Goals and Gaps: Educational Careers of Immigrant Children (2022) PDF

Econometrica, Vol. 90, No. 1 (January, 2022), 1–29 GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN MICHELA CARLANA Harvard Kennedy School, LEAP, IZA, and CEPR...

Econometrica, Vol. 90, No. 1 (January, 2022), 1–29 GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN MICHELA CARLANA Harvard Kennedy School, LEAP, IZA, and CEPR ELIANA LA FERRARA Bocconi University, IGIER, LEAP, and CEPR PAOLO PINOTTI Bocconi University, DONDENA, LEAP, CReAM, and CEPR We study the educational choices of children of immigrants in a tracked school sys- tem. We first show that immigrants in Italy enroll disproportionately into vocational high schools, as opposed to technical and academically-oriented ones, compared to na- tives of similar ability. The gap is greater for male students and it mirrors an analogous differential in grade retention. We then estimate the impact of a large-scale, random- ized intervention providing tutoring and career counseling to high-ability immigrant students. Male treated students increase their probability of enrolling into the high track to the same level of natives, also closing the gap in grade retention. There are no significant effects on immigrant girls, who exhibit similar choices and performance as native ones in absence of the intervention. Increases in academic motivation and changes in teachers’ recommendation regarding high school choice explain a sizable portion of the effect. Finally, we find positive spillovers on immigrant classmates of treated students, while there is no effect on native classmates. KEYWORDS: Tracking, career choice, immigrants, aspirations, mentoring. 1. INTRODUCTION MIGRANT FLOWS have grown considerably over the past decades, increasingly involv- ing families with children. In 2015, almost one in four 15-year-old students in OECD countries had an immigrant background (OECD (2018)). This has profoundly changed the challenges that schooling systems face in order to ensure skill development in a di- Michela Carlana: [email protected] Eliana La Ferrara: [email protected] Paolo Pinotti: [email protected] The present paper formed the basis of the Walras–Bowley Lecture given by Eliana La Ferrara at the North American Meeting held at University of Washington in Seattle in June 2019. We thank three anonymous referees, Josh Angrist, David Autor, Erich Battistin, Pietro Biroli, Daniele Checchi, Kaivan Munshi, Imran Rasul, Enrico Rettore, Miguel Urquiola and seminar participants at several universities for helpful comments. We are particularly grateful to Gianpaolo Barbetta, Paolo Canino, Stefano Cima, and Andrea Trisoglio for continuous support during the project. We thank Gianna Barbieri and Lucia De Fabrizio from MIUR and Patrizia Falzetti and Michele Cardone from INVALSI for giving us access to the administrative data used in this paper. We are grateful to the schools that took part into the intervention for their collaboration. Serena Cocciolo, Rosa De Vivo, and Cristina Clerici provided excellent research assistance. This project was partly funded by Fondazione CARIPLO, Compagnia di San Paolo, and Fondazione Cassa di Risparmio di Padova e Rovigo. Carlana acknowledges financial support from the “Policy Design and Evaluation Research in Devel- oping Countries” Initial Training Network (PODER), which is financed under the Marie Curie Actions of the EU’s Seventh Framework Programme (Contract 608109). La Ferrara acknowledges financial support from the ERC Advanced Grant “Aspirations, Social Norms and Development” (ASNODEV, Contract 694882). AER RCT Registry ID: AEARCTR-0002148. © 2022 The Authors. Econometrica published by John Wiley & Sons Ltd on behalf of The Econometric Society. Eliana La Ferrara is the corresponding author on this paper. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made. 2 M. CARLANA, E. LA FERRARA, AND P. PINOTTI verse student population and promote social cohesion. The ethnic gap in achievement test scores and socio-emotional abilities increases substantially during childhood (Fryer and Levitt (2004), Cunha and Heckman (2007)). The problem is exacerbated in schooling systems characterized by stratification in high school tracks, as early tracking may lead to the educational segregation of children from disadvantaged backgrounds in schools of lower quality. This could ultimately have long-term effects on the skills and occupa- tional careers of children from immigrant families, reducing social mobility and creat- ing unequal opportunities (Guyon, Maurin, and McNally (2012), Brunello and Checchi (2007)). This paper documents the extent of educational segregation of immigrant students and evaluates the effectiveness of an innovative program aimed at steering high-achieving im- migrants towards high schools that fit their academic potential. Throughout the paper, we use the term “immigrant” to indicate children without Italian citizenship: around half of them are first-generation immigrants (i.e., they are born outside Italy) and the other half are second-generation immigrants (i.e., they are born in Italy from non-native parents). We perform our analysis in the context of Italy, where the schooling system is character- ized by tracking in the transition from middle to high school. While uncommon in the Anglo–Saxon world, this type of stratification is the norm among OECD countries, with the age of selection varying from 10 to 16 and an average of three high school tracks per country (OECD (2013)). We start by documenting that immigrants are more likely to enroll in vocational over technical or academic-oriented curricula, relative to native students with similar ability, as measured by a standardized test administered at the beginning of middle school. We denote this phenomenon as “educational segregation.” The gap in track choice for boys persists along the entire distribution of ability, while for girls it is found at the low end of the distribution but not at the high end. The presence of a bigger gap for boys, on aver- age, is consistent with evidence from various countries that boys increasingly lag behind in educational attainment and that the female-male educational advantage is larger for low-socioeconomic status (SES) families (Chetty, Hendren, Lin, Majerovitz, and Scud- eri (2016), Autor, Figlio, Karbownik, Roth, and Wasserman (2019), Bertrand and Pan (2013)). We then estimate the impact of an innovative program called “Equality of Opportunity for Immigrant Students” (EOP henceforth) that provided tutoring and career counsel- ing to immigrant children displaying high academic potential. The curriculum of EOP included a number of meetings that helped students reflect on their aspirations and their potential through a series of psychological exercises based on Social Cognitive Career Theory (Lent, Brown, and Hackett (1994)). We evaluate the effects of EOP leveraging random assignment of the program across schools as well as uni data on students’ careers, cognitive and soft skills, and parental background. The program was offered in 70 middle schools (grades 6 to 8) randomly chosen from a sample of 145 in Northern Italy. We selected the 10 immigrant students with the highest standardized test scores in grade 6: in treatment schools these students were invited to participate in EOP through grades 7 and 8; in control schools they were not offered such program.1 1 For the sake of exposition, in the rest of the paper we refer to these two groups as “treated” and “con- trol,” though we really mean “assigned to treatment” and “assigned to control.” In other words, we present intention-to-treat estimates. We assess compliance with treatment assignment in Appendix B of the Supple- mental Material (Carlana, La Ferrara, and Pinotti (2022)).  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 3 We find that EOP was remarkably successful in reducing educational segregation. Treated males have a 44 percent lower probability to be retained and a 12 percent higher probability of attending an academic or technical high school (as opposed to a vocational one) relative to males in the control group. Indeed, treated immigrant males chose de- manding tracks in the same proportion as native males of comparable ability. In other words, EOP completely closed the immigrant-native track choice gap at the end of grade 8. The effects are in the same direction but smaller and not significant for girls (for whom no educational segregation had been identified in the first place). Although multiple fac- tors could explain the difference in impact across genders, it is suggestive that EOP in- creases immigrants’ enrollment into technical or academic schools only when counterfac- tual enrollment rates lie below those of comparable native students. To shed light on mechanisms, we collected first-hand data on psychological traits. We find that male treated students have higher aspirations and more confidence in their own abilities, and they perceive that environmental barriers will play a smaller role in their future choices, relative to the control group. They also display an improvement in cogni- tive skills, as measured by their standardized test score at the end of grade 8. For girls we detect a reduction in perceived barriers but no relevant effect on aspirations or cognitive skills. Program effects seem to have been internalized by teachers, who recommend treated boys for more demanding high schools. Importantly, no statistically significant effect is found on teachers’ recommendations for treated girls, suggesting that teachers do not automatically recommend the high track to all students involved in EOP, but selectively update their beliefs based on actual changes in motivation and performance—greater for males, lower for females. We estimate heterogeneous treatment effects using a causal forest following Athey and Imbens (2016) and Wager and Athey (2018) to understand who benefits most from the treatment. We find that, although on average EOP did not have a statistically significant impact on girls’ track choice, all male and female students in the middle of the initial test score distribution and those from a low socioeconomic background benefited from the intervention. This pattern of heterogeneity could be used to target future scaled-up interventions. Following the students in our sample through the first two years of high school, we assess the longer term effects of EOP. We find that treated students did equally well as control ones, despite attending (on average) more demanding high schools. Finally, we find evidence of positive spillovers of the intervention on immigrant classmates of treated students, while