The Efficiency of Race-Neutral Alternatives to Race-Based Affirmative Action: Evidence from Chicago’s Exam Schools∗ Glenn Ellison Parag A. Pathak† August 2016 Abstract Several public K-12 and university systems have recently shifted from race-based affirmative action plans to race-neutral alternatives. This paper explores the degree to which race-neutral alternatives are effective substitutes for racial quotas using data from the Chicago Public Schools (CPS),wherearace-neutral, place-basedaffirmativeactionsystemisusedforadmissionsathighly competitiveexamhighschools. Wedevelopatheoreticalframeworkthatmotivatesquantifyingthe efficiency cost of race-neutral policies by the extent admissions decisions are distorted more than needed to achieve a given level of diversity. According to our metric, CPS’s race-neutral system is 24%and20%efficientasatoolforincreasingminorityrepresentationatthetoptwoexamschools, i.e. about three-fourths of the reduction in composite scores could have been avoided by explicitly consideringrace. EventhoughCPS’ssystemisbasedonsocioeconomicdisadvantage, itisactually less effective than racial quotas at increasing the number of low-income students. We examine several alternative race-neutral policies and find some to be more efficient than the CPS policy. Whatisfeasiblevarieswiththeschool’ssurroundingneighborhoodcharacteristicsandthetargeted level of minority representation. However, no race-neutral policy restores minority representation topriorlevelswithoutsubstantialinefficiency,implyingsignificantefficiencycostsfromprohibitions on affirmative action policies that explicitly consider race. ∗WethankKatieEllis,SusanRyanandthestaffatChicagoPublicSchoolsfortheirexpertiseandhelpwiththedata. Vivek Bhattacharya, Alex Olssen, Rahul Singh provided excellent research assistance. We’ve also benefitted from the assistanceofundergraduatesJackieBredenbergandBrandonEnriquez. EllisonacknowledgessupportfromtheToulouse Network for Information Technology. Pathak acknowledges support of National Science Foundation grant SES-1056325 and SES-1426566. Pathak is a member of the scientific advisory board of the Institute for Innovation in Public School Choice. †Department of Economics, Massachusetts Institute of Technology, Cambridge MA 02142 and NBER, e-mail: gelli- [email protected] and [email protected] 1 1 Introduction Affirmative action is one of the most contentious issues in American social policy, nowhere more so than in the context of admission to selective educational institutions. Originally a 1965 executive order requiring federal contracts to “take affirmative action” to ensure that minorities and women are employed where available, affirmative action led to increasingly widespread efforts to boost minority participation in public K-12 school systems and universities in the 1970s and 1980s. More recently, such plans have come under attack on two fronts. Some states have banned race-based affirmative action.1 And in 2003 the US Supreme Court ruled in Grutter vs. Bollinger that “strict scrutiny” must be applied to race-based plans: they must serve a compelling government interest that cannot be effectively achieved in a race-neutral manner. A June 2007 Supreme Court decision applied the earlier decision to strike down race-based admissions plans for public schools in Seattle and Jefferson County, Kentucky.2 Against this backdrop, a number of public K-12 and university systems have shifted from race-based affirmative action plans to race-neutral alternatives.3 This paper explores the degree to which race-neutral alternatives are an effective substitute for racial quotas using data from Chicago Public Schools (CPS). CPS has adopted a new race-neutral admission policy for their selective high schools, including the district’s flagship Northside and Walter Paytonhighschools. From1980-2007,aconsentdecreemandatedrace-basedadmissionsforallChicago magnet and selective enrollment schools. Admissions decisions were and are based on a composite scorethatcombinesmiddleschoolgrades, standardizedISATtestscores, andaspecialentranceexam. Through 2009, the admissions procedure employed racial quotas that placed a cap on the number of white students in each school. In2010, CPSreplaced this systemwith a place-basedaffirmativeaction systemwheredisadvantageismeasuredbyaneighborhoodproxyofsocioeconomicstatus. Eachcensus tract in the city was assigned to one of four tiers using an index combining five variables: median family income; a measure of adult educational attainment; home ownership rates; and the prevalence ofsingle-parenthouseholdsandnon-nativeEnglishspeakers. Schoolsfirstfilled40%oftheirslotswith the applicants having the highest composite scores. The remaining 60% of slots were then filled by 1This includes constitutional amendments passed by ballot initiative in California in 1996 and Michigan in 2006. 2Department of Education guidelines suggest that a school district could draw attendance zones based on the racial composition of particular neighborhoods, as well as on race-neutral factors such as the average household income and averageparentaleducationlevelofparticularneighborhoodswithintheschooldistrict,providedthatallstudentswithin those zones would be treated the same regardless of their race (OCR 2011). 3Kahlenberg(2003)describessocioeconomicadmissionscriteriaacrossanumberofdistricts,whileKahlenberg(2008) lists 60 US school districts using socioeconomic status as a factor in student assignment. 