Finding a certain person

American Sociological Review 77(3) 463 –485 © American Sociological Association 2012 DOI: 10.1177/0003122412440802 http://asr.sagepub.com

Throughout adolescence, boys are overrepre- sented among high school dropouts, special education students, and every failed or special needs category. Boys’ notorious underperfor- mance in school and their tendency to disrupt the learning process in classrooms has sparked intense academic and public debates about the causes of what many now call the “prob- lem with boys.” Some see the gender gap as largely biological in origin. Others blame schools for a de-masculinized learning envi- ronment and a tendency to evaluate boys negatively for fitting into this environment less well than girls. Yet, the true impact of

school context on the size of the gender gap in academic performance remains controver- sial. The 1966 Coleman Report raised the profile of research on school effects, and much attention since then has been motivated by a concern for equality of educational opportunity by social class and race. Now that

440802ASRXXX10.1177/0003122412440802Lege wie and DiPreteAmerican Sociological Review 2012

aColumbia University

Corresponding Author: Joscha Legewie, Columbia University, Department of Sociology–MC9649, 606 W. 122nd Street, New York, NY 10027 E-mail: [email protected]

School Context and the Gender Gap in Educational Achievement

Joscha Legewiea and Thomas A. DiPretea

Abstract Today, boys generally underperform relative to girls in schools throughout the industrialized world. Building on theories about gender identity and reports from prior ethnographic classroom observations, we argue that school environment channels conceptions of masculinity in peer culture, fostering or inhibiting boys’ development of anti-school attitudes and behavior. Girls’ peer groups, by contrast, vary less strongly with the social environment in the extent to which school engagement is stigmatized as un-feminine. As a consequence, boys are more sensitive than girls to school resources that create a learning-oriented environment. To evaluate this argument, we use a quasi-experimental research design and estimate the gender difference in the causal effect of peer socioeconomic status (SES) as an important school resource on test scores. Our design is based on the assumption that assignment to 5th-grade classrooms within Berlin’s schools is as good as random, and we evaluate this selection process with an examination of Berlin’s school regulations, a simulation analysis, and qualitative interviews with school principals. Estimates of the effect of SES composition on male and female performance strongly support our central hypothesis, and other analyses support our proposed mechanism as the likely explanation for gender differences in the causal effect.

Keywords causal inference, education, gender, gender gap, peer effects

464 American Sociological Review 77(3)

a growing gender gap in educational attain- ment has emerged, it is important to extend this line of research and ask whether schools affect gender inequality, and if so, what are the mechanisms by which this occurs.

Integrating theories about gender identity, adolescent culture, and findings from prior ethnographic classroom observations, we argue that the school environment channels conceptions of masculinity in the peer cul- ture, fostering or inhibiting boys’ develop- ment of anti-school attitudes and behavior. An academically oriented environment sup- presses the construction of masculinity as oppositional and instead facilitates boys’ commitment by promoting academic compe- tition as an aspect of masculine identity. Lower quality schools, by contrast, implicitly encourage—or at least do not inhibit—devel- opment of a peer culture that constructs resist- ance to schools and teachers as valued masculine traits. Girls’ peer groups, by con- trast, vary less strongly with the social envi- ronment in the extent to which school engagement is stigmatized as un-feminine. As a result, boys, in particular, benefit from school resources that create a learning-ori- ented peer culture, and the size of the gender gap in educational performance depends on environmental factors connected to the qual- ity of schools.

We evaluate our argument with a quasi- experimental research design using reading test scores as an outcome variable and the socioeconomic composition of the student body as the focal treatment variable. This design is based on within-school variation across classes using the ELEMENT data from one German city-state (i.e., Berlin). In con- trast to the United States, the lack of perfor- mance-based tracking in Berlin’s elementary schools, and parents’ smaller influence on classroom assignment, makes it plausible that student assignment to elementary school classrooms in Berlin is almost random. To develop a detailed understanding of the actual selection process, we examine official school regulations, provide statistical evidence from simulation analyses, and conduct qualitative

interviews with school principals. Results suggest that randomness indeed plays an important role in the assignment process, but they also point at potential sources of bias. We address these potential biases statistically with targeted sensitivity analyses using instru- mental variable and sample restriction meth- ods. We supplement the ELEMENT analysis with estimates from a large-scale nationally representative German dataset (PISA-I-Plus 2003) to address potential concerns about generalizability of the findings.

Results of our investigation support our core hypothesis. In addition, a systematic comparison of our preferred explanation with alternative accounts suggests that our hypoth- esized mechanism brings about the gender difference in the causal effect of SES compo- sition on student achievement. Our findings speak to recent political debates about the educational shortcomings of boys by deepen- ing our understanding of their notorious underperformance. Our analytic strategy also makes a methodological contribution by illus- trating how a detailed study of selection pro- cesses using simulations and qualitative interviews can assist estimation of causal effects.

EDuCATionAL ouTComES AnD SChooLS The 1966 Coleman Report (Coleman 1966) claimed that family is the most important determinant of achievement, but performance improves when classroom peers have greater socioeconomic resources and are racially integrated (see also Coleman 1961; Jencks and Mayer 1990; Kahlenberg 2001). As Coleman and others have subsequently argued, students are motivated to invest more heavily in their studies when adolescent cul- ture rewards academic performance and thereby supports parents’ and teachers’ reward systems. But when adolescent culture values other behaviors more highly (e.g., sports, being popular with the opposite sex, or oppo- sition to school authority), and especially when adolescent culture denigrates academic

Legewie and DiPrete 465

achievement, it inhibits academic investment and weakens academic achievement. Simply put, students who are highly motivated and capable (attributes more common at higher SES levels) create a learning-oriented peer culture (Jencks and Mayer 1990; Rumberger and Palardy 2005; Sewell, Haller, and Portes 1969).

For about 20 years following the release of the Coleman Report, the literature reported that school effects were relatively small in comparison to family effects, and therefore “schools are not an effective agent for the redistribution of societal resources” (Hallinan 1988:255; see also Hanushek 1989). This pes- simistic view of schools began to change with the rise of the accountability and standards movement to improve schools (Schneider and Keesler 2007). Reanalysis of earlier studies suggested a more consistently positive rela- tionship between school resources and stu- dent achievement (Greenwald, Hedges, and Laine 1996) and found that teacher quality, in particular, was a major input into student learning (see also Murnane 1983).

The renewed focus on schools’ impact on learning has not obscured attention to the central conclusion of the Coleman Report that “the social composition of the student body is more highly related to achievement, inde- pendent of the student’s own social back- ground, than is any school factor”(Coleman 1966:325). Far more than was historically appreciated, estimation of peer effects is chal- lenging (Angrist and Pischke 2008) because of non-random selection and unmeasured confounding variables (e.g., teacher quality) that affect student outcomes. The most per- suasive recent studies use natural experiments to estimate the impact of changes in class composition on outcomes (e.g., Imberman, Kugler, and Sacerdote 2009). A second strat- egy is to exploit potentially random assign- ment of students to classes within schools. This strategy is only persuasive when applied in school districts that make it difficult for parents to teacher shop (Ammermueller and Pischke 2009). A third strategy examines arguably random fluctuations in adjacent

cohorts (e.g., of gender or race composition) for the same school and grade (Gould, Lavy, and Pawerman 2009; Hoxby 2000), although these studies have not looked at peer effects related to socioeconomic characteristics. The magnitude of estimated effects is not large (about .15 standard deviations), but it is about the same as some of the most believable estimates of teacher effects, whether for academic, social, or behavioral outcomes (Jennings and DiPrete 2010; Rockoff 2004). Meanwhile, recent studies whose primary estimation strategies control for observable potential confounders have found a similar effect size on test scores (Crosnoe 2009; Rumberger and Palardy 2005).

SChooL ConTExT AnD ThE GEnDEr GAP in EDuCATion The original focus on school effects devel- oped out of a concern for equality of educa- tional opportunity by social class and race. Now that a growing gender gap in educa- tional attainment has emerged, it is natural to ask whether schools also affect gender inequality, and if so, what are the mechanisms by which this occurs. Starting in the 1970s and early 1980s (Spender 1982; Stanworth 1984), ethnographic studies documented girls’ and boys’ gendered behavior at school as well as the different ways that teachers treat girls and boys. Although overt discrimi- nation against girls in the classroom has declined over the past three decades, recent studies suggest that boys still verbally domi- nate the classroom (Jovanovic and King 1998; Sadker and Zittleman 2009). Meanwhile, the once celebrated coeducation of boys and girls as a pivotal step toward gender equality is now challenged by the increasing popularity of single-sex private schools, the opening of girls-only public schools, and the claimed educational short- comings of coeducation for girls (Morse 1998; Salomone 2003).

Despite these important strands of research and the general recognition that schools are an important context for socialization of

466 American Sociological Review 77(3)

young adolescents, literature on the educa- tional gender gap has widely ignored the school as a potential source of variation in this gap. To our knowledge, Dresel, Stöger, and Ziegler (2006), Machin and McNally (2005), and Schöps and colleagues (2004) are the only studies that examine variation in the size of the gender gap across a number of schools. Using data from a specific region in Germany (i.e., Baden-Württemberg), Dresel and colleagues (2006) found substantial vari- ation in the educational gender gap across schools and classes, while Schöps and col- leagues (2004) obtained a similar finding using the German PISA data. Machin and McNally (2005), by contrast, argue that spe- cific school-based characteristics, such as school inputs, teaching practices, and the examination system, have no effect on the gender gap. We extend this line of research by building on reports from prior ethnographic classroom observations and theories about gender identity to understand the role of school context for boys’ underachievement.