there is no effect on native classmates. Our work is related to several strands of literature. The first comprises evaluations of interventions aimed at reducing inequality in educational achievement and opportu- nities. Several interventions have targeted low-achieving students and provided a com- bination of information on school options and mentorship on soft skills. Some of these programs were successful in reducing grade retention and high school dropout rates (e.g., Goux, Gurgand, and Maurin (2017), Martins (2010)); others had zero or negative ef- fects (Rodriguez-Planas (2012)). Our program can be seen as complementary to the ones cited above, as it targets a different population—high-achieving students—with the aim of aligning their potential to more ambitious goals.2 2 Card and Giuliano (2016a) studied an intervention for high-achieving students in 4th grade that has a significant positive effect on minority students but not on whites. Their treatment closes an achievement gap with whites of similar baseline ability and makes minority students more likely to participate in an accelerated 4 M. CARLANA, E. LA FERRARA, AND P. PINOTTI A second strand of literature looks at the role of soft skills (e.g., Heckman, Stixrud, and Urzua (2006), Heckman and Kautz (2012)). Low aspirations and high perceived so- cioeconomic barriers can lead students to choose less demanding educational paths, per- petuating a negative cycle (Dalton, Ghosal, and Mani (2016), Genicot and Ray (2017), Mookherjee, Ray, and Napel (2010)). Recent contributions show that students from dis- advantaged backgrounds often lack ambition and suffer from negative stereotypes (e.g., Hoxby and Avery (2013), Guyon and Huillery (2016)). Our paper shows that it is possi- ble to modify aspirations and soft skills through a program that combines academically relevant information with psychological tools. Finally, our work speaks to the literature on tracking within education systems.3 For example, some authors found that postponing school stratification or increasing the pro- portion of seats in academic tracks leads to improved educational outcomes (Malamud and Pop-Eleches (2011), Guyon, Maurin, and McNally (2012)). Our goal is not to assess what would be the effects of postponing or modifying high school tracking, given that tracking is held constant in our experiment. However, one interpretation for the gap in choice that we see between immigrant boys and native students with similar performance is that there is a mismatch. Under that interpretation, our findings suggest that—within an existing tracking system—it is possible to align ability and high school choice for a population that may be particularly misinformed and disadvantaged: immigrant students. The remainder of the paper is organized as follows. Section 2 describes the institutional setting and provides evidence of the educational segregation of immigrant children in Italy. Section 3 illustrates the components of the intervention and our evaluation design. Section 4 describes the data and Section 5 the results. Finally, Section 6 concludes by discussing the policy implications of our findings. 2. INSTITUTIONAL BACKGROUND 2.1. Immigrants in Italian Schools Immigration is a relatively recent phenomenon in Italy. The number of (legal) foreign residents increased from 781,000 to 5 million between 1990 and 2015—14 and 83 per- cent of total residents, respectively. The majority of immigrants in Italy come from low- and middle-income countries, and are characterized on average by lower socioeconomic background than native households.4 At the beginning of the 2016–2017 school year, immigrant children represented 108 percent of students in primary school, 97 percent in middle school, and 72 percent in math track in middle school. Different types of interventions have targeted students’ inaccurate beliefs about their chances of succeeding in competitive academic environments. Bobba and Frisancho (2016) provided students with feedback on their performance in a (mock) exam for admission to selective high schools in Mexico, while Goodman (2016) estimated the effect of mandated exams on selective college admission in the United States. 3 For a review, see Betts (2011). Differently from systems where tracking takes the form of sorting higher- ability students within the same type of education into specialized instruction (e.g., gifted programs, Card and Giuliano (2016a)) or into magnet schools (e.g., Pop-Eleches and Urquiola (2013)), our context is one where tracking involves sorting into high schools with very different curricula and differential access to college. Brunello and Checchi (2007) showed that parental background has stronger effects on labor market outcomes when tracking starts earlier. 4 Figure A.1 in the Supplemental Material Appendix shows the number of immigrants by nationality in 2015 (first 20 nationalities). Figure A.2 compares the income distribution across immigrant and native families, respectively.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 5 high school (see Appendix Figure A.3). The decline in immigrants’ presence from middle to high school reflects higher dropout rates of immigrants in later grades compared to natives. Importantly, the share of immigrant students also differs between different types of high schools. Immigrants represent 125 percent of the student population in vocational schools, 85 percent in technical schools, and only 41 percent in academic schools. As we detail in the next section, these three types of high schools offer very different educational and employment opportunities. 2.2. Secondary Education in Italy Italian pupils normally enter formal schooling the year they turn 6 and the compulsory schooling age is 16. Pre-university education comprises five grades in elementary school, three grades in middle school, and five grades in high school. Within each type of school (elementary, middle, and high school), pupils take all their subjects within the same class and with the same set of peers. In elementary and middle school, they typically also re- tain the same teacher for a given subject throughout the relevant education cycle (e.g., a pupil’s maths and science teacher will be the same throughout middle school). At the end of middle school, students must choose among three different types of high schools: vocational schools (istituto professionale and formazione professionale), technical schools (istituto tecnico), and academically-oriented schools (liceo). Students are free to enroll in any track, as there is no tracking by ability. The three tracks have the same duration, 5 years, but differ widely in terms of cur- riculum, difficulty, and prestige.5 Vocational schools focus on practical training in spe- cialized manual, low-skilled jobs (e.g., plumber or hairdresser), and are meant to prepare students for immediate employment at the end of high school. Technical and academic schools offer instead a comprehensive curriculum in math, humanities, and science. In principle, academic schools are primarily intended for students who want to pursue a uni- versity degree, whereas technical schools complement theory with practical training in specific non-manual jobs (e.g., accountant or graphic designer). Enrollment in college is always possible from the academic-oriented and the technical tracks; in principle, it is also possible from some schools within the vocational track, although very few students who attended vocational education typically obtain further education.6 In practice, academic and technical schools offer much better educational and employment prospects than vo- cational schools. Therefore, in what follows, we refer to vocational schools as the “low track,” and we group technical and academic schools together into the “high track.” We also run some robustness checks on the three tracks separately. Appendix Table A.I of the Supplemental Material (Carlana, La Ferrara, and Pinotti (2022)) compares average outcomes by track four years after graduation, separately by gender and for native versus immigrant students.7 Panel A shows that only 145 percent of Italians graduated from vocational schools, with no relevant differences by gender. They exhibit a much lower probability of pursuing tertiary education compared to high-track 5 Regarding duration, the only exception to the 5-year rule is a sub-track of the vocational track (formazione professionale) that lasts 3 years. 6 Enrollment in college is not possible for students in the sub-track of the vocational track that lasts 3 years (formazione professionale). In the 2015–2016 school year, less than 4 percent of students enrolled in college had attended a vocational track. 7 The source of these data is the “Survey on Educational and Professional Paths of Upper Secondary School Graduates,” conducted in 2015 by the Italian National Statistical Institute (ISTAT) on a representative sample of about 26,000 students graduating from high school in 2011. 6 M. CARLANA, E. LA FERRARA, AND P. PINOTTI graduates: 205 percent, as opposed to 704 percent. College dropout rates in university also differ dramatically between the two groups, at 306 and 118 percent, respectively. In light of these figures, employment rates and salaries four years after graduation are not really informative about labor market prospects, as only a selected group of high-track graduates has already entered the labor market. However, we can compare the share of those “Not in Education, Employment, or Training”(NEET), which reaches 29 percent among low-track graduates, 10 percentage points higher than among other graduates. This is particularly surprising, as in principle vocational schools should prepare students for immediate employment. Finally, graduates from vocational schools also have a higher probability—about one third—of regretting their choice, as elicited in the ISTAT survey. These figures are fairly similar by gender. Panel B of Appendix Table A.I shows comparable statistics for immigrant students. Conditional on completing the same high school track, educational and occupational out- comes are remarkably similar to those experienced by natives. Also among immigrants, graduates from vocational schools exhibit lower enrollment into (and higher dropout rates from) tertiary education, as well as a higher prevalence of NEETs. A stark difference emerges, however, when we compare high school choice: 37 percent of immigrants (42 percent among males) graduate from the low track, compared to the aforementioned 145 (156 for males) of natives. The fact that the share of NEETs is not lower among im- migrants graduating from the low track helps dispel a commonly held belief, that is, that vocational tracks may be better for immigrants who may be seeking immediate employ- ment. In light of the (worse) outcomes experienced on average by low-track graduates, the over-representation of immigrants in this group raises concerns about immigrants’ future career opportunities and, eventually, their prospects for successful integration and up- ward social mobility. The spirit of the intervention we evaluate was to improve the future of talented immigrant students, by making them more aware of a career path that—based on the above data—on average leads to better subjective and objective outcomes. Of course, enrollment rates and outcomes in Appendix Table A.I reflect endogenous sorting by ability and socioeconomic background. Below, we provide a more informative comparison of transitions to high school between immigrants and natives, exploiting a unique data set that matches administrative data on educational careers with standardized test scores and information on parental background. 