Does the Concentration of Immigrant Pupils Affect the School Performance of Natives?

Does the Concentration of Immigrant Pupils Affect the School Performance of Natives? Inés Hardoy and Pål Schøne Institute for Social Research May 2011 Preliminary please do not quote Abstract In this paper we analyse whether immigrant concentration affects the school performance of natives in secondary school. To reach to causal statements we exploit potential random variation in the number of immigrants between grades within the same school. Performance is measured in terms of the likelihood of dropping out of secondary school. The tentative results reveal a positive and significant correlation between the fraction of immigrants and the dropout rate of natives. The coefficient suggests that a 10 percentage point increase in the immigrant share leads to a 2 percentage point increase in the dropout rate. Regarding the mechanisms, our results point to the importance of peer quality. Our interpretation of this is that it is not an ethnical effect per se; rather it is the skill deficit of immigrants that creates the peer effect. JEL classification: Peer effects, dropout rates, immigrants Keywords: I20, J24 * Address: Inés Hardoy: Institute for Social Research, PB 3233 Elisenberg, N-0208 Oslo, Norway. Tel.: + 47 23 08 61 35. E-mail: iha@samfunnsforskning.no. : Pål Schøne. Institute for Social Research, PB 3233 Elisenberg, N-0208 Oslo, Norway. Tel.: + 47 23 08 61 82. E-mail: psc@samfunnsforskning.no. The authors thank seminar participants at the Institute for Social Research for fruitful comments and suggestions. The work is financed by the Norwegian Research Council, project: The educational system in Norway: Putting to the test of the labour market. The financial support is gratefully acknowledged. 1

1. Introduction During the past three decades the immigrant share of the Norwegian population increased from two to ten per cent. And, in line with trends in other high-income countries, the composition of immigrant inflows changed radically with regard to country of origin (OECD, 2008). Prior to the 1980s, the majority of immigrants came from countries that are geographically and culturally close. Today the majority of the immigrant population comes from countries much more distant in both respects. The increased immigrant share of the population has spurred a large economic research literature analysing labour market impacts of immigration on receiving countries (Card 2001, 2009; Borjas 2003, Ottaviano and Peri 2008). In this paper we report to a more scant but still fast growing literature focusing on the impact of immigration on native pupils school performances. Parallel with the increasing number of immigrants the educational policy has to larger extent been concerned with impact of segregated schools, especially in the primary and secondary level. Many politicians have expressed concern of whether ethnical segregated schools have negative impact on the learning environment for the pupils. In a period were many western countries are expected to absorb increasing number of immigrants, this is also an important topic for the future. The concern related to segregated schools is part of a general concern that the level of dropout from secondary school is too high. In Norway, approximately 30 per cent of those that start at secondary school never finish. This has spurred the need for explanation. In this paper we analyse whether peer effects from schoolmates can contribute to the explanation. Concretely, we analyse the impact of immigrant concentration on native pupils performances in secondary school. The dependent variable is whether the pupil drops out of secondary school, measured by the likelihood of not completing secondary school within five years after entry. 2

The research literature has in recent years shown an increasing interest in modelling and measuring the effects of social interactions between pupils so called peer effects at different levels in the educational system; both at the class room level and at the school level. Studies have presented peer effects from race, gender, and immigrants (Angrist and Lang 2004, Hoxby 2000, Gould et al 2009, Lavy and Schlosser 2007). We return to this literature in the next section. There are several mechanisms that can explain why there might be a relationship between immigrant share of schoolmates and performance of native pupils. One potential mechanism is peer group quality. Since immigrants grades are generally lower than native grades (See Figure 1.1 below), a rise in the fraction of immigrant pupils lowers the mean grade potential of the student body. Figure 1.1 presents mean achievement distribution from national exams at the 8 th grade in Norwegian reading and calculus, taken from Statistics Norway (five achievement levels, 5 is highest). Immigrants here include all individuals born outside Norway by two foreign born parents. We see that immigrant pupils are clearly overrepresented in the lower two achievement levels compared to the rest of the pupil population. 3

Figure 1.1. National exams in Norwegian reading and calculus in the 8th grade. Distribution according to five achievement levels (5=highest). Overall results 2007 100 % 90 % 80 % 70 % 60 % 50 % 40 % 30 % 20 % 10 % 0 % Immigrants Source: Statistics Norway Reading Rest of the population 100 % 90 % 80 % 70 % 60 % 50 % 40 % 30 % 20 % 10 % 0 % Immigrants Calculus Rest of the population Level 5 Level 4 Level 3 Level 3 Level 1 If individual achievement is affected by the average test potential of classmates, an increase in the fraction of immigrants will lower the expected performance of students (holding constant their own characteristics). A second channel is school quality. Boyd et al. (2003), for example, argue that teachers prefer to teach in schools with lower minority enrolment, and that schools compete for more effective teachers, which might lead to an equilibrium in which better teachers are assigned to schools with fewer immigrants. An important goal of this paper is to disentangle between these effects, and identify causal effects. Tin the identification of causal effects of peers our paper relates to papers by Hoxby (2000) and Gould et al. (2009) both in the ethnical focus and because we as them use potential random variation in the number of immigrants between grades within the same school. Hoxby (2000) exploited across grade variation in peers within the same school that arise because of random demographic between cohorts. In our case: conditional on the number of immigrants in 11 th and 12 th (first and second year at secondary school), the share of immigrants in 11 th grade can be considered as being as determined random variation in the distribution of grades in the immigrant pool in the local area. This is the first part of our 4