Boys’ Underachievement, Gender Identity, and School Climate

In a classic study, Willis (1981) argued that working for academic success is in conflict with adolescent conceptions of masculinity. He portrayed working-class white boys’ anti- school attitudes and behavior as arising from peer dynamics and a belief that their opportu- nity to use education to achieve success in the labor market was blocked (see also Kao, Tienda, and Schneider 1996; MacLeod 2008). In line with Willis’s early findings, much of the literature on boys’ underachievement focuses on disincentives to engage with school that stem from adolescent conceptions of masculinity, which are developed and rein- forced in peer groups. Gender differentiation and creation of stereotypical gender identities begin in early childhood before children have had any experience with school (Davies 2003; Maccoby 1998; Thorne 1993). Gender- differentiated childhood cultures become the basis for gender-differentiated adolescent

cultures, which are important influences on how children view school, whether they take school seriously, and how hard they work as students (Steinberg et al. 1997).

Classroom observations and other ethno- graphic studies document how gender identi- ties are constructed in the classroom and how these gender cultures affect boys’ and girls’ interactions and approach to education (Fran- cis 2000; Pickering 1997; Salisbury and Jack- son 1996; Skelton 1997). Boys tend to be noisier, more physically active, and more easily distracted than are girls (Francis 2000; Howe 1997; Spender 1982; Younger, War- rington, and Williams 1999). Studies also find that masculine stereotypes portray boys as competitive, active, aggressive, and dominat- ing, while girls are viewed as conciliatory and cooperative (Francis 2000). Other scholars argue that stereotypical gender identities per- petuate the belief that girls have to work hard to learn in school, whereas boys are naturally gifted (Cohen 1998; Epstein 1998; Mac an Ghaill 1994; Power et al. 1998; Quenzel and Hurrelmann 2010). Cohen (1998) shows that these gendered beliefs are reflected in boys’ casual and detached attitude toward school, which accords with the ethnographic studies referenced earlier. Despite the transformation of gender relations in modern societies, ste- reotypical gender identities continue to shape orientations toward school and produce behaviors that reinforce these identities while potentially affecting children’s academic suc- cess. Morris (2008:736) observed this process at a rural high school: “girls tended to direct considerable effort and attention to school” whereas “boys . . . took pride in their lack of academic effort” as an aspect of their mascu- line identity.1

Peers and the adolescent reward system reinforce gender identities and gendered behavior patterns. In some contexts, disrup- tive behavior produces status gains in lower SES students’ peer groups. Working for aca- demic achievement, by contrast, is labeled as feminine and thereby stigmatized. Girls, how- ever, typically view school work as accepta- ble and sometimes even encouraged. In a lack

Legewie and DiPrete 467

of parallelism with male peer groups, work- ing-class and lower-class female peer groups do not consider resistance to authority and disengagement from school to be core aspects of feminine identity (Maccoby 1998). As a result, girls’ peer culture more readily encour- ages attachment to teachers and school.2

A diverse group of studies supports the role of peers in shaping attitudes toward school. Coleman (1961), Eitzen (1975), Steinberg and colleagues (1997), and, more recently, Bishop and colleagues (2003) argue that adolescents value attributes that make one cool or popular because these attributes are linked with high status. Based on her own and others’ ethnographic work, Epstein (1998:106) argues that “the main demand on boys from within their peer culture . . . is to appear to do little or no work” whereas for girls “it seems as if working hard at school is not only accepted, but is, in fact, wholly desirable.” Morris (2008:738; for other exam- ples, see Epstein 1998) documents this atti- tude in a conversation between three boys in an English class:

Kevin: “I don’t want to put in a lot of extra effort like that. I’ll just do the basic stuff and get a B.” “I got an 87 in here,” he says proudly. Warren chimes in, “Yeah, I hate these pussies who make like an A minus and then they whine about it.” Kevin says, “Yeah it’s like why do you care? Why does it have to be better? Nothin’ wrong with a normal grade!”

Although ethnographic studies document substantial within-gender diversity in the con- struction of gender identities, evidence on typical gender differences is rather persua- sive. Young boys tend to construct masculin- ity at least partly in terms of resistance to school. This conception of masculinity may be partially responsible for male under- achievement (Francis 2000; Pickering 1997; Salisbury and Jackson 1996; Skelton 1997). The conception of female identity and female peer culture, by contrast, is not as closely tied to resistance to school, and indeed may even

support schoolwork as a positive attribute of femininity. As a result, girls consistently have better work habits and a stronger pro-school orientation.

While Willis and others mainly focus on consequences of lower- and working-class backgrounds for anti-school attitudes among boys, we are interested in the school and class environment as a context that either encour- ages or limits development of anti-school attitudes and behavior. High-status parents generally foster an orientation for their boys that is at least instrumentally focused on high performance in school. These parents also have resources to intervene in their children’s lives to counter signs of educational detach- ment or poor performance. As Coleman and others argue, schools can play a similar role in enhancing students’ incentives to be engaged with academics by creating a learn- ing-oriented peer culture. In this line, many scholars argue that the success of some char- ter schools, such as KIPP and the Harlem Children’s Zone, comes from their ability to foster a learning-oriented environment (Rav- itch 2010).

We argue that boys gain more than girls from a learning-oriented environment because it channels how masculinity is constructed in the school culture. Such an environment pro- motes academic competition as an aspect of masculinity and encourages development of adaptive strategies that enable boys to maintain a show of emotional coolness toward school while being instrumentally engaged in the schooling process. In other words, academic competition as one of the “different ways of ‘doing’ masculinity” (Francis 2000:60; see also Mac an Ghaill 1994) becomes a more important part of the construction of masculine identity in certain environments.

As is true in the family, production of an academically oriented environment in school is not effortless. It requires resources. Better facilities, better curriculum, better teachers, and better support staff all can produce more value-added in school. Both boys and girls will generally benefit from better schooling, of course, but we expect school inputs that

468 American Sociological Review 77(3)

strengthen a learning orientation in the stu- dent culture have the potential to enhance educational outcomes especially strongly for boys. Teachers, for example, can promote a learning-oriented student culture. Accord- ingly, we expect teachers with the right col- lection of skills might have especially positive effects on boys’ achievement.

Here, the school resource of central inter- est is the socioeconomic composition of the student body. Peers’ impact on school climate and student achievement has played a crucial role in the literature on schools ever since Coleman (1966:325) claimed that “the social composition of the student body is more highly related to achievement, independent of the student’s own social background, than is any school factor.” The mechanism behind this association is cultural; students with high motivation and achievement from high-SES backgrounds create a learning-oriented peer culture and assist teachers in the process of education (Jencks and Mayer 1990; Rum- berger and Palardy 2005; Sewell et al. 1969). We expect the disadvantages of low-SES composition to be larger for boys than for girls because of the evidence that lower SES student bodies create a stronger oppositional culture in male than in female peer groups. Conversely, an academically oriented envi- ronment in schools channels conceptions of adolescent and pre-adolescent masculinity, suppresses boys’ negative attitudes toward school, and facilitates academic competition as an aspect of masculine identity. Girls’ peer groups, on the other hand, more readily and independently of school context encourage attachment to teachers and school, and do not identify femininity with disengagement from school. Based on this argument, we hypothe- size that the female advantage in academic achievement is bigger in schools with a lower SES composition in their student body.3

DATA AnD mEThoDS We address our core hypothesis with the German ELEMENT dataset, using reading test scores as an outcome variable and SES

composition of classroom peers as our focal treatment variable. The ELEMENT dataset is a longitudinal study that assessed develop- ment of reading and math ability in the 4th, 5th, and 6th grades in Berlin schools (Lehmann and Lenkeit 2008). It includes about 3,300 students who attended 4th grade during the 2002 to 2003 school year in 71 randomly selected elementary schools in Berlin, and all 1,700 students who attended 5th grade in 2003 to 2004 in one of Berlin’s 31 upper-secondary schools that begin with 5th grade.4 In our final models, we combine these two ELEMENT samples and control for school type through school-level fixed effects. We also examine whether relevant effects vary by school type using interaction terms (they do not). Part A in the Appendix provides a short introduction to the German educational system.

The ELEMENT dataset includes at least two classrooms for every school.5 This fea- ture provides the basis for a quasi-experimen- tal design. It allows us to estimate contextual effects of 5th-grade class composition by gender using school-level fixed-effects mod- els, because the original assignment to ele- mentary school classes in 1st grade within schools is not subject to self-selection or parental control.6 This estimation strategy provides a clear advantage over similar esti- mates based on data from U.S. schools, where performance-based tracking in elementary schools and parents’ influence on assignment to classes are more pronounced.