2.3. Educational Segregation As we discuss in detail in Section 4.1, Italian students take a series of standardized tests of proficiency in reading and math at various points of their careers. These tests are known as INVALSI, from the name of the agency that administers them. The tests are identical for all students in a given grade and are blindly scored, so results are fully comparable across schools. Throughout the paper, we use the standardized test score obtained in grade 6 (INVALSI6) as a measure of students’ ability at the beginning of middle school. Using a unique data set that matches the above scores with students’ educational ca- reers, we can compare the average probability of enrolling into the high track for native and immigrant students conditioning on their initial ability.8 Figure 1 plots the probability 8 As we explain in Section 4, this data set was constructed for the evaluation of the EOP program, which took place in five large cities of Northern Italy. The data used in Figure 1 are those from schools in the control group, which were unaffected by our intervention.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 7 FIGURE 1.—Probability of enrolling in the high track at the end of middle school, by quintile of standardized test score in grade 6 (INVALS6). of enrolling in a “high track” (academic or technical) by quintile of INVALSI6, separately for male and female students (left and right panel, respectively).9 Not surprisingly, the probability of choosing the high track is increasing in INVALSI6 for all groups. However, such probability remains significantly lower for immigrant males than for native ones. The gap is larger in the upper part of the ability distribution, reach- ing 16 percentage points in the top quintile. By contrast, the gap between immigrant and native females is much smaller and it is negligible in the upper part of the ability distri- bution. Appendix Table A.II shows that the gap in educational choices between native and immigrant males persists when conditioning in addition on family background, as measured by parents’ education, employment status, and occupation. This suggests that migration status plays a role, on top of general socioeconomic disadvantage. Why is the native-immigrant gap larger for boys than for girls? The gender gap we uncover is consistent with a recent literature that has documented significant gender dif- ferences in the effect of family background on child development and outcomes dur- ing adulthood. Unlike girls, boys’ development and behavior appear to be extremely re- sponsive to worse parental inputs. Low-income boys have lower performance in stan- dardized test scores, lower high school graduation rates, and more disruptive behavior than their sisters—or girls who grew up in similar contexts (see, among others, Autor et al. (2019), Almås, Cappelen, Salvanes, Sørensen, and Tungodden (2016), Bertrand and Pan (2013), DiPrete and Jennings (2012), Conger and Long (2010), Jackson and Moore (2008)). These effects translate into long-term negative consequences: minority boys are less likely to attend university, work disproportionally less, and have higher crime rates 9 Appendix Figure A.4 reports corresponding graphs for other outcomes that we consider in our analysis, namely: grade retention, teachers’ recommendations, and test scores at the end of grade 8 (INVALSI8). 8 M. CARLANA, E. LA FERRARA, AND P. PINOTTI than girls (Chetty, Hendren, Jones, and Porter (2018), Chetty et al. (2016), Brenøe and Lundberg (2019)). An additional reason behind the gender gap in favor of girls among immigrant children may be related to gender roles and norms imposing social control of daughters and more lax regulations for sons (Lopez (2003)), which may lead to better outcomes for high-achieving immigrant girls compared to boys. Although the focus of our paper is not on the gender gap among minority students— but, rather, on the gap between natives and immigrants—the fact that this gap differs between boys and girls will help us to understand the differential impact of our interven- tion across genders. 3. THE INTERVENTION The intervention we evaluate was developed in collaboration with the Italian Ministry of Education and three bank foundations.10 The program was called “Equality of Oppor- tunity for Immigrant Students” (EOP) and aimed at aligning the goals and aspirations of high-achieving immigrant students with their ability, in order to favor congruous educa- tional choices at the end of middle school. The intervention took place during the last two years of middle school (grades 7 and 8) and was administered in a randomized fashion in five large cities of Northern Italy: Milan, Turin, Genoa, Brescia, and Padua. The first dimension of targeting involved the definition of the school sample: schools were eligible to receive the program if they had at least 20 immigrant students, where ‘immigrant’ was defined as being a citizen of a country with lower GDP than Italy.11 In the five cities, there were 145 such schools: 70 were randomized into the treatment group and 75 into control. To enhance comparability between the two groups of schools, we stratified randomization by province and school size. The second step was the definition of the target students. Because the goal of EOP was to reduce mismatch in track choice for high-achieving immigrants, within each school we defined as ‘high achievers’ the 10 immigrant students with the highest standardized test score in grade 6 (INVALSI6). In the treatment schools, these 10 students took part in EOP, while in the control schools, they did not. In both sets of schools, these top 10 im- migrant students were surveyed and their academic performance and school choices were followed through administrative records. In our empirical analysis, we will thus compare outcomes between the 10 immigrant students with the highest INVALSI6 scores in treated and control schools. The EOP program consisted of two components: (i) a psychologically grounded career choice consultancy, and (ii) an academic tutoring one. The career choice consultancy was developed based on Social Cognitive Career Theory, a paradigm that views career devel- opment as a choice subject to contextual influences and constraints (Lent, Brown, and Hackett (1994)). Under this view, goals and self-efficacy are as important as cognitive skills in shaping individual careers.12 Specifically, “persons with adequate skills but weak 10 The program was financed by three philanthropic institutions operating in Northern Italy, namely Fon- dazione CARIPLO, Compagnia di San Paolo, and Fondazione Cassa di Risparmio di Padova e Rovigo. 11 This excludes from the program immigrant children from high-income European countries, for whom no educational segregation exists. When we use the term “immigrant,” we include both first-generation im- migrants (i.e., children born outside Italy and without Italian citizenship) and second-generation immigrants (i.e., children born in Italy from parents without Italian citizenship). In Italy, citizenship is mainly acquired through ius sanguinis, that is, if an individual’s parents have Italian citizenship. It is possible to obtain the Italian citizenship when first-generation immigrants turn 18 years old. 12 Self-efficacy can be defined as people’s beliefs about their capabilities “to organize and execute courses of action required to attain designated types of performances”; see Bandura (1986), p. 391.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 9 FIGURE 2.—Time line. self-efficacy beliefs in a particular performance domain may prematurely rule out that do- main from further occupational or academic choice consideration” (Brown (2002)). The goal of EOP was to help high-achieving immigrant students to identify educational and occupational goals congruous with their talents. The approach was not to unconditionally push students towards high tracks, but to make them aware of existing opportunities and of their own skills, so they could make more informed choices. The protocol of the career choice consultancy involved a total of 13 meetings, and the program guidelines required participants to attend at least 75 percent of the meetings. All meetings were administered by career counselors with graduate degrees in psychology and significant experience in career choice guidance for secondary school, especially with immigrant children. Some of the meetings were one-to-one, while others were in groups. Students had five (one-to-one) meetings with a counselor and worked on tasks that prompted them to reflect on their goals, the personal resources needed to achieve such goals, and whether they already had or they needed to develop such resources. Exam- ples of the tasks that students worked on include: (i) “Think about your past life, indicate five study experiences and five other experiences that you have completed successfully. Consider now such experiences one by one and briefly indicate where and with whom it happened, what you did and which personal resources helped you doing well in that thing—your knowledge, skills, personality traits, motivations and everything you believe it was important to have”; (ii) “Choose a number of professions that interest you. For each of them, indicate which resources are needed (knowledge, skills, personality traits, motivations,... ) and divide them into ‘I have it’ and ‘I need to develop it’;” (iii) “List the results you would like to achieve with your job, from the most to the least important.” Five other meetings were held in groups, where counselors provided students with infor- mation about the Italian education system and showed videos with success stories of older immigrant students. Two further meetings, respectively at the beginning and at the end of the intervention, were intended for parents. In the first meeting counselors described the content of the program, while in the second they shared with parents aspirations and barriers perceived by students. Parents also received a brochure, translated into their mother tongue, sum- marizing the main options for secondary education in Italy.13 Finally, towards the end of the intervention, career counselors met with teachers and discussed the educational path and high school track chosen by the students involved in the intervention. Figure 2 shows the timeline of the intervention and the realization of the main out- comes of interest. In grade 6, we selected the eligible students on the basis of the stan- dardized test score they got that year. EOP meetings started at the beginning of grade 13 While it would be very interesting to disentangle the role that parents play in high school choice (e.g., Dustmann (2004), Giustinelli (2016)), it was not possible to survey parents within our experiment. 