in England is higher compared to that between a pair of religiously co-affiliated youth. Furthermore, separate analyses (Table 8 in “Appendix 1”) show that boundaries between Muslims and non-Muslims are evident in two countries; the odds of a negative tie increase by approximately 28% in the Netherlands (p < 0.05) and 24% in Sweden when ego (alter) is Muslim and alter(ego) is non-Muslim, compared to other possible ego-alter combinations. These findings give only partial support to Hypothesis 3b, suggesting that the boundaries between Muslims and non-Muslims are the strongest in terms of negative ties.