Our quasi-experimental research design provides high internal validity and allows us to make a strong case for causal inference, but the analysis is geographically limited to a single German state. To address this limita- tion, we supplement the ELEMENT data with the German PISA-I-Plus 2003 data, a German extension of the international PISA study.7 The PISA-I-Plus includes a nationally repre- sentative sample of 9,000 students in at least two 9th-grade classrooms in 220 schools (PISA-Konsortium Deutschland 2006).8 The two datasets complement each other and pro- vide strong internal and external validity for the estimation of causal effects.

Legewie and DiPrete 469

School-Level Fixed Effects as a Quasi-Experimental Identification Strategy

Regression or matching estimates of school effects based on conditioning on observable variables as an identification strategy poten- tially suffer from endogeneity problems. These strategies rely on the assumption that students are randomly assigned to schools conditional on observable covariates in the model (Sørensen and Morgan 2006). This common identification strategy is especially problematic for estimation of school effects with cross-sectional data. Students clearly are not randomly assigned to schools, and it is unlikely that this non-random assignment can be perfectly modeled with the observed covariates.

To avoid these potential endogeneity prob- lems, we estimate school-level fixed-effects models using ELEMENT and PISA-I-Plus data. Both datasets contain an additional level of analysis, the classroom. We argue that most students are randomly assigned to classrooms conditional on their school in Berlin’s elemen- tary and 5th-grade upper-secondary schools (for a similar strategy, see Ammerm ueller and Pischke 2009). Assuming students’ ran- dom assignment to classes within schools, we can estimate the causal effect using school fixed-effect models and a measure of SES composition on the classroom level (we dis- cuss the variables in detail below). We specify these models as

y female SESComp

SESComp female

y

ijk j i k

k i

i

= + +

+ ×

+

α θ

δ

β

γ ( ) ( )

(( ) )

1

4tth grade

i k ijk + + +X Uβ β

2 3 ε

(1)

where i, j, and k are indices for individuals, schools, and classes, respectively; αj are school fixed-effects; y

i

4th grade is a student’s prior achievement measured in 4th grade; and X

i and U

k are sets of control variables on the

individual and class level, respectively.9 Due

to data limitations, analysis with the PISA-I- Plus dataset omits the variable of prior achievement on the right-hand side.10

These models examine whether class-to- class variation in performance is systemati- cally related to class-to-class variation in socioeconomic composition, controlling for all unobserved school characteristics (and therefore the non-random selection of stu- dents into schools). Coefficients of interest are θ, which captures the causal effect of the socioeconomic class composition, and δ, which captures the difference in this effect between boys and girls. We expect to find a positive effect of SES composition, as previ- ously documented, and, more important for our theory, a negative estimate of the interac- tion term, indicating that boys are more sensi- tive than girls to peer SES. Pre-treatment control variables on student and class levels are of secondary interest; we include them to increase balance between the treatment and control groups (for a description of the con- trol variables, see Table 1).

Assignment of Students to Classrooms within Schools

Our estimation strategy relies on the assump- tion that selection of students into different classes within schools is practically random. While students self-select into schools, their allocation to different classes within schools is arguably less selective, but it might not be completely random. In particular, the alloca- tion process, and therefore selection into treatment, might involve three potential biases: (1) parents might influence which class their children attend; (2) schools might allocate students based on certain characteris- tics (e.g., performance-based tracking or sub- ject choice); and (3) children might self-select over time when certain children have to repeat a class or change schools. Even if stu- dents are randomly assigned to classes, cer- tain teachers might be assigned to specific classes based on a classroom’s composition, which could create a bias in relevant esti- mates of classroom composition.

470 American Sociological Review 77(3)

To develop a deeper understanding of the actual selection process, we conducted a three-part analysis. First, we studied official school regulations in Berlin. Second, we used a simulation-based approach to compare observed composition of classes with simula- tions involving random assignment of stu- dents to classrooms within schools. Third, we conducted qualitative interviews with school principals in Berlin. The resulting detailed

picture of the actual selection process allows us to evaluate our argument that self-selection is practically random and to design targeted statistical sensitivity analyses that address potential sources of biases.

School regulations and general con- siderations. Primary school regulations in Berlin (Grundschulverordnung Berlin, §8) prohibit allocation of students based on gender,

Table 1. Variables in Main Analysis at Individual and Class Level

Variables Description Mean SD Mean (Male)

Mean (Female)

Diff. in Means

Dependent Variable Achievement 5th-grade reading test

scores 109.92 13.65 108.79 111.05 −.17**

Independent Variables (Individual Level) Prior achievement 4th-grade reading test

scores 103.22 15.96 101.39 105.07 −.23**

Female 0 = Male; 1 = Female .50 .50 Family background ISEI scale 51.45 16.63 51.09 51.80 −.04 Migration background Categorical Variable: 1 = both parents born

in Germany .64 .48 .64 .65 −.01

2 = one parent not born in Germany

.12 .33 .12 .13 −.01

3 = both parents not born in Germany

.17 .38 .18 .16 .02*

4 = child not born in Germany

.06 .24 .06 .06 −.01

Class repeater 0 = not repeated; 1 = repeated at least once

.06 .23 .07 .04 .03**

Interaction terms Female x Family background

Female x Migration background

Female x SES composition

Independent Variables (Class Level) Size of class Number of students

in the class 23.33 3.41 23.33 23.34 −.00

SES composition Average ISEI at class level (aggregated)

51.45 9.40 51.08 51.10 −.01

Source: ELEMENT data.

Note: N = 4,372. The difference in means refers to the mean for boys minus the mean for girls divided by the pooled standard deviation. The difference in means is not standardized for the binary variables (i.e., migration background dummies and indicator for class repeater). All continuous variables are standardized for the final analysis. * p < .05; ** p < .01; *** p < .001 (two-tailed t- and z-tests).

Legewie and DiPrete 471

first language, or performance and emphasize heterogeneity of classes in regard to these characteristics. These legal constraints rule out performance-based tracking, set limits on parental influence over classroom assignment, and provide guidelines for classroom assign- ment of grade repeaters and newcomers. As a consequence, allocation of students to class- rooms based on family background is unlikely. Regulations also mention, however, that schools can consider existing friendships between new students and assign them to the same classroom. This practice, if common, might create a bias in the assignment process that could pose a problem for estimation of the causal effect.

In secondary schools such as those in the PISA-I-Plus data, class-specific tracking based on subject choice (e.g., foreign lan- guage) is more common, and a higher number of students have to repeat a class compared with elementary school. This creates poten- tially non-random allocation of students to classrooms, so the selection problem might be more pronounced in secondary schools. In Berlin, however, 5th-grade upper-secondary schools (grundständige Gymnasien) are dif- ferent from other secondary schools. Their student population is more homogeneous compared to other secondary schools, which makes a purposeful allocation to different

classes relatively inconsequential. In addi- tion, assignment to 5th grade is not subject to selection over time through grade retention because students enter these schools for the first time at 5th grade.

Based on these considerations, we expect that assignment to 5th-grade classrooms is practically random in Berlin’s elementary and upper-secondary schools, whereas assign- ment to 9th-grade classrooms in secondary schools is subject to more pronounced selec- tion processes.

Simulation of random assignment. We use a simulation-based approach to evaluate whether within-school variation in socioeco- nomic composition across classrooms created by the actual (unknown) allocation process is consistent with random assignment. Figure 1 compares socioeconomic composition across classrooms obtained from simulations that randomly assign students to classrooms (histo- gram) with the observed composition (vertical line), that is, the average variation of class means within schools (see Part B in the Appen- dix for details on the simulation).

For the two ELEMENT samples, the observed mean is consistent with a random assignment process. This is in line with our expectation about assignment to classrooms in 5th grade. As expected, however, the

Figure 1. Comparison of Observed Classroom Composition with Simulations Involving Random Assignment Note: Graphs show average variation of class means within schools for the observed samples (vertical line) together with the sampling distribution of this statistic obtained from 1,000 random simulations (histogram). Vertical grey lines indicate 95% confidence intervals from the simulations.

472 American Sociological Review 77(3)

observed value for secondary schools in the PISA-I-Plus is relatively unlikely to occur under random assignment. Similar simula- tions for the proportion of students with migration backgrounds suggest that assign- ment in regard to this characteristic is consist- ent with randomness for all three datasets. Finally, variation in gender composition across classrooms within schools is smaller in the actual data than in the simulated distribu- tion (see the online supplement [http://asr. sagepub.com/supplemental]). This result sug- gests that schools distribute boys and girls equally across classrooms.

These results provide statistical evidence to support the previously described institu- tional evidence that assignment to classrooms within schools with respect to family back- ground is as good as random in the ELE- MENT dataset. By contrast, some non-random selection process seems to play a role for 9th-grade classrooms in secondary schools.

Interviews with school principals. Although simulations are informative, they do not provide information about the actual assignment process. It is still conceivable that non-random selection processes are at work that produce a distribution of students in terms of socioeconomic status that is consistent with a random assignment process. To develop a deeper understanding about the actual assignment process, we conducted 12 inter- views with school principals in Berlin as the central actors in the allocation process (nine for elementary schools and three for upper- secondary schools). We selected schools using a random sample that we then supple- mented with specific schools to ensure diversity in regard to neighborhood and ethnic composition. Interviews lasted about 15 to 20 minutes and focused on the actual procedure schools use to assign students to classes, crite- ria that play a role in assignment, the extent to which parents try to influence this process, and how schools deal with parental requests. Interviews also solicited information about how schools assign students who repeat a

class or who transfer from other schools and about how teachers are assigned to class- rooms. The online supplement contains a detailed description of the sampling proce- dure and a translation of the interview questions.