Dropout rate identification strategy. In addition, to control for potential time varying school quality (time fixed effect school quality is controlled for by school dummies) we use a three year moving average approach including mean level characteristics at the school. The sample we study in this paper is all pupils starting at secondary schooling in the period 1996-2003. This group is followed until 2008, when all pupils have had five years to complete the studies. Figure 1.2 shows the relationship between the immigrant share and the drop out share in our material. It reveals a strong positive relationship between the immigrant share and the dropout share. The dropout share increases from 0.3 to 0.52 when you go from an all-native school to a school with 60 per cent immigrants. 1 It is especially for natives we find a strong positive correlation between the immigrant share and drop-out share. The relationships in Figure 1.2 are correlations only. To analyse whether this pattern also hides causal effects is the core object of this paper. Figure 1.2. Average dropout rate by share of immigrants in secondary school. 1996-2003 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0,1 0 0 0,1 0,2 0,3 0,4 0,5 0,6 Share immigrants All Immigrants Natives 1 There are no schools where all students are immigrants. 5

Our contribution to the literature is threefold. First, we present evidence from the secondary school level. The majority of studies have focussed on earlier schooling. We argue that more research on the secondary level is warranted. Second, we use a comprehensive set of controls aiming to reach causal statements, including within-school idiosyncratic variation and a moving average approach. Third, we present some tentative evidence of what kind of mechanism that lays behind the peer effect. The results show that the concentration of immigrants has a positive impact on the dropout likelihood of natives. The preferred estimate suggests that a 10 percentage point increase in the fraction of immigrants lead to a 3 percentage point increase in the dropout rate. This effect is sustained after controlling for a set of controls aiming at presenting causal effects. The result is also strengthened after running a placebo test constructing peers from another grade and another field of study. The paper proceeds as follows: in the next section we give a brief overview of related literature. Section 3 gives a brief overview over the Norwegian educational system and immigration history. Section 4 presents the data, the variables, and the sample. Section 5 presents the modelling and identification strategy, section 6 presents the results, and section 7 concludes. 2. Related literature Our study relates to the economic literature analysing peer effect in education (Ammermueller and Pischke 2006, Black et al. 2010, Lavy and Schlosser 2006, Gibbons and Talhaj 2008, Lavy et al. 2009) and especially to that part of the literature analysing ethnical peer effects (Hoxby 2000, Gould et al. 2009, Card and Rothstein 2006, Hanushek et al. 2009). Gould et al. (2009) analyse the impact of immigrant concentration in elementary school on the long-term academic performance of native students in high school in Israel. They control for the 6

endogeneity of immigrant placement between schools by conditioning on the total number of immigrants in a school and exploit random variation in the number of immigrants across grades within the same school. Their results suggest that the overall presence of immigrants in a grade has a significant and large negative effect on the dropout rate and also a negative effect on the chances of passing the high school matriculation exam which is necessary to attend college. Hoxby (2000a) exploits idiosyncratic variation in gender and race composition of neighboring cohorts in public schools in Texas. Her results show that children s elementary school test scores are affected by those of their peers, with intra-race peer effects appearing to be particularly strong. Card and Rothenstein (2006) analyse the relationship between racial segregation and black-white test score gap in the US. They compare black-white test score gaps across metropolitan areas that differ in the extent of school segregation. They decompose the effects of school segregation into three components: one arising from differences in residential sorting patterns, one due to court ordered school desegregation, and a residual. The results show that the residential component has a large effect on test score gaps, while the remaining components do not. They conclude that the composition of neighborhoods matters, but not that of schools. Hanushek et al. (2009) use a rich panel data on the achievement of Texas students and disentangle racial composition effects from other aspects of school quality and from differences in abilities and family background. Their results suggest that a higher percentage of black schoolmates reduce achievement for blacks, while it implies a much smaller and generally insignificant effect on whites. 7

3. The Norwegian educational system and immigration to Norway The educational system The compulsory school starting age in Norway has been six years since the 1997 Educational Reform. The reform of 1997 also increased the years of compulsory education from 9 to 10. However, for our sample of pupils the school starting age is 7 and they go through 9 years of compulsory education. The municipalities are in charge of operating schools to provide compulsory education (1st 10th grade, ages 6 15). There are no ability school tracking in compulsory school. Residence determines which compulsory school the children go to. Furthermore, the share of private schools at the primary level is very low in Norway. After completed compulsory school all pupils are as from 1997 guaranteed at least three years of secondary level education. The sample we study is pupils that are registered at the first year at secondary school, which is the 11 th grade. At secondary level the pupils can choose between an academically oriented track and a vocational track. The vocational track leads to occupation and professional qualifications. The vvocational track usually consists of two years learning in school followed by two years on the-job training through an apprenticeship or other scheme. The three year academic track qualifies for college or university admission certification. However, the system permits that also pupils following a vocational track can switch to the academic track after two years. In other words, the system permits to change the track along the way. Immigration to Norway Since World War II, Norway has had considerable restrictions on labour immigration from non-western countries. The exception is a period of liberalisation between 1957 and 1975. In this period, and especially in the beginning of the 1970 s, there was a considerable influx of low skilled labour immigrants from countries like Pakistan, Turkey and Morocco. In 1975, an 8