10 M. CARLANA, E. LA FERRARA, AND P. PINOTTI 7 and continued through grade 8, until the month of March. By the month of January of grade 8, students receive a formal ‘recommendation’ by their teachers about the high school track that teachers deem most appropriate for them. This recommendation is not binding, but it serves as a signal to the students and their families. In February, all stu- dents have to pre-register for the high school they wish to attend through a web portal of the Ministry of Education. This choice can later be modified (though this is not very common), so we use the high school track in which students actually enroll at the end of grade 8 as our variable of interest. The second component of the intervention, the academic tutoring part, was motivated by the concern that participants who enrolled in the high track as a result of the pro- gram may subsequently experience difficulties in completing this (more demanding) high school. For this reason, the counseling and career choice module was accompanied by a module on Cognitive Academic Language Proficiency (CALP). CALP was not aimed at improving students’ knowledge or their cognitive skills but, rather, at teaching them a method for studying several subjects—Italian grammar, geography, algebra, and geome- try. Since the main motivation of the CALP module was to decrease the risk of subsequent failure for immigrant students enrolling in demanding tracks, students with a lower IN- VALSI6 were offered a higher number of CALP meetings. Specifically, students scoring below 65 out of 100 in INVALSI6 were invited to 29 meetings (55 hours tutoring). This group constituted 66 percent of all treated students. Students scoring between 65 and 80 (30 percent of the treatment group) were invited to 17 meetings (32 hours tutoring); fi- nally, students scoring above 80 (4 percent of the treatment group) were not invited to CALP sessions. Due to ethical concerns, we could not implement a fully factorial design, as our sponsors were concerned that encouraging students to pursue ambitious goals without endowing them with the tools for succeeding in demanding high schools could have created a risk of harm. However, we can assess the relative effectiveness of career counseling and CALP at the intensive margin by exploiting the different cutoffs for the number of CALP meetings and employing a regression discontinuity design (RDD). We do this in Section 5.3. 4. DATA We build a unique data set including survey data and administrative information from the Italian Ministry of Education (MIUR) and the Institute for the Evaluation of the Ital- ian Schooling System (INVALSI), an independent public agency that monitors students’ performance. 4.1. School Choice and Academic Performance Administrative data from school registry. From the MIUR administrative registry we take the following variables that we use as outcomes for each student: (i) track choice at the end of grade 8; (ii) track recommended by teachers halfway through grade 8; (iii) grade retention for all grades between 6 and 9 (included); (iv) number of retakes after the sum- mer of grade 9;14 and (v) teachers’ assessment of student’s behavioral conduct during grades 8 and 9. The registry also contains some information on students’ background, 14 While through middle school pupils are either admitted to the next grade or retained, in high school they can also be admitted to the next grade conditional on retaking (and passing) after the summer an examination in one or more subjects in which they were deficient during the year.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 11 in particular: citizenship, country of origin, date of birth, and school and class attended throughout their careers. INVALSI tests. Since 2010, INVALSI administers standardized reading and math profi- ciency tests to all students at the end of grades 2, 5, 6, 8, and 10. Such tests resemble those administered by the OECD Programme for International Student Assessment (PISA) to representative samples of 15-year-old students. They consist of a series of questions in- cluding multiple choice as well as open-ended questions, the exact structure of the test varying by grade.15 Importantly, the test is identical for all students in a given grade, it is administered to everyone on the same day (towards the end of school year), and it is blindly scored, so results are fully comparable across schools in Italy. This is crucial for the purposes of our analysis, because it allows us to compare the educational choices of immigrant and native students holding constant their academic proficiency. We use two test scores for the cohort of students who were in grade 6 in 2012. The first is the standardized test score for grade 6 (INVALSI6), which we include in all specifications to control for students’ initial ability. The second is the test score for grade 8 (INVALSI8), which is one of our outcomes. 4.2. Soft Skills We complement the above data set with original survey data on soft skills collected at the end of grade 8. The goal of this survey is to allow us to better understand what mech- anisms shape career-related interests and high school track choice. The survey was ad- ministered to all treated students and to a random 50 percent sample of control schools.16 The questionnaire was developed with a team of psychologists and includes three main sections: (i) Goals. This section comprises both educational (e.g., university degree, diploma, or less) and occupational targets (e.g., blue collar, white collar, managerial, or en- trepreneurial jobs) that the student aims to achieve. For example, questions in this mod- ule include: “Thinking about your future, education-wise, what objectives do you intend to achieve?” (Options: Work as soon as I finish this school; Study until age 16 then work; Study until age 18 and get a high school diploma; Enroll in university); and “Thinking about your future, work-wise, what objectives do you intend to achieve?” (Options: Blue collar job; White collar job; Managerial job; Entrepreneur). (ii) Self-efficacy. This section includes a student’s own assessment of the extent to which he or she possesses the skills and resources required to achieve the goals stated above, as well as broad notions of self-esteem. Questions in this module are worded as: “Indepen- dently of your educational aim but thinking about your abilities, do you think you could get a... (university degree/ white collar job/ managerial job, as worded above)?”. Answers in this module are on a scale of 1 to 4, with 1 meaning ‘not at all’, and 4 meaning ‘very much’. (iii) Barriers. A series of questions elicit students’ perceptions of environmental barriers, be they related to economic constraints, racial prejudice, or family preferences that differ from a student’s own plans. Typical questions are worded as: “Do you think the following 15 In general, math questions are related to calculus, geometry, probability, and algebra, while reading ques- tions are related to text comprehension and grammar. 16 We chose not to administer the survey to half of the control schools because we wanted to be able to test if filling in a questionnaire on goals and perceived barriers may constitute a ‘treatment’ in itself. In Appendix Table A.V, we show that students in control schools involved and not involved in the soft skills questionnaire do not systematically differ in terms of high school choice, grade retention, and test scores in grade 8. 12 M. CARLANA, E. LA FERRARA, AND P. PINOTTI barriers could be an obstacle in the achievement of your educational aims?” (Options: Economic resources; The needs and ideas of your family; Racial prejudice; Family plans related to children or marriage; Not feeling up to the standards). Answers in this module also range from 1 to 4. Following Thompson (2004), we summarize the individual variables described above into interpretable aggregates using first exploratory factor analysis and then confirmatory factor analysis. This method extracts latent factors from subsets of psychological measures by maximizing (minimizing) the correlation across measures within (between) subsets. The measures associated with each factor and their respective loadings are reported in Appendix Table A.VI. 4.3. Sample and Randomization Check Our working sample at the inception of EOP comprises 1217 students: 597 in treated schools and 620 in control ones.17 Table I reports average characteristics of the treatment and control group at the start of our intervention. We distinguish between individual student characteristics (Panel A) and family background (Panel B). Half of the students in our sample are girls, 56 percent are first-generation immigrants, and 26 percent were born before 1999 (the typical birth year of the cohort in our study). The mean of the standardized test score INVALSI6 is 609 in the treatment group and 607 in the control group (not significantly different). To get a more complete picture, Figure 3 plots the distribution of INVALSI6 across three groups of students in our 145 schools: native students, all immigrant students, and the 10 immigrant students with the highest score (i.e., our treated and control groups). Although immigrants generally exhibit lower schooling performance than native students, the top 10 immigrant students in each school are comparable to natives in the medium- upper part of the distribution. INVALSI also provides information on parents’ education and occupation which, how- ever, is