While the schools under study use differ- ent procedures to assign students to classes, a number of findings emerged from the inter- views. First, none of the principals reported that they directly take family background or performance into account in the assignment process, and most schools do not respond to parents who try to influence the process (for an exception, see below). Second, schools try to have classes of similar size. This plays an important role in assignment of students who either repeat a grade or transfer from another school. Third, assignment of teachers to class- rooms is generally not connected to socioeco- nomic composition or other characteristics of the class. Teacher assignment is based on scheduling issues and past experience with the teacher.11

A number of potential biases, however, do exist. First, while all school principals empha- sized that a desire to equalize classroom size is the main criterion, principals also reported that students who repeat a grade are some- times assigned to specific classes based on expectations about social dynamics. Second, some principals reported that they take into account whether groups of children attended the same kindergarten and try to assign these students to the same 1st-grade classroom. Other principals said they follow parent requests when they are related to friendships between two new students, which often devel- oped because the children attended the same kindergarten. Third, while most principals reported distributing children with immigra- tion backgrounds equally across classes, two principals said they create a separate class for children who are learning German. While the simulations suggest the contrary, this finding makes it unclear how common the practice of sorting students by migration background or language skills is. We take special care to

Legewie and DiPrete 473

address this potential issue statistically. Fourth, all principals reported that they try to ensure gender balance between the class- rooms. This practice is consistent with results from the simulation insofar as variation in the proportion of female students across classes within schools is smaller than what we would expect from random assignment.

Except for the last criterion, which is irrel- evant because boys and girls are equally dis- tributed across families, these selection criteria might induce some systematic bias in the composition of classrooms. The impor- tance of these selection criteria, however, seems to be limited. Most school principals independently and without knowledge of our study concluded that randomness plays an important role in the assignment process because they simply have little prior knowl- edge about entering students and because the whole process is not very systematic. One elementary school assistant principal and teacher, for example, emphasized that even decades of experience at elementary schools could not remove the inherent unpredictabil- ity about classroom dynamics, given the lim- ited prior knowledge about entering students that schools have to work with:

We have realized again and again that even if we try to make sense of the classroom com- position based on names or other attributes we know about, there is no way to know how the class actually turns out in regard to its social composition. Even though I have been working at schools for 40 years now, there are always unexpectedly difficult or bal- anced classes, which really depends on the personalities of the students inside the class- room so that in the end randomness plays a big role. (translation by authors)

We elicited these and similar concluding remarks from interviewees at the end of inter- views by asking how they would weigh the importance of different criteria and whether they thought randomness also plays a role. These observations are particularly interest- ing considering that we expected a social

desirability bias in favor of principals report- ing a sophisticated assignment procedure.

Conclusions about selection process. Based on evidence from school regulations, the simulations, and interviews with school principals, we conclude that the role of poten- tial selection biases is limited. Results justify our quasi-experimental design and support our argument that using within-school varia- tion across classrooms in Berlin elementary schools greatly improves our estimates com- pared to estimates based only on between- school variation. We also recognize the poten- tial selection biases documented by the interviews, and we address these problems statistically by conducting a set of targeted sensitivity analyses. These robustness checks are based on instrumental variable analyses and sample restrictions specifically designed to address each potential source of bias.

Finally, we note that in contrast to most research on compositional school effects, we are not fundamentally interested in school performance as an outcome. Rather, we address contextual determinants of the gender gap in school performance. While evidence from the interviews indicates that students might select into certain classrooms, it seems unlikely there is differential selection of boys and girls into different classrooms. Non-ran- dom assignment to classrooms only matters for our key estimation results to the extent that schools treat boys and girls differently during the assignment process. Interviews did not provide any indication of differential treatment of boys and girls, even though school principals were asked directly about such a possibility. This fact enhances our con- fidence in the validity of our estimates.

Variables and Treatment of Missing Data

Our analysis uses reading test scores in 5th grade (ELEMENT) and 9th grade (PISA-I- Plus) as the main outcome variable (see Table 1 for descriptive statistics). Campbell and col- leagues (2001:1) describe reading scores as

474 American Sociological Review 77(3)

“one of the most important abilities students acquire through their early school years. It is the foundation for learning across all sub- jects.” Reading literacy also figures impor- tantly in research on the gender gap in education, because boys’ reading test scores lag notably behind that of females (Buchmann, DiPrete, and McDaniel 2008). Some research- ers even argue that boys’ failure in general is due to their deficits in reading (Whitmire 2010). We measure test scores on a common scale using item response theory and stan- dardize them with a mean of zero and a stan- dard deviation of one.

Our focal treatment variable is the student body’s socioeconomic (SES) composition, which we measure at the classroom level as the average social status on the ISEI scale (Ganzeboom, Treiman, and Ultee 1991).12 One could argue that peers’ prior achievement is a more natural contextual measure for test- ing our core hypothesis. However, peer achievement is endogenous in our data because it is measured after random assign- ment. Moreover, the correlation between peer achievement and SES is too high to reliably distinguish effects of the two variables. Accordingly, SES composition provides a stronger test (i.e., one resting on weaker assumptions) of our theory than could be obtained using peer achievement. In addition, a long tradition in sociology, going back to the Coleman Report, sees SES composition as connected to a peer group’s learning orienta- tion because attributes such as high motiva- tion and capability are more common among students from high-SES families. Conse- quently, a student body’s SES composition is a school resource that fosters a learning orien- tation, and it is highly relevant for our study.

Along with SES composition, we use a comprehensive set of control variables, includ- ing 4th-grade test scores, as a measure of prior performance. Table 1 describes these variables together with descriptive statistics. All inde- pendent continuous variables are standardized to have a mean of zero and a standard devia- tion of one across the combined sample of males and females in both datasets.

The Forschungsdatenzentrum at the IQB provides five imputed versions of the ELE- MENT dataset (see Lehmann and Lenkeit 2008). We performed each analysis separately for the five imputed datasets and then com- bined the different estimates to obtain the final results presented here. We employed a similar imputation strategy based on the chained equations approach for the PISA-I- Plus dataset.

rESuLTS Variation of the Gender Gap across Schools

In an average school, the female advantage in reading scores is about .12 standard deviations in 5th grade and .21 standard deviations in 9th grade. It ranges from –.04 to .28 standard deviations in 5th grade and from .07 to .35 standard deviations in 9th grade for 95 percent of schools. Expressed in terms of years of education, girls are .36 school years ahead in 5th-grade reading test scores in an average school, but the gap ranges across schools from a male advantage of .12 years to a female advantage of .83 years.13 Figure 2 plots this variation in the gender gap on the school level against average performance at a school. The striking pattern in the figure indicates that schools with higher than average performance also have the smallest gender gap. This find- ing is consistent with our theoretical predic- tion; it suggests that boys do not fall as far behind in performance-oriented schools.

SES Composition and the Gender Gap in Education

Table 2 presents estimates from school-level fixed-effect regression of reading test scores in 5th grade on classroom-level SES compo- sition, 4th-grade scores, and other control variables. The table shows the main effects of gender and SES composition on the class- room level together with the interaction between SES composition and gender (all coefficients are in standard deviation units).

Legewie and DiPrete 475

Other coefficients are omitted from the table (for the full regression results, see Table S1 in the online supplement). The table also shows fixed-effect (FE) estimates from the PISA-I- Plus data for 9th-grade reading test scores without a measure of prior performance and estimates from a multilevel (MLM) model on the school level with a broad set of control variables. We include MLM estimates as a

comparison because they reflect one of the most common estimation strategies (i.e., con- ditioning on observable covariates) used in sociology to identify compositional peer effects (e.g., Rumberger and Palardy 2005).

Results in Table 2 show that SES composi- tion has a positive and highly significant effect on reading test scores in all models for gain scores (top row) and raw scores. This

Figure 2. Gender Gap and Average Performance across Schools in Standard Deviation Note: Estimates shown in the figure are based on a multilevel model with two levels (student and schools) and with a random intercept and a random slope for female on the school level so that average performance and effect of gender are allowed to vary across schools. Dots represent empirical Bayes predictions for the random intercept (i.e., average school performance) against the prediction for the random slope (i.e., the female advantage).

Table 2. Effect of SES Composition for Boys and Girls in Standard Deviations

Female SES Comp. SES Comp. x Female

Model Prior

Performance Coef. (se) Coef. (se) Coef. (se)

1. FE Estimate (ELEMENT)

yes .007 (.02) .091* (.04) −.060** (.02)

2. FE Estimate (ELEMENT)

no .120*** (.03) .178*** (.06) −.057* (.03)

3. FE Estimate (PISA-I-Plus 2003)

no .196*** (.03) .237*** (.03) −.052* (.02)

4. MLM Estimate (PISA-I-Plus 2003)

no .143 (.11) .303*** (.05) −.099* (.04)

Note: FE = fixed effect. Table 1 describes control variables. The full set of coefficient estimates for Models 1 and 2 are in Table S1 in the online supplement. The number of students for models based on ELEMENT is 4,372, the number of schools is 101, and the average number of students per school is 43.3. N for PISA-I-Plus is 8,559. * p < .05; ** p < .01; *** p < .001 (two-tailed tests); standard errors adjusted for clustering on class level.

