Immigrant background peer effects in Italian schools Dalit

Immigrant background peer effects in Italian schools Dalit Contini University of Torino Improving Education through Accountability and Evaluation, Roma 3 -5 October 2012

The research question Do high concentrations of immigrant background children in schools hamper the learning of native children (and of other immigrant children)?

Motivation Italy is a recent immigration country. Immigrant background children in primary and lower secondary schools have increased from 3% to 9% at the national level over the last decade. Widespread concern that immigrant children could be detrimental for the learning of natives. Is the concern supported by empirical evidence? The research question is relevant for the quality and equity of the schooling system and for social cohesion. It has implications on the distribution of children into schools and allocations of resources.

Data § INVALSI standardized learning assessment 2010 § Reading comprehension and math § Administered to the entire populations at the national level (~ 500. 000 students per grade) § Info on family background provided by student questionnaire and school administrations § Children nested into classes nested into schools § I analyze children in grades 5 and 6 in the North and Centre (where the majority of immigrants live)

State of the art o Existing literature on peer effects mainly focuses on socio-economic status, gender and ethnic differences. Less effort directed to the estimation of peer effects related to immigrant background. o Findings from previous studies on ethnic composition of schools may not be relevant for the more recent immigrants. o EU papers on immigrant background peer effects: Cebolla-Boado (2007) achievement in lower secondary school in France generally small effects Van der Silk et al. (2006) and Dumay (2008): achievement in the Netherlands not always significant Agirdag et al (2011) achievement of lower secondary school in Flemish Belgium Cebolla-Boado and Medina (2011) primary education in Spain no research on Italy Fekjaer and Birkelund (2007) on upper secondary graduates in Norway Brunello and Rocco (2011) upper secondary achievement (PISA. Not on Italy) Gould et al. (2009) 5° grade achievement on later educational outcomes in Israel

Descriptive evidence • Large immigrant/native achievement gaps. Gaps are larger for first generation, but are also large for second generation. • On average scores (of natives and of immigrants) are lower in schools with high concentrations of immigrant children. Causal effect? • Schools with many immigrant children are attended by lower SES native and immigrant children: possible confounding effect. Allocation of children in schools.

Assumption: Structural model other characteristics of peers peer effects operate at the class level achievement of peers Causal effects achievement of peers characteristics of peers individual characteristics Spurious effects school and class unobserved effects school and class characteristics

Reduced form model g* is the parameter of interest • measures class composition effects • captures peer achievement and characteristics effects • policy relevant composite error term Problem: Why should school or class unobserved specific effects be correlated with peer characteristics? • • school selection (freedom of choice/area of residence) class allocation (is it random? )

Addressing selection in the peer effects literature Hoxby (2000) exploits idiosyncratic within-school variation in peer characteristics between adjacent cohorts in given grades. Ammermueller, Pischke (2009) rely on differences in the compositions of individual classes within a school. Gould et al. (2009) study later educational outcomes and exploit random variation in the number of immigrants in grade 5, conditional on the number of immigrants in grades 4 -6. Black et al. (2010) study post-school and labor-market outcomes, exploiting random variation in cohort composition within schools. Hanushek et al. (2003) use panel data to estimate peer effects on test score gains over time using student and school-by-grade fixed effects in a value-added specification. Identification is achieved by exploiting the fact that students change schools.

Addressing selection INVALSI data allow this strategy (impossible with PISA, difficult with PIRLS, TIMSS. . ) By exploiting within-school variability in class composition we remove school-specific effects, hence solve the school selection problem. The class allocation problem is less relevant. Yet: § despite broad recommendations to maximize class heterogeneity there are no binding rules, so school boards may use other criteria (segregate disadvantaged children, limited ability streaming) § families are sometimes allowed to express preferences for particular classes Random allocation of children into classes: error independent of explanatory variables

Random allocation? Random allocation of immigrant background children implies school-level independence between immigrant status and class. System-level (X 2 test): random assignnment rejected School-level (Fisher‘s exact test) with a=0. 10: random assignnment rejected in ~ 20% schools I analyze schools passing both tests: ~ 60% School-level with respect to SES (Anova) with a=0. 10: random assignnment rejected in ~ 30% of schools Underlying hypothesis: the class formation process is not related to performance, given class composition.

Possible biases What if non-random allocating schools are not completely eliminated? • no bias if teachers randomly assigned to classes • overestimate peer effects if better teachers to “better” classes • underestimate peer effects if better teachers to “worse” classes Rationale of this option: Ability streaming + better resources to the more in need. Highly unlikely in Italy. Streaming is not a popular pedagogical practice in primary and lower secondary school.