missing for more than a third of our sample. In Appendix Table A.III, we show that the share of records with missing information on either or both parents’ education correlates negatively with socioeconomic conditions across Census tracts. This finding suggests that missing information on parents background may be a proxy for low socioe- conomic status. When information on education is provided, 22 (20) percent of mothers (fathers) have not completed high school, 29 (30) percent have a high school diploma, and 12 (10) percent have post-secondary education. Unemployment rates are about 6 percent for mothers and 6 percent for fathers. The share of mothers only working at home is 26 percent, while that of fathers is negligible. Among those working outside the home, 24 (36) percent of mothers (fathers) have a blue collar job, and 12 (20) percent a white collar job. Importantly, none of the student or family characteristics differ significantly between the treatment and control sample, indicating that our randomization was successful.18 17 By construction, the sample should have comprised 1451 students, that is, 10 × (70 + 75) = 1450, but in one school the 10th and 11th students obtained the same INVALSI6 score and were both eligible for the program. For 70 students, it was impossible to match the MIUR and INVALSI identifiers, which reduced the sample to 1381 students. In addition, some students were retained in grade 6, moved to another school, or dropped out between the moment when they took the INVALSI6 test and the beginning of grade 7, leading to the above sample of 1217. Appendix Table A.VII shows that missingness is not selective across treatment and control schools. 18 Children in the treatment and control groups are from 72 different nationalities, out of which the most represented countries are Romania, Philippines, Albania, Morocco, Peru, Ecuador, and China. The balance table for these characteristics is reported in Appendix Table A.VIII.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 13 TABLE I TREATED AND CONTROL STUDENTS, BALANCE TEST. Full Sample Treated Controls Difference p-value Std. Difference Panel A: Student characteristics Female 0506 0508 0505 0003 [0.93] −0006 Test score in grade 6 (INVALSI6) 6082 6093 6071 0224 [0.86] 0005 First-generation immigrant 0555 0547 0561 −0014 [0.73] 0028 Born before 1999 0257 0242 0273 −0031 [0.25] 0071 Brescia 0179 0165 0194 −0029 [0.67] 0076 Genova 0067 0074 006 0014 [0.75] −0056 Milan 0496 0476 0516 −004 [0.65] 0080 Padua 0055 0064 0047 0017 [0.68] −0074 Turin 0203 0223 0184 0039 [0.58] −0097 Panel B: Family characteristics Mother Less than high school 0224 0228 0221 0007 [0.87] −0017 High school 0464 0315 0281 0034 [0.36] −0074 Some post-secondary 0119 0112 0126 −0014 [0.63] 0043 education Missing education 0359 0345 0373 −0028 [0.65] 0058 Blue collar 0237 0243 0231 0012 [0.76] −0028 White collar 0124 0124 0124 −0000 [1.00] 0000 Unemployed 0055 0062 0048 0014 [0.59] −0061 At home 0259 0256 0261 −0005 [0.91] 0011 Missing occupation 0325 0314 0335 −0021 [0.72] 0045 Father Less than high school 0197 0198 0197 0001 [0.98] −0003 High school 0296 0301 0290 0011 [0.79] −0024 Some post-secondary 0099 0099 010 −0001 [0.98] 0003 education Missing education 0407 0402 0413 −0011 [0.86] 0022 Blue collar 0357 0351 0361 −001 [0.80] 0021 White collar 0198 0211 0185 0026 [0.47] −0065 Unemployed 0062 0058 0066 −0008 [0.70] 0033 At home 0012 0012 0013 −0001 [0.93] 0009 Missing occupation 0371 0367 0374 −0007 [0.90] 0014 Note: This table shows the number and characteristics of treated and control students in our sample. p-values for difference in means are reported in square brackets. The last column also reports the standardized difference between group averages. 5. RESULTS In this section, we estimate the impact of EOP on educational choices and grade re- tention, separately for males and females, and we investigate heterogeneous effects using causal forest. In addition, we decompose treatment effects into several mediating factors, estimate impacts on longer-term outcomes, and spillover effects on non-eligible students in treatment schools. 5.1. Educational Choices and Grade Retention In Table II, we estimate the impact of EOP on high school track choice. The dependent variable is a dummy equal to 1 for students who choose the high track and 0 for those who choose the low track. The explanatory variable of interest is EOP, an indicator for whether a student attends a school that was (randomly) selected to receive our interven- tion. The coefficient of this dummy should thus be interpreted as the intention-to-treat 14 M. CARLANA, E. LA FERRARA, AND P. PINOTTI FIGURE 3.—Distribution of standardized test score in grade 6 (INVALS6). (ITT) effect. Odd-numbered columns condition on treatment only, while the specifica- tions in even-numbered columns include a squared polynomial in INVALSI6, a dummy for first-generation immigrants, and province fixed effects. In all cases, we cluster stan- dard errors at the school level, the unit of randomization. In column 1, assignment to EOP increases the probability of choosing the high track by 5 percentage points, on a baseline rate of 75 percent in the control group. Given random assignment, this estimate is largely unaffected when controlling for student characteristics and province fixed effects (column 2). This average effect, however, masks important differences by gender. EOP increases males’ enrollment into the high track by 8 to 9 percentage points, up from a baseline rate of 674 percent (columns 3 and 4). This is a 12 to 13 percent increase over the mean. By contrast, there is no effect on female students, who start from a higher baseline enroll- ment rate of 824 percent (columns 5 and 6). Figure 4 compares the educational choices of students eligible for EOP, namely the ten immigrant students obtaining the highest score in INVALSI6 within each school, with the educational choices of other students down in the ranking. The left graph confirms that (eligibility to) EOP increases the probability of choosing the high track for eligible male students compared to non-eligible students near the cutoff for eligibility. The “difference- in-discontinuities” between treated and control schools in the probability of choosing the high track at the eligibility cutoff amounts to 18 percentage points (estimates are reported in Appendix Table A.IV, columns 1–3). By contrast, there is no significant discontinuity in schooling choices between eligible and non-eligible female students near the cutoff.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 15 TABLE II THE EFFECT OF EOP ON EDUCATIONAL CHOICES. Dependent Variable: Choosing the High Track (= 1 if Choose High Track) (1) (2) (3) (4) (5) (6) (7) (8) All Immigrants Male Immigrants Female Immigrants All Immigrants EOP 0051 0043 0091 0080 0011 0009 0091 0077 (0027) (0024) (0037) (0035) (0034) (0031) (0037) (0034) Female X EOP −0080 −0067 (0046) (0043) Constant 0750 0683 0674 0651 0824 0720 0674 0621 (0019) (0038) (0028) (0048) (0022) (0049) (0028) (0041) Mean dep. var. control 0750 0750 0674 0674 0824 0824 0750 0750 Observations 1217 1217 601 601 616 616 1217 1217 Controls No Yes No Yes No Yes No Yes R-squared 0004 0087 0010 0086 0000 0100 0024 0105 Note: This table shows the effect of EOP on immigrant students’ educational choices by the end of middle school. The dependent variable is a dummy equal to 1 for students choosing the high track (academic or technical schools) and equal to zero otherwise. EOP is a dummy equal to 1 for students in schools assigned to the treatment group and equal to zero for schools assigned to the control group. Specifications in columns (2), (4), (6), and (8) control in addition for a squared polynomial in INVALSI6, a dummy equal to 1 for first-generation immigrants, and province fixed effects. Standard errors clustered by school are reported in parentheses. While the above analysis aggregates academic and technical tracks into a single out- come category, in Appendix Table A.IX we maintain the distinction among tracks and investigate the impacts of EOP using an ordered probit model.19 We find that the pro- gram had an effect for boys throughout the distribution of high school track choices, most FIGURE 4.—The effect of EOP on educational choices, difference-in-discontinuities at the eligibility cutoff. 19 In the Italian schooling context, as suggested by Carlana (2019) and supported in our context by Appendix Figure A.5, high schools can be clearly divided into three groups using the average test score of students in grade 8 who self-select into the different tracks: top-tier schools (scientific and classical academic tracks), 16 M. CARLANA, E. LA FERRARA, AND P. PINOTTI FIGURE 5.—Track choice and grade retention of immigrants and comparable natives. likely by moving students from vocational to the middle track and from the middle track to the top tier schools. The coefficients for girls are very small and never significant. Panel A of Figure 5 compares enrollment in the high track for high-achieving immigrant students randomized into the control group (leftmost bar), treatment group (middle bar), and for a group of native students with comparable ability in the first year of middle school (rightmost bar). Specifically, we match each immigrant student in our sample with one native student of the same gender who obtained an identical score in INVALSI6. By construction, these three groups of students had the same standardized test score in the first year of middle school. The figure shows that, two years later, the immigrant boys who received EOP make similar choices as natives who started off like them, while the un- treated immigrants have a significantly lower probability of choosing the high track. EOP thus prevented the type of educational segregation that we documented in Section 2.3. Interestingly, Figure 5 confirms that immigrant girls make similar choices as native girls even in the absence of intervention. middle-tier schools (including linguistic, artistic, pedagogical academic tracks as well as technical tracks), bottom-tier schools (vocational).  