immigration stop from countries outside of the Nordic countries was implemented. Exceptions were made for those with specialized skills not available in the Norwegian labour market. The immigrants share of the Norwegian population has increased considerably since the immigration stop of 1975, from approximately 2 per cent of the population in 1980 to approximately 8 per cent in 2005. During the same period, the composition of the immigrant population has changed, from being dominated by Nordic and western immigrants to at present being dominated by immigrants from non-western countries. By 2004, almost 75 per cent of the immigrants in Norway were non-western immigrants compared to 25 per cent in 1980 (Statistics Norway 2006). The increase in immigration from non-western countries is due to the influx of refugees and asylum seekers, and because of family reunifications. A common immigration pattern for the refugees are that the male arrives first, then the wife and the children arrive later through family reunification. The largest recent groups of refugees and asylum seeks have come from Bosnia-Herzegovina and Kosovo, in the aftermath of the Balkan wars at the start of the 1990 s. Refugees and their families are granted a residence permit if their application for asylum is approved. They may apply after entering the country on their own, or the UN may apply for them (Quota refugees). When the residence permit has been granted, the new immigrants should be settled in a municipality within six months. 4. Data, variables, and sample All the analyses are based on a comprehensive set of individual register data collected and organized by Statistics Norway. The starting point is individual register information from The Norwegian educational data base (NUDB) containing detailed longitudinal information on all pupils from compulsory school, secondary school, and higher education. With a unique personal identifier NUDB is linked to other registers containing demographic information. 9

The sample The starting point is information on all pupils starting at secondary school each year in the period 1996-2003. That means that we have seven fresh cohorts of pupils. We follow these pupils until 2008. By then all cohorts have had five years to complete their secondary level education. To the seven cohorts of pupils we link data on individual characteristics (fixed and time varying), school related characteristics, regional information (neighbourhood and municipality), and information on their peers. We focus on men and women that are 16 years old when they start their first year of secondary school, i.e., we limit the analyses to pupils that start at secondary school when they are supposed according to their time of birth. Variables Our dependent variable is a dummy variable measuring whether the student has dropped out of secondary school, or not, during the five years period we follow them. Our criterion is that if the pupil is not registered as having passed all the final exams five years after starting at secondary school, he or she is classified as a dropout pupil. This is the standard definition of dropout in the Norwegian school system. Individual information includes: Age, gender, number of siblings, whether her or she is the oldest child, whether the student started an academic or vocational track at secondary school, mothers years of education, fathers years of education, mothers yearly income, fathers yearly income, and. Parents education is measures by six dummy variables; i) compulsory school, ii) secondary school low level (one or two years completed education after compulsory school), iii) secondary school high level (completed full secondary education), iv) college/university low level (lower than master degree), v) college/university high level,(master degree and higher, and vi) unknown education. Mother and father s yearly income is measured as average yearly total income from the period here the pupil is 7 years 10

old till he or she is 16 years old, deflated to 2003-Norwegian kroner. For the peers (nonwestern immigrants) we include peer level mean values of all the same variables as for the individual variables. Control variables at the school level include information on number of pupils in each cohort, and fixed school effects. Time varying regional controls include aggregated mean values of the average wage level, the educational attainment and the share of non-western immigrants. This is measured at the neighbourhood ( grunnkrets ) level. Neighbourhood is the most detailed regional unit in Norway. There are approximately 13000 neighbourhoods in Norway. We also control for local unemployment rate in the municipality (there are 435 municipalities in Norway). The key explanatory variable is the share of immigrant pupils in the same cohort in the first year of secondary school. Immigrants in this paper are synonymous with non-western immigrants, and includes individuals born in Asia (including Turkey), Africa, South America and Central- and Eastern European countries, of two foreign born parents. In our sample the six largest country of births are (per cent of all immigrants in our sample in parenthesis) Bosnia-Herzegovina (13.4), Iran (8.7), Pakistan (7.9), Kosovo (6.8), Vietnam (6.4), and Chile (5.9). Refugees from Bosnia-Herzegovina and Kosovo came as a consequence of the Balkan wars in the 1990 s. The average age at arrival in our sample is 7.5 years. 5. Identification strategy A positive relationship between individual behaviour and the behaviour of a group the individual interacts with many exist for several reasons. Manski (1993) distinguishes between three types of reasons: i) endogenous social interaction effects, arising from the mechanism that behaviour of persons in the group directly affects the behaviour of an individual member 11