476 American Sociological Review 77(3)

result conforms with previous findings in the literature on effects of SES composition (Jencks and Mayer 1990; Rumberger and Palardy 2005). In all models, the point esti- mate for the interaction between SES compo- sition and female is negative and significant. Most important, estimates from the fixed- effect model using the ELEMENT data along with a control variable for prior performance show that boys learn more in classes with higher average SES. Adding additional peer characteristics, such as the proportion of for- eign-born students, to this specification does not affect this finding (results not shown here). Results from the two FE models based on ELEMENT and PISA-I-Plus data without 4th-grade performance show the same results (we include ELEMENT results for direct comparison). In particular, the main effect of SES composition in the model based on PISA-I-Plus data seems to be upwardly biased (.237 compared to .178), and both estimates are somewhat larger than the .15 effect size estimated by Crosnoe (2009). However, the estimated size of the interaction between female and SES composition is very similar across the three fixed-effect models. This finding supports our argument that even if students self-select into classes (and self- selection appears to be more important in 9th grade), boys and girls are unlikely to differ in this selection process, which increases our confidence in the ELEMENT estimates. Results from the MLM model point in the same direction but appear to be upwardly biased. In particular, the estimate for the interaction is about 90 percent higher in the MLM model compared to the corresponding FE model. This could reflect the fact that the MLM estimate is based on non-random school-level variation, while the fixed-effect estimate is based on almost-random class- room-level variation within schools. The school-based estimate’s larger size might also reflect spillover effects between SES compo- sition of two classrooms in the same school. Given the possibility of selection bias in the MLM estimates, we consider the fixed-effects classroom-based estimates to be a more definitive test of our theoretical prediction.

Overall, our estimates provide strong evi- dence that boys are more sensitive than girls to the important school resource of classroom SES composition. Our statistical evidence is strengthened by the fact that institutional, simulation-based, and qualitative evidence indicates that randomness plays a central role in the allocation of students to 5th-grade classrooms in Berlin.

Targeted Sensitivity Analysis

In this section, we investigate whether our results are sensitive to three potential selection biases documented in interviews with school principals. Our detailed knowledge about the assignment process allows us to design a set of sensitivity analyses based on instrumental variables (IV) and certain sample restrictions targeted to address these potential biases. The FE model specified in Equation 1 and shown in the top row of Table 2 serves as the starting point. Table 3 presents results from the differ- ent sensitivity analysis and repeats estimates from the school FE model based on ELEMENT data for direct comparison.

The first selection process documented in the interviews refers to the non-random assignment of students who have repeated a grade to specific classrooms. While all princi- pals reported that classroom size plays an important role, some principals also men- tioned that they take potential implications for classroom culture into account. To address this potential selection problem, we treat SES composition on the class level as endogenous and instrument it with the average SES of the subset of students who never repeated a grade. This instrument is highly correlated with the total composition (the treatment indicator), and is arguably not affected by potentially non-random selection of grade repeaters because it is based only on students who never repeated a grade. The instrument should only be connected with the outcome through the actual class composition (i.e., it satisfies the exclusion restriction). Model 1 in Table 3 presents results and shows that the interaction between SES composition and female remains negative and significant. This indicates that

Legewie and DiPrete 477

selection of students who repeat a grade into specific classes does not significantly bias the estimated effects.

The second potential selection process is assignment of students to the same class who attended the same kindergarten or who were friends before entering school. Using a similar strategy as in the last sensitivity anal- ysis, we instrument peer SES by SES com- position calculated for the subset of students who either did not attend kindergarten or who skipped a grade or transferred from another school. This set of students was certainly not assigned to classrooms based on the kindergarten criterion, and students who skip a class or transfer from a different school are most likely assigned to class- rooms based on the number of students in different classrooms. For these reasons, the instrument is unaffected by the kindergarten criteria and (for the most part) by friendship self-selection. Results, presented in Model 2 of Table 3, again support our previous find- ing and indicate that the estimated causal effect is not sensitive to selection of con- nected students (either through the same kindergarten or through friendship) into the same class.

Finally, some principals reported that, in violation of school regulations, they assign students with migration backgrounds to the same class. To address this potential selection bias, we estimated the fixed-effect model reported earlier on a restricted sample. For this purpose, we assessed which schools allo- cate students with migration backgrounds non-randomly to classes, and we exclude these schools from the analysis.14 Results, presented in Model 3 of Table 3, show that self-selection of students with migration backgrounds into specific classrooms in some schools does not affect our results.

Overall, results from the targeted sensitiv- ity analyses specifically designed to address the potential selection processes identified in the interviews provide strong evidence that our estimates of gender-specific effects of classroom composition are not biased by these selection processes.

Explaining the Observed Difference in the Causal Effect between Boys and Girls

The theoretical argument presented earlier sug- gests that school context plays an important

Table 3. Sensitivity Analysis

Female SES Comp. SES Comp. x Female

Model Coef. (se) Coef. (se) Coef. (se)

FE Estimate (full sample) .007 (.02) .091* (.04) −.060** (.02)

(1) FE/IV Estimate Instrument: SES comp. of students who

never repeated a class

.008 (.02) .108* (.05) −.066** (.02)

(2) FE/IV Estimate Instrument: SES comp. of students who

did not go to kindergarten, skipped a class, or transferred to school

.009 (.02) .113 (.06) −.068* (.03)

(3) FE Estimate (restricted sample) Sample restriction: only schools that

do not allocate based on ethnicity (24 schools excluded)

.008 (.03) .117* (.05) −.052* (.03)

Note: N = 4,372. First stage results show that the two instruments are highly correlated with SES composition (i.e., the treatment). F-statistics are over 700 (highly significant), which is far above the commonly used threshold of 10. Table 1 describes control variables. * p < .05; ** p < .01; *** p < .001 (two-tailed tests); standard errors adjusted for clustering on class level.

478 American Sociological Review 77(3)

role for the size of the gender gap. An aca- demically oriented environment in schools with high-SES peers shapes how masculinity is constructed; this suppresses boys’ negative attitudes toward school, facilitates their com- mitment, and enhances students’ incentives to be engaged with academics. Other mecha- nisms, however, may account at least in part for the observed difference in the causal effect of SES composition for male and female students.

The literature on compositional school and classroom effects offers an alternative explana- tion for the relationship between SES compo- sition and student performance, which focuses on social comparison processes (Jencks and Mayer 1990; Rumberger and Palardy 2005; Thrupp, Lauder, and Robinson 2002). This alternative account argues that students use their classmates as a reference group to evalu- ate their own performance and thereby develop academic self-perceptions that in turn may affect their performance (Crosnoe 2009; Dai and Rinn 2008). To adjudicate between our proposed explanation and this alternative account, we estimate models based on ELE- MENT data that are identical to the school- level fixed-effects regression described in Equation 1, but that replace the reading score dependent variable with measures of student

attitudes, student behavior, and self-perception about academic ability.15 Our core hypothesis implies that class environment has a more pro- nounced effect on attitudes toward school, learning orientation, and academic effort for boys than for girls. Accordingly, a higher posi- tive effect of SES composition on these out- comes for boys than for girls would provide further evidence for this mechanism. An expla- nation for gender differences based on refer- ence group processes, however, would imply that a class’s SES composition affects boys’ and girls’ academic self-perceptions differ- ently. In other words, this alternative account suggests that boys and girls react differently to their reference group.

Table 4 shows results from school-level fixed-effect models of the indicated variables on classroom socioeconomic composition, controlling for variables described in Table 1. Panel A, which reports regression results using attitudes toward school, learning orien- tation, and work habits as dependent varia- bles, provides further evidence for our core hypothesis. Point estimates for SES composi- tion and the interaction with female are not all significant but consistently point in the expected direction. This pattern of results suggests that boys’ attitudes toward school, their learning orientation, and their work

Table 4. Effects of Gender and SES Composition on School-Related Attitudes and Behavior

Female SES Comp. SES Comp. x

Female

Coef. (se) Coef. (se) Coef. (se)

Panel A Attitudes toward school

.300*** (.04) .057 (.07) −.075* (.04)

Learning orientation

.131*** (.04) .044 (.07) −.033 (.03)

Work habits .166*** (.04) .147* (.07) −.086* (.04)

Panel B Self-evaluation reading

.140*** (.04) −.098 (.06) −.028 (.03)

Self-evaluation German

.207*** (.04) .012 (.08) −.056 (.03)

Self-evaluation general

−.294*** (.04) −.020 (.07) −.025 (.03)

Note: N = 4,372. Table 1 describes control variables. *p < .05; **p < .01; ***p < .001 (two-tailed tests); standard errors adjusted for clustering on class level.

Legewie and DiPrete 479

habits are more sensitive to the school envi- ronment than are girls’ attitudes and work habits. Panel B, by contrast, reports small and insignificant interaction effects between gen- der and social classroom composition on self- evaluations of performance in reading, per- formance in German, and performance “in general.” The lack of gender differences in the effect of SES composition on self-percep- tions of ability favors our preferred explana- tion over the alternative account based on reference group processes.