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 17 TABLE III THE EFFECT OF EOP ON GRADE RETENTION. Dependent Variable: Grade Retention (= 1 if Repeat a Grade) (1) (2) (3) (4) (5) (6) (7) (8) All Immigrants Male Immigrants Female Immigrants All Immigrants EOP −0013 −0013 −0037 −0037 0011 0009 −0037 −0035 (0016) (0016) (0021) (0021) (0019) (0019) (0021) (0020) Female X EOP 0048 0044 (0024) (0023) Constant 0056 0073 0085 0097 0029 0048 0085 0097 (0010) (0028) (0016) (0028) (0010) (0043) (0016) (0030) Mean dep. var. control 0056 0056 0085 0085 0029 0029 0056 0056 Observations 1217 1217 601 601 616 616 1217 1217 Controls No Yes No Yes No Yes No Yes R-squared 0001 0017 0006 0022 0001 0023 0009 0024 Note: This table shows the effect of EOP on immigrant students’ grade retention during middle school. The dependent variable is a dummy equal to 1 for students retained in grade 7 or 8, and equal to zero otherwise. EOP is a dummy equal to 1 for students in middle schools assigned to the treatment group and equal to zero for schools assigned to the control group. Specifications in columns (2), (4), (6), and (8) control in addition for a squared polynomial in INVALSI6, a dummy equal to 1 for first-generation immigrants, and province fixed effects. Standard errors clustered by school are reported in parentheses. Taken together, the evidence in Figure 1, Table II, and Figure 5 highlights a remarkable feature of the intervention: EOP influences educational choices only when counterfactual enrollment rates into the high track lie below those of comparable native students. There- fore, EOP seems to align immigrant students’ goals and aspirations to those of native students when there is an initial misalignment, as opposed to just pushing all immigrant students towards the high track. In Section 5.4, we provide direct evidence in this respect. Table III and Panel B of Figure 5 convey similar evidence for grade retention in grade 7 or 8. Grade retention is surprisingly high among male immigrant students in our sample: absent the intervention, it reaches 85 percent, as compared to only 42 percent for native males with a similar INVALSI6 score. This gap disappears in EOP schools, whereas there is neither a significant gap nor an effect for female students. In both respects, the effect on grade retention across genders is very similar to that on high school choice. Since there is one-sided non-compliance with treatment assignment, we can also esti- mate the average treatment-on-the-treated (ATT) effect on the subset of compliers. We do so in Appendix B of the Supplemental Material. In particular, Appendix Table B.I shows that, if we classify as treated all students attending at least 75 percent of the meet- ings (in accordance with the program guidelines discussed in Section 3), the ATT effects on males are a 125 percentage point increase in enrollment in the high track, and a 57 percentage point decrease in grade retention. 5.2. Heterogeneous Effects: Multiple Hypothesis Testing and Causal Forest Given the strong gender differences in educational segregation, we focused our main analysis on the impact of EOP on track choice by gender and showed that the effect is not significantly different from zero for girls. However, estimating heterogeneous effects by splitting the sample may increase the probability of Type I errors. We take a number of steps to ensure that the results are not driven by our ex ante choice of the relevant subgroups. 18 M. CARLANA, E. LA FERRARA, AND P. PINOTTI First, we adjust the p-values following the bootstrap-based procedure, as recommended by List, Shaikh, and Xu (2019).20 The results in Table II retain similar levels of statistical significance, with an adjusted p-value of 0024 for boys and 0723 for girls. Second, in Appendix Table A.X, we examine heterogeneity of the ITT effect along addi- tional dimensions beyond gender, reporting the coefficient for each subgroup in column 1, the p-value in column 2, and the p-value adjusted for multiple hypothesis testing in column 3. Each panel of the table focuses on a different, additional dimension of hetero- geneity (on top of gender): mother’s education in Panel A, terciles of INVALSI6 in Panel B, and EU citizenship in Panel C. While most p-values increase above standard levels once we adjust for multiple hypothesis testing, the pattern of coefficients suggests that boys and girls from low socioeconomic background benefited the most from EOP. Finally, to capture high-dimensional combination of covariates that the researcher- specified interactions may miss, we explore heterogeneity using machine learning tools, following the recent literature on heterogeneous treatment effects (Athey and Imbens (2016), Davis and Heller (2017), Bertrand, Crépon, Marguerie, and Premand (2017)). Appendix C of the Supplemental Material describes our methodology in detail. In a nutshell, we estimate the Conditional Average Treatment Effect (CATE), includ- ing in the causal forest the following baseline characteristics: gender, INVALSI6 (squared polynomial), generation of immigration, mother’s and father’s education and occupation, school province, and region of citizenship. We use the predictions on the expected treat- ment effect for each individual, given the covariates, to investigate treatment heterogene- ity. We divide the sample in two groups, that is, top and bottom half of the predictions. Appendix Table C.I reports the balance test for the CATE and p-values adjusted for multiple hypothesis testing. Figure 6 shows a heatmap of the CATE by gender, mother’s education, and INVALSI6. Overall, the results on gender differences are consistent with our main analysis: girls are over-represented among students with lower CATE. However, additional interesting results emerge. Boys and girls with low or missing levels of mother’s education and stu- dents in the central part of the ability distribution benefit the most from participating in EOP.21 This suggests that targeting only boys in future scaled-up interventions may miss girls that could substantially benefit from EOP, such as girls in the middle of the initial test score distribution and those with parents with low socioeconomic status. 5.3. Mechanisms In this section, we try to disentangle the effects of different components of the program and explore the mechanisms through which they impacted educational choices. Career Counseling versus Academic Tutoring As explained in Section 3, the EOP intervention comprised two components: motiva- tional and career counseling (the Social Cognitive Career Theory module), and academic tutoring to provide students with a method for studying (the CALP module). Ideally, one 20 As shown in List, Shaikh, and Xu (2019), under weak assumptions, this procedure asymptotically controls for Familywise Error Rate (FWER), that is, for the probability of getting one or more false rejections. 21 As shown in Appendix Table A.III and discussed in Section 4.3, missing information on parents’ education is a proxy for low socioeconomic status.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 19 FIGURE 6.—Heatmap of the Conditional Average Treatment Effect by gender, mother education, and IN- VALSI6. would want to disentangle the contribution of each of these two components using a fac- torial treatment design. However, ethical constraints prevented us from delivering the motivational treatment without also delivering help for studying. This means we cannot disentangle the contribution of CALP on the extensive margin, but we can assess the ef- fectiveness of CALP on the intensive margin by comparing students who scored below and above the cutoff used to determine the number of CALP meetings. In particular, while virtually every treated student was invited to at least 17 CALP meetings in grade 8, students who had scored below 65 in the entry level test (INVALSI6) were invited to an additional 12 CALP meetings in grade 7. Achieving a score of 65 does not impact stu- dents’ careers in any other way, so this rule provides a suitable regression discontinuity design.22 The first two graphs in Figure 7 show that students with a score below 65 attended on average 55 more CALP meetings than students with a score above 65, whereas there are no differences in the number of career counseling meetings. Therefore, be- ing invited to additional tutoring meetings increased the ‘dosage’ of CALP. The two graphs at the bottom of Figure 7 show that there is no significant discontinuity in the probability of choosing the high track or in grade retention between students on one side or the other of the cutoff. This implies that neither one of the main out- comes was significantly affected by an increase in the number of CALP meetings.23 Therefore, at the intensive margin, EOP seems to operate mostly through motivational and career counseling, as opposed to specialized help for studying. This, of course, does not exclude that academic tutoring may have played a role at the extensive mar- gin. 22 Appendix Figure A.7 confirms that the density of INVALSI6 does not change discontinuously around the cutoff, using a formal McCrary test. 23 The estimated changes at the cutoff are reported in Panel A of Appendix Table A.XI. 20 M. CARLANA, E. LA FERRARA, AND P. PINOTTI FIGURE 7.—Effect of additional CALP meetings, regression discontinuity estimates. Personality and Cognitive Skills In Table IV, we report the effects of EOP on the cognitive and personality skills de- scribed in Section 4.2.24 The dependent variables include the indexes of aspirations and perceived barriers whose components and loading factors are reported in Appendix Ta- ble A.VI, as well as measures of academic performance. Starting with personality skills, the intervention substantially increased students’ aspi- rations, especially for males (+031 standard deviations), whereas the effect is weaker and not statistically significant for females. EOP also reduced students’ perceptions that their choices would be limited by barriers such as financial constraints, prejudice, or fam- ily plans. The effect is sizable (a reduction in perceived barriers of about 04 standard deviations) and virtually identical between males and females. In the last row of each panel, we also report the p-values adjusted for multiple hypothesis testing (List, Shaikh, and Xu (2019)), considering jointly the four outcomes and the two subgroups by gender. Appendix Table A.XII shows that the effect is generally statistically significant for the 24 The sample differs across columns because questionnaires measuring personality skills were administered only to a 50 percent random sample of control students; see Section 4.2.