of the group, ii) contextual interactions, where the behaviour of a person in some way varies with exogenous characteristics of the group members, and iii) correlated effects, where persons in the same group tend to act in the same way because they have equal individual characteristics or face similar institutional environments. Our target in this paper is to present evidence on the first of these mechanisms. Five main identification strategies dominate the economic literature on this topic. One strategy is randomized experiments where peers are distributed randomly, sweeping away all problems related to unobserved selection. Analyses of randomized room-mates in college are probably the most well know studies exploiting this variation (Sacerdote 2001). A second strategy has been to estimate individual fixed effects models, using movers between school to identify causal effects (Hanushek et al. 2003). A potential problem with this approach is that mobility between schools may be endogenous with respect to school characteristics. A third approach has been to use instrument variables (Goux and Marin 2007). A fourth approach has been to aggregate to a level where sorting is reduced or eliminated (Evans et al 1992, Card and Rothenstein 2006). A fifth approach has been to use idiosyncratic variation arising from within-school variation in immigrant exposure by exploiting variation in the immigrant composition of each cohort within each school (Hoxby 2000, Gould et al. 2009, Hanushek et al 2009. They assume that while pupils may sort themselves between schools based on factor like the immigrant composition, these choices are unaffected by cohort-specific variation in these factors. It is the latter approach we draw on in this paper. The modelling strategy is described in the detail in the next section 12

5.1 Model The model we want to estimate can be explained as follows: (1) A ics a 6 c 1 1 a x a 5 d c 2 ics a a 6 x a IMM 2 NIMM _11_12 3 ics a g ics 4 rc sc u s ics where i stands for individual, c for cohort, and s for school. A ics is the outcome variable, taking the value 1 if the individual has not completed secondary school five years after first being registered and 0 otherwise, xics is a vector of individual characteristics (gender, number of siblings, whether the pupil the oldest child, whether the pupil starts on a vocational or academic track, mother s educational attainment, father s educational attainment, mother s yearly income, and father s yearly income), x is a vector of the average values of the corresponding variables for the peers, g sc is a measure the number of pupils in that cohort (in the freshman year), d c are cohort dummies (1996-2003) and rc accounts for the time varying characteristics of the neighbourhood. Neighbourhood is the smallest administrative geographical unit in Norway. In total there are 13000 neighbourhoods in Norway. Fixed neighbourhood effects are taken care of by the fixed school effects. The time varying variables include the mean level of yearly labour market earnings, the local skill distribution (six dummy variables), and the share of non-western immigrants, all the neighbourhood level, which are measured for all inhabitants in the neighbourhood 20-60 years of age. Furthermore, s picks up school fixed effects. IMM is the key variable and measures the share of nonwestern in the cohort. NIMM_11_12 measures the total number of immigrants in the pupils own grade (grade 11) and the grade above (grade 12). This variable is used for identification reason. We assume that conditional on the number of peer immigrants in both grade 11 and 12, the fraction of immigrants in grade 11 is determined by random variation in the grade distribution 13

among the pool of immigrant pupils in the neighbourhood. t. Finally, although not shown in equation (1), we also include controls for the annual local unemployment rate. The coefficient in equation (1) that we are especially interested in is a 3. If the identification criteria holds and the control for time varying school quality (explained below) is successful, a 3 this gives us the causal impact of peer immigrants on the schools performance of native pupils. In (1) we assume that, conditional on the number of immigrants across grades (IMM_11_12), grade dummies and school fixed effects, peer characteristics in a particular cohort are potentially randomly assigned. This is our identification strategy. We know that selective residential sorting takes places but such effect should be picked up by the conditioning of the number of immigrants across grades and the school fixed effects (λ). However, to further control for potential time varying school quality effects (for example schools may get bad reputations which may lead them to get lower quality pupils in the future) we use a moving average approach (Black et al. 2009), by which instead of comparing over a prolonged period of time, we narrow the timespan. For every year we construct a measure of the average characteristics of pupils that same year, the previous year and the following year. Since within each three year window changes in mean pupil characteristics cannot be due to a linear trend over this three year period. Therefore it can be treated as idiosyncratic variation. In equation (1) we add the three year moving average approach as additional explanatory variables. Therefore, our identification strategy combines approaches from both Gould et al. (2009) and Black et al. (2010) in the sense that we incorporate both the within-school across grades approach and the moving average approach. 14

6. Results First, Table 6.1 presents some descriptive statistics of the individual, of the peers, as well as school characteristics, and regional characteristics. Statistics are presented for all, and separately for those that completed secondary school and for school dropouts. In our data, 29 per cent of the pupils drop out, i.e., they do not complete secondary school within a five year window. Female pupils complete secondary school more than male pupils. Furthermore, if you are the oldest sibling you seem to have a somewhat higher likelihood of completing. The advantage of being the oldest sibling is well funded in the empirical literature of birth order (see for example Black et al. 2007). The dropout rate is also higher among those on the vocational track than among those that are on the academic track. Regarding parents education, dropouts have a higher fraction of parents with low formal skills compared to those that complete. This is the case for both mothers and fathers education. The same picture applies for parents income. Parents of pupils that drop out have on average lower income compared to parents to pupils that complete secondary school. 15