We further extend this examination of mechanisms by building on the initial FE model for 5th-grade performance (defined in Equation 1), and add school-related attitudes and behavior as independent variables in a stepwise fashion. Compared to the models presented so far, the elaborated model is less rigorous from a causal point of view because the causal ordering of performance and school-related attitudes and behavior is not clear-cut. Nonetheless, it can be informative about potential mechanisms. Results in Table 5 suggest that the effect of SES composition is clearly reduced by the addition of variables for school-related attitudes and behavior (Model 2). They also suggest that part of the gender difference in the effect of SES compo- sition (33 percent) may be explained by its

gender-specific effect on school-related atti- tudes and behavior; this provides further sup- port for our proposed mechanism.

Finally, we investigate the possibility that boys benefit from a stronger academic peer culture not because they are boys, but rather because underperforming students benefit in general, and boys are a disproportionate frac- tion of underperforming students. Accord- ingly, we again extend the model described in Equation 1 by adding an interaction term between performance in 4th grade (the year prior to our measured outcomes in the regres- sions) and SES composition in 5th grade. Results (available from the authors) show that the impact of SES composition is signifi- cantly stronger for low-performing students, which is in line with findings from other stud- ies (Bryk, Lee, and Holland 1993; Coleman 1966, 1970). Inclusion of this interaction also weakens the direct benefit of being male in a high-SES class by about 27 percent (from –.060 to –.044). However, the interaction between SES composition and gender remains statistically significant ( p = .021) and sub- stantively important. Results suggest that boys do indeed benefit indirectly from a stronger academic climate because they are disproportionately low-performing students. Nonetheless, the bulk of the effect stems from

Table 5. Fixed-Effects Models with School-Related Attitudes and Behavior

Model 1 Model 2 Model 3

Variable Coef. (se) Coef. (se) Coef. (se)

Female .007 (.02) −.009 (.03) −.009 (.03) SES Composition .091* (.04) .037 (.04) .033 (.04) SES Composition x Female −.060** (.02) −.061** (.02) −.040* (.02) Attitude Toward School .041* (.01) .040* (.02) Learning Orientation .019 (.01) .017 (.02) Work Habits .072*** (.01) .092*** (.01) Attitude Toward School x Female −.047* (.02) Learning Orientation x Female .002 (.02) Work Habits x Female −.059*** (.02) Control Variables yes yes yes Constant −.673*** (.35) −.538 (.27) −.533* (.25)

Note: N = 4,372. Table 1 describes control variables. *p < .05; **p < .01; ***p < .001 (two-tailed tests); standard errors adjusted for clustering on class level.

480 American Sociological Review 77(3)

boys’ greater sensitivity than girls to class- rooms’ academic orientation.

DiSCuSSion Throughout the industrialized world, girls have made dramatic gains in educational attainment, while boys’ underperformance and their tendency to disrupt the learning process have sparked intense academic and public debates about the causes of what many now call the “problem with boys.” Some scholars and pundits blame schools for foster- ing a de-masculinized learning environment. Yet, the role of school context and the con- nection between school resources and the gender gap is underdeveloped in the literature to date. In this article, we have extended research on schools’ effect on class and race inequality by asking whether schools affect gender inequality as well, and if so, what are the mechanisms by which this occurs.

Building on theories about gender identity, adolescent culture, and prior ethnographic classroom observations, we developed a theo- retical argument about the role of environ- mental factors for the educational gender gap and boys’ underachievement. In particular, we argue that school and class environments shape how masculinity is constructed in peer culture and thereby influence boys’ orienta- tion toward school. Resources that create a learning-oriented environment raise the valu- ation of academics in adolescent male culture and facilitate commitment. Girls’ peer groups, by contrast, do not vary as strongly with the social environment in the extent to which they encourage academic engagement, and they are less likely to stigmatize school engagement as un-feminine. As a conse- quence, boys differentially benefit from these school resources, and the female advantage in test scores shrinks in higher quality schools.

Results from our analysis of German ELE- MENT and PISA-I-Plus 2003 data provide clear support for this hypothesis. We find substantial variation in the gender gap in aca- demic performance across schools and that this variation is related to average school perfor- mance. We then used a quasi-experimental

research design to establish that boys are more sensitive to peer SES composition as an important dimension of school quality related to the learning environment. This quasi- experimental research design is based on the argument that randomness plays an important role for student assignment to classes within Berlin’s elementary and 5th-grade upper- secondary schools. To evaluate this argument, we examined Berlin’s school regulations, compared observed classroom composition with simulations of random assignment, and conducted qualitative interviews with school principals in Berlin. Findings from this evalu- ation of the selection process generally sup- port our argument but also point at potential biases, which we addressed with targeted sensitivity analyses. Results from these anal- yses show little effect of potential selection biases on our core results. In addition, we considered alternative mechanisms that might explain the observed difference in the causal effect between boys and girls. Results from this analysis provide further support for our own explanation. They suggest that boys benefit indirectly (because low-performing students benefit in general) and directly (because the effect is bigger for boys than for girls) from being in a classroom with high- SES composition.

Our findings contribute to several areas of research. First, our study makes an important contribution to the debate about boys’ well- publicized underperformance. The outlined cultural mechanism explains why boys are more sensitive to human and cultural capital resources in schools, which plays an impor- tant role for their underperformance and the gender gap in educational achievement. This argument suggests that boys’ resistance to school is not purely a function of either their class background—as many studies sug- gest—or the fact of their masculinity—as other research suggests—but instead depends on schools’ and classrooms’ local cultural environment. As such, the findings broaden our understanding of boys’ notorious under- performance. Results point at an important mechanism connected to how school and class environments shape boys’ and girls’

Legewie and DiPrete 481

learning orientations, and in the process reveal a pattern similar to what has been found in families (Buchmann and DiPrete 2006). In both cases, boys seem to be more sensitive to the level of resources in the local environment, so that the size of the gender gap is a function of environmental resources.

Second, our results point to useful direc- tions for new research on policies to raise boys’ achievement levels. It is obviously important to know that boys respond espe- cially positively to an academic orientation among their peers. However, while local gov- ernments could decide to invest more resources in their schools, they cannot, as a practical matter, produce more high-SES chil- dren for their school systems. An important unanswered question raised by our research concerns whether schools can accomplish the same cultural enrichment through alternative means. The most obvious alternative resource would be better teachers. Teachers directly influence schools’ academic environment and can raise academic performance. They have the potential to modify student behavior and produce a stronger academic student culture, even without socioeconomic enrichment of a school’s student body. At present, however, too little is known about what makes a quality teacher, or the extent of the effect of better teachers on higher academic performance and the academic climate. Our research suggests, for example, that teaching methods that emphasize academic competition are particu- larly beneficial for boys. These are important questions for future research.

Finally, we make a methodological contri- bution to the literature on estimation of causal effects. Our work illustrates how a detailed study of the relevant selection process—in our case, examination of official regulations, statistical simulations, and qualitative inter- views—can facilitate the estimation of causal effects. This detailed understanding of the actual selection process not only allows researchers to evaluate the extent of bias but also enables the design of targeted sensitivity analysis (in our case, based on instrumental variables and sample restrictions). Overall, we believe that knowledge about the selection

process can help researchers improve the accuracy of causal effect estimates. Consider- ing these benefits, we invite sociologists to take selection processes seriously as an inde- pendent object of study—an argument previ- ously made by Sampson (2008:189), who conceptualizes “selection bias as a fundamen- tal social process worthy of study in its own right rather than a statistical nuisance” (for an earlier statement of this argument, see DiPrete 1993).

Our findings are limited in some regards. Most important, our theoretical argument applies to all kinds of school resources that create a learning-oriented environment. Our empirical analysis, however, only focuses on one (although important) dimension, peer socioeconomic composition. Given this limi- tation, future studies should establish the extent to which conclusions from this study apply to other kinds of school-based resources. Additionally, due to lack of adequate data, our study neglects teachers’ role in shaping boys’ and girls’ learning orientations. While our interviews indicate that teachers are not assigned to classrooms based on classroom composition, teachers might react to class- room dynamics in certain ways that play an important role for the processes studied here. Finally, our study focuses on only one major dimension of cognitive achievement, reading. On average, boys do as well or better than girls in mathematics, and the male advantage is larger on the right tail of the distribution. Whether boys gain a stronger advantage than girls from being in a classroom with higher mean SES, or whether their special advantage occurs only for academic subjects where they otherwise lag behind girls, is an important question for future research.

