The data analysed in this paper is from the first two waves of the Children of Immigrants Longitudinal Survey in Four European Countries (CILS4EU) data-set, which focuses on the integration of first and second generation youths in four European countries: England, Sweden, Germany, and the Netherlands (Kalter et al. 2014, 2015). This is the first comprehensive dataset that is fully standardised, longitudinal, and generalisable to the survey country populations, providing data on youths, teachers, and parents, as well as information on class-room social networks. Wave 1 of this survey interviewed students at age 14–15, and wave 2 at age 15–16. The survey includes information on minority and majority ethnic groups, oversampling immigrant-dense schools in order to increase power in the analyses that focus on ethnic minority students.
This study uses data from England, Germany, and Sweden. Since the education system in the Netherlands is not strictly classroom-based to the same extent as in the other countries, the measure of social exclusion through the classroom-networks is regarded to be less valid than in the other countries; thus, this country is excluded from the analysis. In total, the sample used here has information on 15,017 students in 731 classrooms, in 381 schools. Mostly, two classrooms per school were included in the sample, with an average of 20.6 students in each classroom.
Descriptives of all variables used are reported in Table 1, in the full data set and in the analytical sample, respectively. Around one third of all respondents had missing information on one or more variables that were included in the analysis, mainly due to missingness on the outcome variable, grade point average (GPA; 3328 of 15,017 cases). When comparing the variable distributions of the original sample to the analytical sample, that is the sample of all complete cases, there are no substantial differences between the two samples, therefore it is assumed that there is no systematic attrition that results from list-wise deletion.
The outcome variable GPA was collected in wave 2, while the predictor variables are taken from wave 1.
Outcome Variable: Grade Point Average (GPA)
The outcome is GPA, measured as the average of the school grades in mathematics and the survey country language. This is available as self-reported grades for England and Germany, and as teacher-assigned grades taken from school-register data for Sweden. This renders the data for Sweden more objective, and presumably of higher quality. The measure for GPA is z-standardised within each country, because countries use different grading scales. For interpretation, this means that a unit change in the outcome variable refers to a change in the distribution of grades, expressed in standard deviations.
Main Predictor 1: Migration Background as Generational Status
Migration background is included in the models as a categorical variable, based on the official generation status variable available in the CILS4EU data-set, constructed according to the “ancestral distance” of the respondents to their immigrating ancestors (Dollmann et al. 2014). The official variable was constructed using information on country of origin and age of arrival in the survey country of the child, his or her parents as well as grandparents. Based on this, the variable takes four values for the purpose of this analysis.
First, “majority” refers to adolescents whose parents were born in the survey country. Second, “newly arrived” captures those that arrived after the age of 10, who have been in the country for only a few years: at the point of time of the survey, youths were mostly between 14 and 15 years old. Third, the “first generation”, i.e. those who migrated themselves, but arrived before the age of 10. Fourth, the “second generation” refers to those who were born in the survey country, but whose parents have been born abroad.
Since the main analysis here pools data from the three countries, differentiating between countries or regions of origin is not feasible, as the survey countries vary substantially in the ethnic composition of their immigrant populations; this would have led to a low number of cases in some categories. Thus, this paper utilises the generation status in the main analysis, and uses the ethnic composition of the immigration population in the country-specific interpretation of the results. The ethnic composition of the generational status categories of the pooled data is presented in Table A2 in the appendix. Additionally, country-specific analyses are included in Tables A3 to A8 in the appendix.
Main Predictor 2: Social Exclusion
In line which what has been discussed in the literature review, social exclusion is conceptualised here both as social isolation and social avoidance, by accounting more closely for the type of isolation, as well as for the migration background of those who avoid the individual. It has been shown that different types of relations between individuals, such as absence of friendship and avoidance are best considered jointly to measure a complex social situation, such as exclusion (Vörös and Snijders 2017).
Isolation from everyone in the class. This dichotomised variable captures social isolation as the complete absence of friendship nominations by classmates. 3.6% of majority students and 3.8% of minority students are isolated from everyone in the class.
Isolation from majority students in the class. This dichotomised variable accounts for individuals not receiving friendship nominations from majority students in the classroom, but from minority students. They are thus not isolated completely, as in (1), but they are isolated from the majority students in the class. 3.5% of majority students and 31.1% of minority students are isolated from majority students in the class, and have no majority friends outside the classroom, either.
Avoidance by majority students in the class. This variable captures the number of incoming avoidance ties from majority students in the classroom, and expresses this as a proportion of all majority students in the classroom. The avoidance tie is based on the survey question “who do you not want to sit next to?”. A value of 0.5 would thus mean that 50% of majority students in the classroom report that they do not want to sit next to the individual.