Table 6.1. Descriptive statistics. All, completed and dropouts All Completed Drop out Individual characteristics: Women 0.487 0.520 0.399 Number of siblings 0.958 0.967 0.934 Oldest child 0.548 0.556 0.526 Academic track 0.479 0.398 0.696 Mothers education: Compulsory school 0.289 0.236 0.431 Secondary school I 0.236 0.238 0.233 Secondary school II 0.166 0.169 0.158 College/university I 0.276 0.319 0.162 College/university II 0.029 0.036 0.010 Unknown education 0.003 0.002 0.006 Fathers education: Compulsory school 0.209 0.165 0.326 Secondary school I 0.201 0.197 0.211 Secondary school II 0.257 0.257 0.259 College/university I 0.226 0.257 0.144 College/university II 0.094 0.116 0.037 Unknown education 0.013 0.009 0.023 Mothers yearly income (in 1000 NOK) 158.519 167.019 135.774 Fathers yearly income (in 1000 NOK) 328.486 347.998 276.277 Peer characteristics: Women 0.473 0.481 0.450 Age 16.305 16.300 16.339 Number of siblings 1.504 1.488 1.545 Oldest child 0.654 0.652 0.660 Peers mothers education: Compulsory school 0.424 0.421 0.430 Secondary school I 0.037 0.036 0.039 Secondary school II 0.163 0.165 0.159 College/university I 0.107 0.113 0.092 College/university II 0.032 0.034 0.028 Unknown education 0.237 0.231 0.252 Peers fathers education: Compulsory school 0.298 0.297 0.301 Secondary school I 0.041 0.041 0.042 Secondary school II 0.145 0.146 0.140 College/university I 0.122 0.125 0.115 College/university II 0.040 0.043 0.032 Unknown education 0.354 0.348 0.370 Mothers yearly income (in 1000 NOK) 55.295 56,685 51.647 Fathers yearly income (in 1000 NOK) 100.685 103.174 94.028 School characteristics: Number of schools 427 Number of students 176.457 175.698 178.479 Fraction non-western immigrants 0.039 0.038 0.041 Number of immigrants at grade 11 and 12 7.043 6.805 7.678 Neighborhood characteristics: Yearly labour market earnings (in 1000 NOK) 210 210 204 Compulsory school 0.239 0.233 0.257 Secondary school I 0.182 0.182 0.183 Secondary school II 0.271 0.271 0.271 College/university I 0.237 0.242 0.224 College/university II 0.050 0.053 0.043 Unknown education 0.018 0.017 0.019 Non-western immigrants 0.033 0.031 0.038 Community characteristics: Unemployment rate in municipality 3.404 3.371 3.472 N 277233 201813 75420 Note: The sample compromises all students enrolled in secondary school 1996-2003. NOK is Norwegian kroner. 16

Regarding the time varying neighbourhood characteristics there are no large differences between dropouts and no-dropouts. Still, the mean level labour market earnings are somewhat lower, the share with higher education is somewhat lower, and the share of non-western immigrant inhabitants is somewhat higher in dropout neighbourhoods. Regarding the peer level variable, it is noteworthy that the share of parents with higher education is much lower among parents of immigrant pupils than among parents of native pupils. The share of parents with missing education is much higher among the immigrant parents than among native parents. This is natural since the vast majority of the parents have their education from abroad. The share missing is especially high for father s education. Regarding the mean income of immigrants parents this is considerably lower compared to parents of natives. Lower income of immigrants compared to natives is a robust finding in the economic assimilation literature. The peer variables do also reveal that immigrant pupils come from larger families, measured by the number of siblings they have. Regarding the mean age level of peers it is just above 16 years, with only a small difference between completed and dropouts. Finally, regarding the mean regional unemployment rate, it is marginally higher in municipalities where the dropout pupils live. Before we proceed to the presentation of the results from the regression analyses we show results from some simple balancing tests. The underlying assumption in the analyses is that conditional on the controls, variation in the share of immigrants should as close to random. Conditional on controls, the share of immigrants should not be significantly related to the predetermined variables as parental education and number of siblings. Table A1 in Appendix presents results from different regressions, where the dependent variables are different individual variables (mothers education, fathers education, number of siblings, and whether the individual is the oldest child), regressed against the peer variable. We present two models; one where only the peer variable is included as explanatory variables, and one where 17