APPEnDix Part A. Education and the Educational Gender Gap in Germany

Although our main focus here is the theoreti- cal argument, background information can help contextualize findings from the German case. In Germany, children usually attend

482 American Sociological Review 77(3)

elementary school from age 6 until age 10 or 12, depending on state (Bundesland) regula- tions. After finishing elementary school, stu- dents transfer to secondary schools, which are distinct from U.S. middle and high schools because of performance-based tracking on the school level. Although the system has become more differentiated in recent decades, three school types have traditionally been of great importance. The Gymnasium is the highest secondary school type, the Realschule is for intermediate students, and the Hauptschule is the low secondary school track. As a response to critiques of this tripartite secondary school system, some states have introduced compre- hensive schools that either integrate all three school tracks or just the Haupt- and Realschule (Gesamtschule and Schule mit mehreren Bildungsgängen). After finishing secondary school, students have the option to obtain a higher education degree, to continue their education in one of the vocational programs (which figure importantly in the German edu- cational system), or to enter the labor market immediately. Overall, the German educational system is distinct from the U.S. system and other countries primarily because of the early school-based tracking in secondary school, the strong vocational track as an alternative to higher education, and the limited role of the federal government, which is evident in the many differences in the specific structure of German schools across German states. Similar to other industrialized countries, the gender gap in Germany has closed over the past decades. Legewie and DiPrete (2009) empha- size, however, that the female advantage in higher education is less pronounced than in the United States due in large part to women’s failure to converge with men in rates of obtaining degrees from Fachhochschulen (universities of applied sciences).

Part B. Simulation of Random Assignment

Our simulation-based approach allows us to evaluate whether within-school variation in the composition of classes is consistent with a

random allocation process. To compare observed composition with composition obtained under complete randomization, we proceed in the following way. For each school, we randomly allocate students to classrooms in the school they attend, keeping the number and size of classrooms constant. We then compare socioeconomic composi- tion across classes obtained from the simu- lation with the observed composition. Accordingly, the simulation evaluates whether the actual (unknown) allocation process is consistent with a completely randomized classroom assignment. The statistic to com- pare the actual and simulated distribution for some variable x (e.g., SES, migration back- ground, or gender) for classroom k in school j is defined as the average square deviation of the classroom means from the school mean

t n

x x k

n

j j

jk j

j = −( )

= ∑

1

1

2

where j and k are indices for schools and classrooms, respectively; x–j is the average for school j; x–jk is the average for classroom k in school j; and nj is the number of class- rooms in school j. If the number of students is the same in each classroom within a school, this measure is simply the variance of the class specific means in a school.

Acknowledgments We thank Claudia Buchmann, Jennifer L. Jennings, Merlin Schaeffer, and participants at seminars at the Mailman School, Columbia University, the University of Wisconsin, Madison, and the 2010 RC28 conference in Haifa, Israel for helpful comments and suggestions. We thank the Institut zur Qualitätsentwicklung im Bildung- swesen (IQB) for providing access to data and Prof. Rainer Lehmann for his help with ELEMENT data.

notes 1. Of course, stereotypical gender identities also affect

girls. Correll (2001), for example, shows how cul- tural beliefs about gender can bias women’s self-perceptions of math ability, controlling for actual performance, and thereby deter women from careers in science, math, or engineering.

Legewie and DiPrete 483

2. These assertions do not imply that girls are always engaged in the learning process. Many studies docu- ment how girls resist teachers and school (e.g., Francis 2000). Nevertheless, one of the most common findings in ethnographic studies is that boys more actively resist the learning process.

3. Our expectations mainly relate to wealthy OECD countries because prior research finds that the role of school context (Chudgar and Luschei 2009) and gender relations differ substantially between high- and low-income countries.

4. In contrast to most other states in Germany, students in Berlin usually attend elementary school until 6th grade, so the 31 5th-grade upper-secondary schools— the grundständige Gymnasien—are different from normal secondary schools.

5. In Berlin, elementary school students assigned to the same classroom take virtually all their classes together, so we use the terms “classroom” and “class” interchangeably in the text.

6. For 5th-grade upper-secondary schools in ELE- MENT, class assignment occurs in 5th grade because students transfer from elementary school after 4th grade.

7. We obtained both datasets from the Forschungs- datenzentrum at the Institute für Qualitätsentwicklung im Bildungswesen (IQB) HU-Berlin.

8. As a substantive matter, 5th-grade culture differs from 9th-grade culture in the obvious sense that 5th- grade students are pre-adolescent while 9th-grade students have generally passed through puberty. At the same time, studies of childhood and adolescent culture find continuity in the emerging masculine culture between middle childhood and high school (Maccoby 1998; Thorne 1993). Thus, for substantive and methodological reasons, we expect comparison of results from 5th and 9th grades to be informative about our core hypothesis.

9. The three-level data structure might imply that error terms of students in the same classroom are corre- lated even after controlling for school fixed-effects. We address this problem by correcting the standard error for clustering on the class level using the Moulton factor (Angrist and Pischke 2008).

10. Although the PISA-I-Plus is a panel study and col- lected achievement data in 9th and 10th grades, the panel component of these data is not yet available.

11. In addition, all schools reported that class changes within a grade level are extremely rare, and resources are generally allocated equally across classes.

12. We explored alternative specifications of SES com- position effects, such as allowing separate effects of SES composition of male and female peers. These alternative specifications yield essentially the same results as those reported in the tables.

13. One additional school year corresponds to the esti- mated test score difference between 5th and 6th grade in the ELEMENT dataset.

14. We use a simple z-test to identify schools in which the difference in the proportion of students with migration backgrounds between classes is higher than what we would expect under randomness. Using conservative criteria, we exclude schools with a p-value smaller than .1 (24 schools).

15. We constructed the measures from a range of indica- tors using exploratory factor analysis (see the online supplement).

references Ammermueller, Andreas and Jörn-Steffen Pischke. 2009.

“Peer Effects in European Primary Schools: Evidence from the Progress in International Reading Literacy Study.” Journal of Labor Economics 27:315–48.

Angrist, Joshua D. and Jörn-Steffen Pischke. 2008. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton, NJ: Princeton University Press.

Bishop, John H., Matthew Bishop, Lara Gelbwasser, Shanna Green, and Andrew Zuckerman. 2003. “Nerds and Freaks: A Theory of Student Culture and Norms.” Brookings Papers on Education Policy 6:141–99.

Bryk, Anthony, Valerie Lee, and Peter Holland. 1993. Catholic Schools and the Common Good. Cambridge, MA: Harvard University Press.

Buchmann, Claudia and Thomas A. DiPrete. 2006. “The Growing Female Advantage in College Comple- tion: The Role of Family Background and Academic Achievement.” American Sociological Review 71: 515–41.

Buchmann, Claudia, Thomas A. DiPrete, and Anne McDaniel. 2008. “Gender Inequalities in Education.” Annual Review of Sociology 34:319–37.

Campbell, Jay R., Dana L. Kelly, Ina V. Mullis, Michael O. Martin, and Marian Sainsbury. 2001. Framework and Specification for PIRLS Assessment 2001. Chest- nut Hill, MA: International Study Center, Lynch School of Education, Boston College.

Chudgar, Amita and Thomas F. Luschei. 2009. “National Income, Income Inequality, and the Importance of Schools: A Hierarchical Cross-National Comparison.” American Educational Research Journal 46:626–58.

Cohen, Michele. 1998. “A Habit of Healthy Idleness: Boys Underachievement in Historical Perspective.” Pp. 19–33 in Failing Boys? Issues in Gender and Achievement, edited by D. Epstein, J. Elwood, V. Hey, and J. Maw. Buckingham, UK: Open University Press.

Coleman, James S. 1961. The Adolescent Society: The Social Life of the Teenager and its Impact on Educa- tion. New York: Free Press.

Coleman, James S. 1966. Equality of Educational Oppor- tunity. Washington, DC: U.S. Dept. of Health, Educa- tion, and Welfare, Office of Education.

Coleman, James S. 1970. “Reply to Cain and Watts.” American Sociological Review 35:242–49.

484 American Sociological Review 77(3)

Correll, Shelley J. 2001. “Gender and the Career Choice Process: The Role of Biased Self-Assessments.” American Journal of Sociology 106:1691–1730.

Crosnoe, Robert. 2009. “Low-Income Students and the Socioeconomic Composition of Public High Schools.” American Sociological Review 74:709–730.

Dai, David Yun and Anne N. Rinn. 2008. “The Big-Fish- Little-Pond Effect: What Do We Know and Where Do We Go from Here?” Educational Psychology Review 20:283–317.

Davies, Bronwyn. 2003. Frogs and Snails and Feminist Tales: Preschool Children and Gender. New York: Hampton Press.

DiPrete, Thomas A. 1993. “Discrimination, Choice, and Group Inequality: A Discussion of How Allocative and Choice-Based Processes Complicate the Standard Decomposition.” Social Science Research 22:415–40.

Dresel, Markus, Heidrun Stöger, and Albert Ziegler. 2006. “Klassen- und Schuleffekte bei Geschlechtsun- terschieden im schulischen Leistungsbereich.” Psy- chologie in Erziehung und Unterricht 1:44–61.

Eitzen, D. Stanley. 1975. “Athletics in the Status System of Male Adolescents: A Replication of Coleman’s ‘The Adolescent Society.’” Adolescence 10:267–76.

Epstein, Debbie. 1998. “Real Boys Don’t Work: Under- achievement, Masculinity and the Harassment of Sis- sies.” Pp. 96–108 in Failing boys? Issues in Gender and Achievement, edited by D. Epstein, J. Elwood, V. Hey, and J. Maw. Buckingham, UK: Open University Press.

Francis, Becky. 2000. Boys, Girls and Achievement: Addressing the Classroom Issues. London: Routledge Falmer.