Avoidance by minority students in the class. Following the same logic as in (3), this variable measures avoidance by minority-students in the classroom.
For measures 3 and 4, avoidance ties from those that the individual considers to be their friend are not counted. Among friends, there can be various reasons of why they do not want to sit together, such as being able to better concentrate on the lessons. Approximately 6% of 37,532 recorded avoidance ties (across all three countries) were dropped based on this reasoning. Additionally, the analyses were carried out without dropping those ties. The results suggest that the measure of social avoidance is more accurately tapping into the dimension of interest here if the ties from friends are dropped (more details are provided in the section on robustness checks).
Figures 1 and 2 show how the measures for avoidance and isolation apply by migration background. These descriptives show two main aspects. First, minority students are a lot more likely to be isolated from majority students than majority students themselves are; this applies in particular to newly arrived immigrants. Second, avoidance ties seem to exist more within groups, i.e. majority and minority, than between groups. Newly arrived minority students are more often than other minority students avoided by majority students. This supports the above-discussed tendency of newly arrived immigrants being particularly vulnerable to social exclusion, especially to social exclusion from majority students.
The analysis furthermore controls for a variety of individual characteristics which are likely to affect grade point average, too. Besides gender and age (in months at the survey date), the models include a measure of cognitive ability. This is the result of a language-free test that is part of the CILS4EU data, and which is based on solving graphical problems in a given time (CILS4EU 2014). This is included in the model to control for baseline differences in ability. Predicting grades controlling for cognitive ability accounts for the individual’s potential to get high grades at school, which is an important, but not perfect predictor.
To account for baseline differences in language ability, which is crucial to include when considering ethnic differences, a dummy of whether another language, besides the survey country language, is spoken at home. This is preferred to a language ability test, which is included in the data as well, because of potential endogeneity in this test variable (those of immigrant origin probably improve their language proficiency if they have majority friends). Additionally, the language test result is highly correlated with the cognitive ability index.
Parental background is controlled for by the highest level of parental educational attainment, which includes four categories: primary school, lower secondary school, higher secondary school, and university degree. A control for parental unemployment is included in the models as well, since past research has shown that youths with fewer economic resources are more likely to be excluded from activities and friendships (Hjalmarsson and Mood 2015). The unemployment variable is a dummy, accounting for whether there is not at least one parent living with the adolescents who is employed.
Resistance to schooling is controlled for, too, as it is an important alternative explanation for low grades. Following Geven et al. (2017), a measure for resistance to schooling is computed based on five survey items that capture the extent of the individual (1) arguing with teachers, (2) getting punishments at school, (3) skipping classes, (4) being late to classes, and (5) putting effort into work for school. These items load on one factor with individual loadings between 0.4 and 0.8 (Crohnbach’s alpha = 0.71). The items were combined additively, with the fifth item being reversed; non-complete cases on all five items were assigned a missing value on the scale.
Furthermore, the models control for number of friends, measured as the in-degree, i.e. the number of people who nominate the individual as a friend. This control is included here since the aim of the analysis is to understand the effect of other types of social exclusion besides the effect of absence of friends. Three of the four measures used in this analysis do not exclude the possibility of having friends. Those who are considered to be isolated from majority students only, have friends who do not belong to the majority group. The two measures of avoidance do not consider friends at all, and only account for avoidance relationships. The effect of isolation from majority students or avoidance net of number of friends can show if it is the exclusion from a certain sub-group of people in the classroom that makes a difference, rather than being completely excluded from everyone.
Official survey weights were used in all models to account for the over-sampling of immigrant-dense classrooms.
Linear regression models with standard errors clustered at classroom level are utilised to predict the grade point average (GPA) from CILS4EU wave 2; predictors are from wave 1, with approximately one year between the two waves. After confirming that there is no country-heterogeneity in the social exclusion effects, which was tested by including country interactions (see Table A2 in the Appendix), analyses were run on pooled data from the three countries. Additionally, all models were run separately for each country (see models A3 to A8 in the Appendix): where country-specific differences are found, they are discussed in the results section. In the first set of models, Model 1 predicts GPA only by migration background. In the next step, the controls described above are added (Model 2), then measures of social exclusion (Model 3). In the second set of models, interaction effects between migration background and all social exclusion variables are added to the full model configuration. This is done with a dummy of migration background (native, i.e. belonging to the majority, vs. first or second generation) and separately for each measure of social exclusion.
All analyses have been carried out in R, version 3.2.2, using the plm package for the regression analysis, and the computation of clustered standard errors. Clustered standard errors are calculated according to Stock and Watson (2008), which is the standard procedure in many statistical packages, such as Stata.