we include the full battery of controls (excluding all the individual variables). The results in the first model show significant and strong relationships with the peer variables. However, when we control for the rest of the variables, all significant relationships disappear (except for oldest child variables which is marginally significant at 10 per cent). The large gap between the gross coefficient in model 1 and the net-coefficient in model 2 suggest that our identification criteria are able to strongly reduce potential bias from selection of immigrant pupils between schools. Table 6.2 presents the first regression results. We present results for the following models: Model 1 includes seven years dummies, in addition to the fraction of immigrants. Model 2 adds individual variables plus variables for the number of immigrants in 11 th and 12 th grade.model3 adds the remaining variables, including school fixed effects. All models are estimated using simple linear probability models. Finally, Model 4 is the same as Model 3, but now instead of only including native pupil that were 16 at the start of secondary school, we include all freshmen, irrespective of age. Table 6.2. Estimates for the effects of immigrants on native pupil s performance. All 1 2 3 All native pupils 0.199** (0.091) The share of non-western immigrants 0.369*** (0.088) 0.409*** (0.070) 0.212** (0.091) Additional controls: Year effects? Yes Yes Yes Yes Number of immigrants in grade 11 and 12? No Yes Yes Yes Individual? No Yes Yes Yes Peer effects? No No Yes Yes School - time varying effects? No No Yes Yes School - fixed effects? No No Yes Yes School quality moving average? No No Yes Yes Regional unemployment? No No Yes Yes Time varying neighbourhood effect No No Yes Yes N 277233 277233 277233 285021 R 2 -adj 0.01 0.08 0.135 0.135 Note: Individual variables include woman, number of siblings, whether you are the oldest sibling, vocational programme, mothers education, fathers education, mothers yearly income, fathers yearly income. Time varying school effects include number of pupils. Level of significance: *** 1 per cent; ** 5 per cent; * 10 per cent. Robust standard error clustered at the school level. 18

Model 1 reveals a positive and significant relationship between the fraction of immigrants and the drop-out rate of natives. The coefficient suggests that a 10 percentage point increase in the immigrant share leads to a 3.7 percentage point increase in the dropout rate. Adding individual controls as well as control for the number of immigrants in grade 11 and 12 in Model 2, does not change the size of the coefficient much.. In Model 3 we add the rest of the controls, including school-fixed effects and the three year moving average variables. This reduces the immigrant share coefficient considerably, but the impact is still significant at level 5 per cent. The coefficient suggests that a 10 percentage point increase in the immigrant share leads to a 2 percentage point increase in the dropout rate. Therefore, our preferred estimate suggests that it is a quite substantial peer effect of immigrants on native performance. Finally, Model 4 re-estimates Model 3, but not limiting the native sample to those that were 16 years old at the beginning of secondary school. This is in a sense a more symmetric sample, since we now relate all native pupils to all immigrant peers. However, the results show that the peer effect is only moderately altered. Therefore, our results are not driven by analysing thee peer effect for native 16 year olds. One natural extension of the estimations in Table 6.2 would be to test if the peer effect is linear or non-linear. Results from a model where we just added a squared element of the fraction of immigrant variable showed no significant impact from the squared variable (available upon request). The endogeneity of school age One might be concerned of the possibility that immigrants more than natives are held back a grade. If so they are placed in grades with better or worse native pupils compared to where they should have been according to their date of birth. This has been raised as a concern in some studies (see for example Gould et al. 2009). This should not be a problem in our data. 19

Holding pupils back in compulsory school is very rare in Norway. Still, in our data we observe that the share of immigrants that start on time in secondary school are lower compared to natives. Approximately 94 per cent of natives in our cohorts start on time, i.e., in the year when they are 16. The comparable fraction of immigrants is 68 per cent. Therefore, a larger share of immigrants starts at secondary school at an older age than natives; the average age of immigrants is higher than the average age of natives. The major share of the late secondary school starters consists of immigrants that arrived Norway after the age of six, i.e., when they have passed primary school start age. 2 Therefore, this is a different explanation compared to the holding-back explanation that was the concern in Gould et al. (2009). In our case - when this is caused by immigrants arriving Norway after primary school start age - it is not obvious that this is a bias that we want to sweep away. If immigrants arriving in early youth are lesser equipped with skills that ease learning this would be part of the peer quality effect mechanism. Then, if individual achievement is affected by the average test potential of schoolmates a rise in the share of this group of immigrant pupils may lower performance of other pupils. Still, to shed light on this issue we run a 2SLS approach much in the spirit of Gould et al. (2009). We instrument the observed share of immigrants in your cohort with the predicted share of immigrants that would have been if all pupils started in secondary school according to their year of birth. Table 6.3 presents the results: 2 For immigrant that arrived Norway before they were six, the share enrolled in secondary school at 16 is 90 per cent. If we look at immigrant arriving before they were five, the on schedule enrollment rate is 92 per cent and almost on par with natives. 20

Table 6.3. Estimates for the effects of immigrants on native pupil s performance. 2SLS Second stage results The share of non-western immigrants 0.207 (0.238) Instruments and first step Predicted share of immigrants 0.563*** (0.034) F-test excl. instr. 233.31 Kleibergen-Paap F 233.31 N 265691 R 2 -adj 0.135 Note: Individual variables include woman, number of siblings, whether you are the oldest sibling, vocational programme, mothers education, fathers education, mothers yearly income, fathers yearly income. Time varying school effects include number of pupils. Level of significance: *** 1 per cent; ** 5 per cent; * 10 per cent. Robust standard error clustered at the school level. The instrument is highly appropriate, and it correlates as expected positively and strongly with the observed share of immigrant peers. The Kleibergen-Paap tests for weak instruments reveal strong instruments. Since we have exact identification we cannot perform the overidentification test. The result for the immigrant share in the second step shows that the size of the peer-effect is only moderately altered compared to the earlier estimates. The standard error has increased though, leading to a non-significant peer-effect. However, since the size of the coefficient is till sizeable, and since the main argument for the control procedure (controlling for holding back pupils) is less relevant in the Norwegian setting, we proceed with original set-up, not using the 2SLS approach. However, in section 6.4 we present analysing separating the peer effect from early and later arriving immigrants. 6.1 Results for subgroups So far we have only reported regression results for the whole group of native pupils. To see whether the effects differ between subgroups we present results depending on choice of track and parents education. The descriptive statistics in Table 6.1 showed that the dropout rate is higher for those following in the vocational track than for those choosing the academic track. The share of immigrants in the two tracks is very similar, approximately 3.9 per cent in both 21