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 21 TABLE IV THE EFFECT OF EOP ON MEDIATING FACTORS. (1) (2) (3) (4) Aspirations Perception of Barriers Dep. var. Males Females Males Females EOP 0312 0105 −0385 −0422 (0102) (0090) (0121) (0101) Mean dep. −0209 0248 0252 0420 Observations 321 366 321 366 R-squared 0156 0110 0096 0086 MHT p-value 0063 0330 0004 0000 INVALSI8 Teachers’ Recomm. Dep. var. Males Females Males Females EOP 0163 0001 0171 0074 (0051) (0050) (0053) (0056) Mean dep. −0129 0057 0371 0578 Observations 520 574 601 616 R-squared 0472 0423 0166 0103 MHT p-value 0044 0757 0000 0138 Note: This table shows the effect of EOP on several mediating factors. Aspirations and perception of barriers are the two principal components extracted from the psychological measures collected through students’ questionnaires. The individual variables included in each index and their loading factors are reported in Appendix Table A.VI. INVALSI8 is the score obtained in the standardized test at the end of middle school (grade 8). Teachers’ recommendation is a dummy equal to 1 when the teacher recommends to enroll in the high track and equal to zero otherwise. All specifications control for a squared polynomial in INVALSI6, a dummy equal to 1 for first-generation immigrants, and province fixed effects. Standard errors clustered by school are reported in parentheses. In the last row of each panel, we report the p-values adjusted for multiple hypothesis testing, considering jointly the four outcomes and the hetereogeneity by gender. individual psychological measures that enter the two indexes, even after we account for multiple hypothesis testing (last column). Columns 1 and 2 of the bottom panel of Table IV show that EOP increased the stan- dardized test score in grade 8 (INVALSI8) for male students but not for females. The effect is smaller than that on personality traits (+016 standard deviations). Appendix Figure A.8 compares the entire distribution of aspirations, perceived barriers, and IN- VALSI8 across treated (solid red line), control (solid blue line), and native (dashed black line) students who had the same standardized test score in grade 6. The pattern in the fig- ure confirms that EOP caused a shift in the distribution of all three outcomes for males, and only of perceived barriers for females. In the last two columns of the bottom panel of Table IV, we estimate the effect of EOP on teachers’ recommendations. Recall that these recommendations are made about halfway through grade 8, and before students make their high school track choice. We find that, on average, the probability that teachers recommend the high track is 171 per- centage points higher for male immigrant students in EOP schools, on a baseline of 371 percent in control schools. Importantly, teachers do not significantly revise their recom- mendations for female students in treated schools: the point estimate for females is less than half of that for males and is not significantly different from zero. This is an impor- tant result to gain insights into what determines teachers’ recommendations. Teachers do 22 M. CARLANA, E. LA FERRARA, AND P. PINOTTI TABLE V DECOMPOSITION OF THE EFFECT OF EOP ON HIGH SCHOOL CHOICE, MALE STUDENTS. (1) (2) (3) Explained p-Value Explained p-value Explained p-Value Aspirations 00413 [0.0005] 00359 [0.002] 00357 [0.001] Barriers −00095 [0.5756] −0006 [0.722] −00024 [0.882] INVALSI8 0013 [0.103] 00101 [0.178] Teachers’ recommendation 00333 [0.046] Total explained 00318 00429 00767 EOP effect on choosing high track 0091 0091 0091 Note: This table decomposes the effect of EOP between changes in personality skills (aspirations and perception of barriers), increased schooling achievement (as measured by INVALSI8), and teachers’ recommendations. As explained in Appendix D, the decomposition follows the method devised by Heckman, Pinto, and Savelyev (2013). All specifications control in addition for a squared polynomial in INVALSI6, a dummy equal to 1 for first-generation immigrants, and province fixed effects. Bootstrap standard errors clustered at school level generated from 1000 iterations. not mechanically adjust their track recommendations simply because they know that a student is enrolled in EOP If teachers were simply reacting to knowledge about the program, mechanically adjust- ing their track recommendations upwards for all students in EOP, we would find similar effects for boys and girls because they were enrolled in EOP in equal proportions. Instead, we find that teachers only revise their recommendations for boys, consistent with teachers reacting to detectable changes in motivation and test scores (recall that aspirations and cognitive skills only improve for boys).25 Decomposing the Treatment Effect Following Heckman, Pinto, and Savelyev (2013), we decompose the treatment effect on educational choices into experimentally induced changes in the mediating factors listed in Table IV and changes in other (unmeasured) factors. This decomposition method re- quires strong assumptions that do not leverage the experimental variation. In particular, one needs to assume that the observable mediating factors are independent of the unob- servables in the control group. The results in this section should thus be interpreted in light of such assumptions. Details on the methodology can be found in Appendix D of the Supplemental Material; here we report the main findings. Table V shows the decomposition of the effect of EOP on the educational choices of boys.26 Changes in personality skills explain about one third of the overall effect (column 1). However, this effect is entirely driven by aspirations, whereas perceptions of barriers do not seem an important mediating factor. This is consistent with the fact that males and females experience a similar decrease in perceived barriers, but educational choices only change for males. In column 2, we add school performance (measured by the standardized test score in grade 8) as an additional mediating factor, and in column 3, we further add teachers’ recommendation. 25 This does not imply that teachers did not contribute to changes in students’ aspirations and performance. For example, EOP may have led teachers to pay more attention to immigrants who had misaligned aspirations (e.g., boys), and this may have contributed to improving students’ aspirations and test scores. 26 We only present results for males because treatment effects on females are not significantly different from zero.  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 23 We find that aspirations and teachers’ recommendation are the most important fac- tors, jointly explaining about two thirds of the treatment effect on educational choices. The importance of the first channel is in line with experimental evidence showing that raising aspirations through role models increases educational investment (Tanguy, Der- con, Orkin, and Taffesse (2014)). Our result on teachers’ recommendations aligns with the literature showing that teachers’ expectations on students’ performance play a cru- cial role in affecting educational choices (Papageorge, Gershenson, and Kang (2018)).27 We obtain very similar results when we employ the alternative decomposition method by Gelbach (2016); see Appendix Table A.XIV. To sum up, although EOP did improve school performance and reduce perceived bar- riers, these two factors play a less crucial role as a direct channel of influence in track choice of high-achieving immigrants. 5.4. Longer-Term Effects The results presented so far show that EOP increased the probability of enrolling into academic and technical schools (high track) after grade 8. A potential concern is that such schools may prove too demanding for the students who were affected by the treatment—a version of the so-called ‘mismatch hypothesis’ (e.g., Arcidiacono and Lovenheim (2016)). One may worry that immigrant students may face additional constraints compared to Ital- ian students of similar ability; for example, they may be less embedded in social networks that could help with studying challenging subjects, or they may face financial constraints in paying for private tutoring. In Table VI, we address this concern by estimating the effect of (assignment to) EOP on performance during the first two years of high school. We consider four different out- comes: (i) the probability of being admitted to grade 10, the second year of high school (columns 1–2); (ii) the number of make-up exams students need to take during the sum- mer in order to avoid repeating the grade (typically no more than three exams);28 (iii) the probability that a student drops out before completing grade 10 (columns 5–6); and (iv) the probability of changing school between grade 9 and 11 (columns 7–8). For all these outcomes, we find that treated students are no more likely to experience difficulties compared to the control group. If anything, they are more likely to be admitted to grade 10, less likely to fail courses, and less likely to drop out, although the estimated coefficients are not significantly different from zero.29 It is worth stressing that the lack of a significant effect should not be seen as a shortcoming: given that our treated students were more likely to enroll in demanding high schools, the fact that they are doing as well as the control group (and if anything, better, given the pattern of coefficients in Table VI) is actually a positive result. 27 This phenomenon has been documented since the seminal work of Rosenthal and Jacobson (1968), whereby high and low expectations bring improvements and worsening in performance, respectively. Card and Giuliano (2016b) showed that teachers tend to overlook the potential of high-ability minority students, especially English language learners. Relatedly, children of minority groups benefit the most from positive ex- pectations of teachers and may be negatively affected by teacher stereotypes (Carlana (2019), Alesina, Carlana, La Ferrara, and Pinotti (2018)). 28 The sample in columns 3–4 is smaller than in columns 1–2 because it only comprises students who did not fail outright and who did not drop out of high school. If a student is below the Pass level in too many subjects, the decision of the school is typically to fail this student rather than give make-up exams after the summer. 29 Ideally, we would also like to estimate the longer-term effects on cognitive skills, as measured by the standardized test score in grade 10 (INVALSI10). Unfortunately, it is not possible to match MIUR registries for middle school students with their INVALSI tests in high school. 