Variables Dependent variables Reading & math scores = % correct answers [0 -1] Explanatory variables Individual Class composition Female SES (n° books, ESCS) Native repeating grade 1°generation 2°generation Sample*1°generation Sample*2°generation % Females heterogenous effects mean SES allowed: % Natives repeating grade % 1°generation - immigrants/natives % 2°generation - natives of different SES % 1 G*native % 2 G*native % 1 G*native*SES % 2 G*native*SES

Immigrant background peer effects A 10 % points increase in the share of immigrants reduces the number of correct answers by less than 1% (=1/20 pop st dev) 5 TH GRADE READING 5 TH GRADE MATH 6 TH GRADE READING 6 TH GRADE MATH -0. 085*** -0. 037** -0. 045*** -0. 035** +0. 002 -0. 005 -0. 075*** -0. 029* +0. 017 -0. 009 -0. 071*** -0. 009 +0. 053** -0. 046*** -0. 005 +0. 036** -0. 021 -0. 072*** -0. 002 +0. 067*** fraction 1 G on: Immigrant Native-SES low Native-SES med Native-SES high fraction 2 G on: Immigrant Native-SES low Native-SES med Native-SES high N° children 120. 000 -140. 000 N° classes 7000+ N° schools 1750+

Main conclusions (i) The concentration of immigrant children in schools should not be an issue of major concern as there is little evidence of substantial detrimental effects on students’ learning. (ii) The effect is somewhat larger for children from disadvantaged backgrounds (immigrants and low SES) and negligible or even positive for high status native children. (iii) On the other hand, the relative disadvantage of immigrant children at the individual level is large.

Thank you for your attention!

Descriptive evidence (1) % immigrants in schools: North-Centre: 11 -15% South-Islands: 3 -4% I focus on North and Centre. 5° grade- Italian scores

Descriptive evidence (2) AREA North. West North. East Centre MEAN SCORES OF N 2 G 1 G 5 TH GRADE ITALIA MATH N -0. 14 -0. 11 -0. 12 -0. 14 -0. 08 -0. 15 -0. 09 -0. 11 -0. 08 -0. 06 -0. 08 -0. 05 -0. 11 -0. 16 -0. 08 -0. 07 6 TH GRADE ITALIAN MATH -0. 32 -0. 20 -0. 21 -0. 15 -0. 20 -0. 04 -0. 13 -0. 20 -0. 26 -0. 15 -0. 13 -0. 15 -0. 20 -0. 05 -0. 13 All negative Almost all highly significant School-level correlations between the % of immigrants and mean scores

Descriptive evidence (3) Area NW NE C 5 TH GRADE SES natives immigrants -0. 17 -0. 11 -0. 24 -0. 11 -0. 18 -0. 11 6 TH GRADE SES natives immigrants -0. 25 -0. 17 -0. 24 -0. 20 -0. 16 All negative and fairly large All highly significant School level correlations between the % of immigrants and SES

Robustness checks pvmig>0. 3 pvmig>0. 1 pvmig>0. 3 pvmig>0. 5 pvescs>0. 1 pvescs>0. 3 pvescs>0. 5 Nstud=155348 Nstud=110908 Nstud=141487 Nstud=78308 Nstud=37523 Nclass=8090 Nclass=5754 Nclass=7425 Nclass=4121 Nclass=1967 Effect of % 1 G on Immig Nat-SES=0 Nat-SES=2 Nat-SES=4 Effect of % 2 G on Immig Nat-SES=0 Nat-SES=2 Nat-SES=4 Effect of mean ESCS pvmig>0. 5 -0. 030* -0. 015 0. 000 -0. 008 -0. 005 -0. 010 -0. 012 -0. 046** 0. 005 0. 056** 0. 009*** -0. 045* 0. 014 0. 072** 0. 009** -0. 021 -0. 072*** -0. 002 0. 067*** 0. 005 -0. 064** -0. 007 0. 076** 0. 004 -0. 084* 0. 012 0. 109** -0. 006 Example. 6° grade math

Robustness checks The results shown are based on schools passing randomness allocation tests with respect to: IB and SES : level a=0. 10 Other subsets IB: level a=0. 30, a=0. 50 IB and SES : level a=0. 30, a=0. 50 Results Ammermueller-Pischke (2009): - peer effects understimated with measurement error - SES affected by substantial m. e. Underestimation of SES peer effects likely to yield to overestimation of IB peer effects No major substantive changes on immigrant background peer effects Relevant changes on peer SES effects: positive but not significant if IB and SES tests positive and significant if only IB test Hanushek et al(2003): When historical family background and school inputs are omitted peer effects are overestimated