tracks. Table 6.4 presents results when we estimate equation (1) separately for academic and the vocational tracks. In the presentation we confine ourselves to the most elaborated model (Model3 in Table 6.2). Table 6.4. Estimates for the effects of immigrants on native pupil s performance. Depending on programme of study Academic programme Vocational programme The share of non-western immigrants 0.352*** (0.119) 0.147 (0.131) Additional controls: Year effects? Yes Yes Number of immigrants in grade 11 and 12? Yes Yes Individual? Yes Yes Peer effects? Yes Yes School - time varying effects? Yes Yes School - fixed effects? Yes Yes School quality moving average? Yes Yes Regional unemployment? Yes Yes Time varying neighbourhood effect? Yes Yes N 147570 129663 R 2 -adj 0.102 0.074 Note: Individual variables include woman, number of siblings, whether you are the oldest sibling, vocational programme, mothers education, fathers education, mothers yearly income, fathers yearly income. Time varying school effects include number of pupils. Level of significance: *** 1 per cent; ** 5 per cent; * 10 per cent. Robust standard error clustered at the school level. The impact of immigrant peers on the likelihood that native pupils will drop out of school is positive for both tracks, but only statistically significant for native pupils in the academic track. One possible explanation is that the peer effect from classmates might be more important in more theoretical and academic classroom courses than in more practically oriented teaching environments. 6.2. Placebo controls In order to put our identification strategy to the test we carry out two so called placebo tests. The first test replaces the actual share of immigrants in own grade with the share of immigrants in the grade above, i.e., the pupils that started the year before. If we pick up peer effects we should get a smaller coefficient when me measure the effect of peer immigrants 22

from the grade above. The second test is to runs a regression among those in the academic track when we construct two peer variables; one with the share of immigrants in the academic track and one with the share of immigrants in the vocational track This exercise is limited to schools with both academic and vocational programmes. The hypothesis is that the peer effect should be much larger from own peers than peers in the vocational track Table 6.5 presents the results. Table 6.5. Placebo regression. Estimates for the effects of immigrants on performance of native pupils. Peers from another grade and peers from another track Peers from another grade The share of non-western immigrants 0.002 (0.116) The share of non-western immigrants at vocational programme The share of non-western immigrants at academic programme Additional controls: Peers from another branch of study. Academic programme 0.085 (0.051) 0.255*** (0.083) Year effects? Yes Yes Number of immigrants in grade 11 and Yes Yes 12? Individual? Yes Yes Peer effects? Yes Yes School - time varying effects? Yes Yes School - fixed effects? Yes Yes School quality moving average? Yes Yes Regional unemployment? Yes Yes Time varying neighbourhood effect Yes Yes N 276450 87832 R 2 -adj 0.134 0.121 Note: Individual variables include woman, number of siblings, whether you are the oldest sibling, vocational programme, mothers education, fathers education, mothers yearly income, fathers yearly income. Time varying school effects include number of pupils. Level of significance: *** 1 per cent; ** 5 per cent; * 10 per cent. Robust standard error clustered at the school level. The first model in Table 6.5 shows that there are no significant effects of peers from the grade above. This is as expected if you believe that the peer effects are coming from school mates in your own cohort. The second model shows that the peer effect is more than twice as strong from peers in your own track than from peers in the other track (vocational track). This 23

finding does also strengthen the interpretation that the previous results are actually picking up ethnic peer effects on the school performance of natives. 6.4. Mechanisms In this final section we turn to the question how do the peer effects arise, i.e., can we identify the mechanisms of social interaction? To answer this question we focus on the importance of peer quality. As mentioned in identification strategy section on potential explanation is peer quality. Using register data this is of course a difficult answer to answer. This is probably one reason why rather few studies have tried to unravel this mechanism (Lavy et al. 2009, Duflo 2008). We know that immigrants grades are generally lower than native grades. If the individual s achievement is affected by the average grade potential of the other pupils, an increase in the fraction of immigrants will lower the expected performance of students We only have information on grades from compulsory school for the two last cohorts of pupils. This leaves us with a very short time span and relatively few observations. Instead, to construct an indicator of peer quality we use two other approaches. The first is parents educational attainment for which we have information over the whole period. We know that it is a positive correlation between the parents education and the offsprings educational attainment. Therefore we use the parents education as an indicator of the grade capacity of the children. We construct two measures: i) the share of immigrant peers that have parents where at least one of them have higher education, and ii) the share of immigrants where none of the parents have higher education. 3 We then re-estimate equation (1) including these two variables. The hypothesis would be that the peer effect - if it picks up peer quality should be stronger from the peers with parents with low education (category ii). The second indicator of 3 This latter group includes parent with missing education 24