24 M. CARLANA, E. LA FERRARA, AND P. PINOTTI TABLE VI EFFECT OF EOP ON LONG-TERM OUTCOMES. (1) (2) (3) (4) (5) (6) (7) (8) Outcomes in Grade 9 Outcomes in Grade 10 Dep. var. Admitted to Grade 10 Retake Courses Dropout Change School EOP 0014 0034 −0099 −0146 −0034 −0057 0000 −0005 (0032) (0042) (0072) (0115) (0027) (0039) (0022) (0034) Female X EOP −0033 0084 0045 0009 (0057) (0146) (0050) (0043) Constant 0380 0300 0346 0429 0343 0406 0161 0169 (0044) (0048) (0126) (0150) (0038) (0041) (0043) (0047) Mean dep. var. control males 0411 0411 0251 0251 0365 0365 0119 0119 Mean dep. var. control females 0578 0578 0075 0075 0211 0211 0100 0100 Observations 933 933 918 918 1157 1157 881 881 R-squared 0054 0077 0025 0028 0063 0083 0011 0011 Note: This table shows the effect of EOP on immigrant students’ outcomes in the first two years of high school, indicated above each column. EOP is a dummy equal to 1 for students in middle schools assigned to the treatment group and equal to zero for schools assigned to the control group. All regressions control in addition for a squared polynomial in INVALSI6, a dummy equal to 1 for first-generation immigrants, and province fixed effects. Standard errors clustered by school are reported in parentheses. 5.5. Spillover Effects Our last piece of evidence relates to spillover effects. Any change in the achievement and educational choices of treated students may influence their peers—particularly those sharing the same immigrant background (Sacerdote et al. (2011)). For instance, treated students could serve as role models to other students in their social network (Patacchini and Zenou (2016)). This may be particularly relevant in our context, because we showed that EOP had a strong impact on aspirations and barriers perceived by treated students. We estimate the effect of EOP on treated students’ classmates exploiting random as- signment of the intervention across schools. Specifically, we include in the sample only classmates of treated and control students and estimate the following equation: Yics = α + βEOPinClasscs + γXics + δZcs + uics  (1) where i denotes the student, c the class, and s the school. EOPinClasscs is a dummy equal to 1 if the student belongs to a class with at least one treated student, and 0 if he/she belongs to a class with at least one control student (i.e., a class in a control school that contains at least one immigrant student whose INVALSI6 score was among the top 10 of the school). In this way, the sample used in the regression contains students from classes that are comparable in terms of having high-achieving immigrant students among them, some of which received EOP and others not. Xics includes our baseline controls (gender, dummy for first-generation immigrants, and second-degree polynomial in INVALSI6); and Zcs includes school size, class size, and percentage of immigrants in the class. Table VII shows the effect of being in class with a treated student on native (columns 1–3) and immigrant (columns 4–6) classmates. Immigrant males who are in class with a treated student do not change their track choice (Panel A) but they experience a de- crease in grade retention (Panel B). On the other hand, immigrant females in the same class with a treated student are more likely to enroll in the high track (Panel A), with no detectable impact on grade retention. Interestingly, the effects on male grade retention  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 25 TABLE VII SPILLOVER EFFECTS IN EOP SCHOOLS. (1) (2) (3) (4) (5) (6) Native Classmates Immigrant Classmates All Males Females All Males Females Panel A—Dependent Variable: Choosing the high track EOP class 0001 0008 −0006 0038 −0009 0091 (0013) (0015) (0017) (0029) (0041) (0038) Mean dep. var. 0767 0755 0780 0421 0423 0419 Observations 8429 4247 4182 1308 686 622 R-squared 0197 0205 0193 0097 0096 0125 Panel B—Dependent Variable: Retained in grade 7 or 8 EOP class −0002 −0005 0002 −0042 −0064 −0019 (0007) (0009) (0007) (0023) (0032) (0027) Mean dep. var. 0044 0053 0035 0148 0192 0099 Observations 8429 4247 4182 1308 686 622 R-squared 0047 0040 0066 0046 0049 0053 Panel C—Dependent Variable: INVALSI8 EOP class −0015 −0011 −0020 −0024 0021 −0063 (0027) (0031) (0030) (0057) (0071) (0075) Mean dep. var. 0092 0098 0085 −0917 −0916 −0918 Observations 7533 3736 3797 1007 500 507 R-squared 0590 0603 0579 0331 0350 0319 Panel D—Dependent Variable: Teachers’ recommendation EOP class −0011 −0012 −0010 0031 0003 0062 (0037) (0038) (0039) (0038) (0039) (0052) Mean dep. var. 0597 0574 0621 0207 0210 0205 Observations 8429 4247 4182 1308 686 622 R-squared 0223 0235 0218 0115 0130 0112 Note: This table shows the effect of being in the same class with an immigrant student randomized into the intervention on several outcomes of interest, indicated in the title of each panel. The sample includes only classmates of treated and control students. The main explanatory variable, EOP class, is a dummy equal to 1 for the classmates of treated students. All regressions control in addition for all the individual characteristics in our baseline specification (dummy for first-generation immigrants, second-degree polynomial of test score in grade 6, and province fixed effects) as well as for class size, percentage of immigrants in the same class, and school size. Standard errors clustered by school are reported in parentheses. and female track choice are comparable in size to the direct effects observed for treated boys. This can be rationalized by observing that, by construction, non-treated immigrant classmates are selected to have lower academic performance at the beginning of middle school (they were not among the top 10 immigrants in terms of INVALSI6). It is thus possible that immigrant students down in the ability distribution are very responsive to (even indirect) treatment effects. This interpretation is reinforced when observing that immigrant females in the mid-to-lower part of the distribution make different choices from comparable native females in the absence of the intervention (see Figure 1). There- fore, it should not be surprising that, while EOP did not affect track choice of treated females (who were selected to be high-performing and already made comparable choices to natives), it had an impact on non-treated females, for whom a ‘choice gap’ existed. Why are male immigrant classmates not affected in terms of track choice (Panel A, col- umn 5)? Our interpretation is that they are too far from the margin at which a demanding high school would become a preferable choice. This can be seen, for example, when com- 26 M. CARLANA, E. LA FERRARA, AND P. PINOTTI paring soft skills of immigrant boys and girls who did not qualify for the program. Female immigrant classmates of students eligible for EOP have aspiration levels that are closer to those of treated male students, compared to other male immigrant classmates.30 At the same time, the reduction in the likelihood of grade retention for untreated male immi- grant classmates (Panel B, column 5) suggests that spillover effects did operate at that (lower) margin. Finally, columns 1–3 of Table VII show that there are no significant differences in out- comes for native classmates of treated students. This is consistent with previous evidence that peer effects are particularly strong within groups with a similar background, where students follow similar norms (see, e.g., Sacerdote et al. (2011), Bursztyn and Jensen (2015)). Overall, the positive spillovers on immigrant classmates uncovered by our anal- ysis have significant implications when assessing the success and cost-effectiveness of the intervention. We briefly discuss the latter in the next section. 5.6. Cost-Benefit Analysis Having established that EOP had sizable and significant impacts, it is important to know if it is also cost-effective. A fully fledged cost-benefit analysis is not possible at this stage: on top of the challenge of quantifying non-pecuniary benefits and costs, students involved in our experiment have not yet completed secondary school and therefore life- time earnings profiles are not observed (Heckman, Humphries, and Veramendi (2018)). The computation of the lifetime rate of return of EOP is therefore naturally based on assumptions about long-term outcomes such as college enrollment, earnings, and unem- ployment. Although EOP may potentially have strong effects on health and criminal be- havior, we present conservative estimates focusing our cost-benefit analysis only on social benefits coming from higher income taxes and public savings on unemployment insurance (Heckman, Moon, Pinto, Savelyev, and Yavitz (2010), Eisenhauer, Heckman, and Vytlacil (2015)). We examine the sensitivity of social rates of returns to a plausible range of assumptions. Appendix Table A.XV reports our calculations, which should be considered tentative for the reasons stated above. If we extrapolate the long-term benefits only on those treated individuals who were directly affected through a reduction in grade retention or a change in their high school choice, we estimate social rates of return between 3 and 5 percent. However, when we include the positive spillovers on the immigrant classmates of treated students, we estimate that the lifetime rate of return of EOP ranges between 66 and 88 percent, close to the historical return on equity. 6. CONCLUSIONS Educational segregation is a significant risk in societies where school tracking occurs at an early age. This risk disproportionately affects students whose parents have less in- formation about or are less integrated in the local education system, such as children of immigrants. We show that it is possible to reduce the mismatch created by early tracking 30 The average index of aspirations for boys and girls in the control group is −011 and 013, respectively, while the average score of their immigrant classmates (i.e., immigrant students outside the best 10 in terms of INVALSI6) is −026 and −013, respectively. Hence, in terms of average academic motivation, female immigrant classmates are closer than male classmates to students selected for the program. Elective affinity in soft skills may ease the diffusion of positive spillover effects of the program through friendship networks (Rapallini and Rustichini (2016)).  GOALS AND GAPS: EDUCATIONAL CAREERS OF IMMIGRANT CHILDREN 27 through an innovative program that provides a mix of career counseling and academic tutoring. The program, known as EOP and implemented in a random sample of mid- dle schools in Northern Italy, targeted high-achieving immigrant students selected on the basis of their test performance in grade 6. Two years later, immigrant boys assigned to treatment were 12 percent more likely than control ones to enroll in academic or tech- nical high schools (as opposed to vocational ones), virtually closing the gap with native boys. No effect was found for treated girls, for whom no mismatch was detected in the first place. The significance and magnitude of our effects are noteworthy when compared to ex- isting evidence from randomized field experiments that study partial derivatives of the human capital production function, recently summarized by Kautz, Heckman, Diris, Ter Weel, and Borghans (2014) and Fryer (20