Ganzeboom, Harry B. G., Donald J. Treiman, and Wout C. Ultee. 1991. “Comparative Intergenerational Strat- ification Research: Three Generations and Beyond.” Annual Review of Sociology 17:277–302.

Gould, Eric D., Victor Lavy, and M. Daniele Paserman. 2009. “Does Immigration Affect the Long-Term Edu- cational Outcomes of Natives? Quasi-Experimental Evidence.” The Economic Journal 119:1243–69.

Greenwald, Rob, Larry V. Hedges, and Richard D. Laine. 1996. “The Effect of School Resources on Student Achievement.” Review of Educational Research 66:361–96.

Hallinan, M. T. 1988. “Equality of Educational Opportu- nity.” Annual Review of Sociology 14:249–68.

Hanushek, Eric A. 1989. “The Impact of Differential Expenditures on School Performance.” Educational Researcher 18:45–62.

Howe, Christine. 1997. Gender and Classroom Inter- action: A Research Review. Glasgow: The SCRE Centre.

Hoxby, Caroline. 2000. Peer Effects in the Classroom: Learning from Gender and Race Variation. NBER Working Paper No. 7867.

Imberman, Scott, Adriana Kugler, and Bruce Sacerdote. 2009. “Katrina’s Children: A Natural Experiment

in Peer Effects from Hurricane Evacuees.” NBER Working Paper No. 15291.

Jencks, Christopher and Susan E. Mayer. 1990. “The Social Consequences of Growing Up in a Poor Neighborhood.” Pp. 111–86 in Inner-City Poverty in the United States, edited by L. E. Lynn and M. G. H. McGeary. Washington, DC: National Academy Press.

Jennings, Jennifer L. and Thomas A. DiPrete. 2010. “Teacher Effects on Social/Behavioral Skills in Early Elementary School.” Sociology of Education 83:135–59.

Jovanovic, Jasna and Sally Steinbach King. 1998. “Boys and Girls in the Performance-Based Science Class- room: Who’s Doing the Performing?” American Educational Research Journal 35:477–96.

Kahlenberg, Richard D. 2001. All Together Now: Cre- ating Middle-Class Schools through Public School Choice. Washington, DC: Brookings Institution Press.

Kao, Grace, Marta Tienda, and Barbara Schneider. 1996. “Racial and Ethnic Variation in Academic Perfor- mance.” Research in Sociology of Education and Socialization 11:263–97.

Legewie, Joscha and Thomas A. DiPrete. 2009. “Family Determinants of the Changing Gender Gap in Edu- cational Attainment: A Comparison of the U.S. and Germany.” Schmollers Jahrbuch (Journal of Applied Social Science Studies) 129:1–13.

Lehmann, Rainer and Jenny Lenkeit. 2008. ELEMENT: Abschlussbericht über die Untersuchungen 2003, 2004 und 2005 an Berliner Grundschulen und grund- ständigen Gymnasien. Berlin: Humboldt Universität zu Berlin.

Mac an Ghaill, Mairtin. 1994. The Making of Men: Mas- culinities, Sexualities and Schooling. Buckingham, UK: Open University Press.

Maccoby, Eleanor E. 1998. The Two Sexes: Growing up Apart, Coming Together. Cambridge, MA: Belknap Press of Harvard University Press.

Machin, Stephen and Sandra McNally. 2005. “Gen- der and Student Achievement in English Schools.” Oxford Review of Economic Policy 21:357–72.

MacLeod, Jay. 2008. Ain’t No Makin’ It: Aspirations and Attainment in a Low-Income Neighborhood. Oxford, UK: Westview Press.

Morris, Edward W. 2008. “‘Rednecks,’ ‘Rutters,’ and ’Rithmetic.” Gender & Society 22:728–51.

Morse, Susan. 1998. Separated by Sex: A Critical Look at Single-Sex Education for Girls. Washington, DC: American Association of University Women Educa- tional Foundation.

Murnane, Richard J. 1983. “Quantitative Studies of Effec- tive Schools: What Have We Learned?” Pp. 193–209 in School Finance and School Improvement, edited by A. Odden and L. D. Webb. Cambridge: Ballinger.

Pickering, Jon. 1997. Raising Boys’ Achievement. Lon- don: Network Continuum Education.

PISA-Konsortium Deutschland. 2006. PISA 2003: Unter- suchungen zur Kompetenzentwicklung im Verlauf eines Schuljahres. Münster: Waxmann Verlag.

Legewie and DiPrete 485

Power, Sally, Geoff Whitty, Tony Edwards, and Valerie Wigfall. 1998. “Schoolboys and Schoolwork: Gender Identification and Academic Achievement.” Inter- national Journal of Inclusive Education 2:135–53.

Quenzel, Gudrun and Klaus Hurrelmann. 2010. “Geschlecht und Schulerfolg: Warum wird die schul- ische Leistungsbilanz von jungen Männern immer schlechter?” Kölner Zeitschrift für Soziologie und Sozialpsychologie 62:61–91.

Ravitch, Diane. 2010. The Death and Life of the Great American School System: How Testing and Choice Are Undermining Education. New York: Basic Books.

Rockoff, Jonah E. 2004. “The Impact of Individual Teachers on Student Achievement: Evidence from Panel Data.” American Economic Review 94:247–52.

Rumberger, Russell W. and Gregory J. Palardy. 2005. “Does Segregation Still Matter? The Impact of Stu- dent Composition on Academic Achievement in High School.” Teachers College Record 107:1999–2045.

Sadker, David and Karen Zittleman. 2009. Still Failing at Fairness: How Gender Bias Cheats Girls and Boys in School and What We Can Do About It. New York: Scribner.

Salisbury, Jonathan and David Jackson. 1996. Challeng- ing Macho Values: Practical Ways of Working with Adolescent Boys. London: Falmer Press.

Salomone, Rosemary C. 2003. Same, Different, Equal: Rethinking Single-Sex Schooling. New Haven, CT: Yale University Press.

Sampson, Robert J. 2008. “Moving to Inequality: Neigh- borhood Effects and Experiments Meet Social Struc- ture.” American Journal of Sociology 114:189–231.

Schneider, Barbara L. and Venessa A. Keesler. 2007. “School Reform 2007: Transforming Education into a Scientific Enterprise.” Annual Review of Sociology 33:197–217.

Schöps, Katrin, Oliver Walter, Karin Zimmer, and Man- fred Prenzel. 2004. “Disparitäten zwischen Jungen und Mädchen in der mathematischen Kompetenz.” In PISA 2003: Der Bildungsstand der Jugendlichen in Deutschland- Ergebnisse des zweiten internationalen Vergleichs, edited by PISA-Konsortium Deutschland. Münster: Waxmann.

Sewell, William H., Archibald O. Haller, and Alejandro Portes. 1969. “The Educational and Early Occupa- tional Attainment Process.” American Sociological Review 34:82–92.

Skelton, Christine. 1997. “Primary Boys and Hegemonic Masculinities.” British Journal of Sociology of Edu- cation 18:349–69.

Sørensen, Aage B. and Stephen L. Morgan. 2006. “School Effects: Theoretical and Methodological Issues.”

Pp. 137–60 in Handbook of the Sociology of Educa- tion, edited by M. T. Hallinan. New York: Springer.

Spender, Dale. 1982. Invisible Women: The Schooling Scandal. London: Writers and Readers.

Stanworth, Michelle. 1984. Gender and Schooling: A Study of Sexual Divisions in the Classroom. London: Routledge.

Steinberg, Laurence D., L. Steinberg, Benson Bradford Brown, and Sanford M. Dornbusch. 1997. Beyond the Classroom: Why School Reform Has Failed and What Parents Need to Do. New York: Simon and Schuster.

Thorne, Barrie. 1993. Gender Play: Girls and Boys in School. Piscataway, NJ: Rutgers University Press.

Thrupp, Martin, Hugh Lauder, and Tony Robinson. 2002. “School Composition and Peer Effects.” Interna- tional Journal of Educational Research 37:483–504.

Whitmire, Richard. 2010. Why Boys Fail: Saving Our Sons from an Educational System That’s Leaving Them Behind. New York: AMACOM.

Willis, Paul. 1981. Learning to Labor: How Work- ing Class Kids Get Working Class Jobs. New York: Columbia University Press.

Younger, Michael, Molly Warrington, and Jacquetta Wil- liams. 1999. “The Gender Gap and Classroom Inter- actions: Reality and Rhetoric?” British Journal of Sociology of Education 20:325–41.

Joscha Legewie is a PhD student in the Department of Sociology at Columbia University. His dissertation research focuses on how school and peer environment shape gender inequalities in education using two quasi- experimental case studies from Germany and the United States. Other research looks at the persisting gender gap in science and engineering education, the role of peer relations and social support for academic work in school, and the effect of terrorist events on the perception of immigrations using a natural experiment.

Thomas A. DiPrete is Giddings Professor of Sociol- ogy at Columbia University. His research interests include social stratification, demography, education, economic sociology, and quantitative methodology. Recent research projects include causes of the widening gender gap in higher education in favor of women, the persisting gender gap in science and engineering educa- tion, social polarization in the United States and its link with segregation in social networks along several poten- tial dimensions of social cleavage, and the role of social comparison and cumulative advantage processes in the trend toward rising inequality at the top of the earnings distribution.