peer quality is based on information on the immigrants age at arrival to Norway. We construct two groups: i) Child immigrants, that is immigrant arriving Norway when they were seven year or younger, and ii) Adult immigrants, that is immigrants arriving Norway when they are older than seven years. Our hypothesis if peer quality is important is that the peer effects should be stronger from adult immigrants, since this group has had shorter time to acquire language and other region specific skills. Table 6.6 presents the results. We only present results from the most elaborated model: The left hand side of the table presents results based on parents education, while the right hand side presents results based on age at arrival. Table 6.6. Mechanisms. Estimates for the effects of immigrants on performance of native pupils. The share of non-western immigrants with parents with higher education The share of non-western immigrants with parents with lower education Parents education 0.042 (0.183) 0.296** (0.111) The share of non-western immigrants that is child immigrants The share of non-western immigrants that is adult Age at arrival 0.050 (0.131) 0.305*** (0.100) immigrants N 277249 277249 R 2 -adj 0.134 0.134 Note: Individual variables include woman, number of siblings, whether you are the oldest sibling, vocational programme, mothers education, fathers education, mothers yearly income, fathers yearly income. Time varying school effects include number of pupils. Level of significance: *** 1 per cent; ** 5 per cent; * 10 per cent. Robust standard error clustered at the school level. The results reveal that it is only from the immigrants that have parents with low education that we find a positive and significant peer effect (0.296). From immigrants with parents with higher education we still find a positive effect, but much smaller and not significant effect. Regarding age at arrival it is only from adult immigrants we find a positive and significant peer effect. In sum, these results suggest that peer quality might be an important mechanism behind the general peer-effect result. Immigrants with parents with lower education, and immigrants arriving as adults may have a skill deficit compared to immigrants with parents with higher education and immigrants arriving as children. A rise in the fraction of immigrants from these two groups may lower the performance potential of the student body. 25

7. Conclusion The immigrant share of the Norwegian population has increased from two to ten per cent during the last 30 years. During the same period the composition of immigrants have changed, from being dominated by immigrants from Nordic and other western countries to at present being dominated by immigrants from non-western countries. This development and change in composition is found in most modern western countries. The increased immigrant share of the population has spurred a large economic research literature analysing labour market impacts of immigration on receiving countries, especially related to consequences for employment and wages. In this paper we report to a more scant but still growing literature focusing on the impact of immigration on native pupils school performances. Parallel with the increasing number of immigrants the educational policy has to larger extent been concerned with impact of segregated schools, especially in the primary and secondary level. We analyse the impact of immigrant concentration on native pupils performances in secondary school. Performance is measured in terms of the likelihood of dropping out of secondary school. Dropping out is defined as not completing the final exam five years after being enrolled. The sample we use is all pupils starting at secondary school each year in the period 1996-2003. We follow these pupils until 2008. By then all cohorts have had five years to complete their secondary level education. To reach to causal statements we exploit potential random variation in the number of immigrants between grades within the same school and we control for time fixed and potential time-varying school quality. The tentative results reveal a positive and significant relationship between the fraction of immigrants and the dropout rate of natives. The preferred coefficient suggests that a 10 percentage point increase in the immigrant share leads to a 2 percentage point increase in the dropout rate. As the results survive our different control procedures we interpret the result as 26

a causal mechanism. The results are strengthened by the results from a simple placebo test, using pseudo -peers from other tracks as peers. Can we identify the mechanisms of social interaction? We look especially at the importance of a peer group quality effect. We use two indicators of peer group quality: the education of the peers parents and information on whether the peers are child immigrants or adult immigrants. The results show that it is only from the immigrants that have parents with low education that we find a positive and significant peer effect. Regarding age at arrival it is only from adult immigrants we find a positive and significant peer effect. These results suggest that peer quality might be an important mechanism behind the general peer-effect result. Our results give arguments to those saying that ethnical segregated schools may be harmful for the performance of pupils. This may be especially important to consider in a period where Norway and other western countries are expected to absorb increasing number of immigrants. Finally, one caveat: we control for school quality by using time fixed and time varying indicators. We cannot rule out that our control procedures do not sweep out all components of school quality. For instance, we lack valuable information on the teachers. Therefore, our results could be a mixture of peer and school effects. References Ammermueller, A and J. Pischke (2009): Peer Effects in European Primary Schools: Evidence from PIRLS, Journal of Labor Economics, 27, 315-348. Angrist, J.D. and Lang, K. (2004). _Does school integration generate peer effects? Evidence from Boston s Metco Program_, American Economic Review, vol. 94(5) (December), pp. 1613 34. Black, S., P. Devreux, and K. G.Salvanes (2007) Older and Wiser? Birth order and IQ for young men. NBER working paper No. 13237. 27