2 dividing the slots equally across the four tiers and filling the slots with the highest-scoring remaining students living in census tracts belonging to each tier. The system was modified in 2012: the number of tier-reserved slots was increased from 60% to 70% and a sixth variable (test scores in the local elementary school) was added to the index. Our focus on Chicago is motivated by several reasons. First, Chicago’s new plan is held up as a national model for achieving racial and ethnic diversity in selective public schools (e.g., Kahlenberg (2014)).4 Followingthe2007USSupremeCourtdecision,federalDepartmentofJusticeandEducation Guidelines (OCR 2011) describe acceptable alternatives for achieving diversity at selective schools, which mirror Chicago’s plan: For students who meet the basic admissions criteria, a school district could give greater weight to the applications of students based on their socioeconomic status, whether they attend underperforming feeder schools, their parents’ level of education, or the average income level of the neighborhood from which the student comes... Second, Chicago employs an affirmative action scheme with explicit rules, which can be analyzed quantitatively. In particular, Chicago’s assignment mechanism is a variant of the student-proposing deferred acceptance algorithm, which is strategy-proof for participants (Dubins and Freedman 1981, Roth 1982). This property motivates simulating alternative assignment rules ignoring how these alternative policies might change how applicants rank schools. Third, the fact that Chicago’s schools are differently situated (with respect to neighborhood segregation) provides an opportunity to explore the effectiveness of race-neutral plans across settings. Finally, Chicago’s exam schools are highly visible and are frequently included in lists of the best U.S. public high schools, which makes them of independent interest. Legal arguments about the effectiveness of race-neutral policies often solely focus on how minority representation would change when race-based policies are replaced. This emphasis is misguided, be- causeitisalwayspossibletoeliminateracialgapsinadmissionsinarace-neutralmannerbyadmitting students completely at random. However, elite schools do not randomly admit students because it conflicts with the main reason for their existence: they provide advanced students access to programs tailored to their knowledge and abilities. Therefore, assessing the “effectiveness” of race-neutral poli- cies inherently requires measuring two components: whether the policy provides desired diversity (or 4Similar place-based schemes also exist in other affirmative actions systems for colleges and universities, including in Israel (Alon and Malamud 2014). For U.S. college admissions, Cashin (2014) argues for prioritizing applicants from disadvantaged neighborhoods rather than by race. 3 other) benefits and how race-neutral restrictions prevent admitting the students schools would most like to admit. To motivate our analysis, Table 1 lists admissions outcomes for selected pairs of students. The top student in each pair is a higher-scoring applicant who is denied admission under the CPS tier plan, while the bottom student in the pair is a lower-scoring student who is admitted. In the first two pairs, underrepresented minority students are denied admission despite having perfect grades in 7th grade, perfect scores on the specialized school entrance exam, and very high ISAT scores. Meanwhile, the CPSprocedureacceptedstudentswithmuchlessimpressiverecords: thecomparisonHispanicstudent has a much lower score on the entrance exam and the comparison black admittee had worse grades and a somewhat lower entrance exam score. In these cases, the admitted students were admitted because they live in lower SES census tracts. The next two pairs provide examples in which the student denied admission qualifies for free/reduced lunch, while the lower-achieving student does not. The high-achieving students in these examples would be admitted if one eliminated affirmative action. They would also be admitted under a race-based policy: when minorities are added more efficiently one does not need to displace as many white and Asian-american students. Race-based policies are also better at accepting minority students with scores just below the cutoffs. The top students in the fifth and sixth pairs are such examples. In each case, the high-achieving students denied admission are free/reduced lunch eligible, yet are missed by the CPS race-neutral policy. The comparison students are admitted because they live in lower-SES census tracts, but are not themselves poor enough to qualify for a subsidized lunch. While these examples are extreme, they clearly illustrate inefficiencies with race-neutral admissions. The paper proposes a methodology to characterize inefficiency and uses it to systematically examine race-neutral admissions in Chicago. Section 2 develops a simple model that motivates our measure of the effectiveness of an affirmative action plan. Students are assumed to benefit both from a curriculum tailored to their knowledge and ability, and from learning within a diverse student body. The optimal admissions policy is race- conscious and can be implemented either as a racial quota or a bonus scheme that gives some number of points to members of the underrepresented group. If race cannot be considered, the effectiveness of policiesbasedonproxiescorrelatedwithracedependsontherelationshipbetweenapplicantattributes and race. We show, via an example, that it is possible that even a highly correlated proxy may be completelyineffective. Ourmainresultshowsthatthesocialwelfareproducedbyanypolicy,nomatter howcomplex, isafunctionoftwosimplepropertiesoftheallocation: theaveragebaselineachievement 4 Free Score Components Comp. Application Student Race Lunch GPA ISAT Exam Score School Admit? A Hispanic No 4.0 96.3 100 98.8 Payton No B Hispanic No 4.0 94.0 77.7 90.6 Payton Yes C Black No 4.0 96.3 100 98.8 Payton No D Black No 3.25 100 93 89.3 Payton Yes E Asian Yes 4.0 98.0 99 99.0 Payton No F Asian No 4.0 100 85 95.0 Payton Yes G White Yes 4.0 98.0 99 99.0 Payton No H White No 3.75 93.0 97 93.9 Payton Yes I Hispanic Yes 4.0 96.0 95 97.0 Payton No J Hispanic No 4.0 94.0 77.7 90.6 Payton Yes K Black Yes 4.0 99.3 89 96.1 Northside No L Black No 4.0 94.3 76.7 90.3 Northside Yes Table 1: Examples of the efficiency costs of race-neutral affirmative action of admitted students and minority representation. There is a Pareto frontier of efficient policies in this two-dimensional space. Race-neutral policies lie inside the frontier. We define a welfare-motivated notion of relative efficiency that represents the policies’ distance to the efficient frontier. Section 3 uses data from CPS to estimate the shape of the Pareto frontier, which represents the fundamental constraints for affirmative action policies. We report on CPS’s two most selective selective-entry high schools: Walter Payton College Preparatory High School (Payton) and Northside College Preparatory High School (Northside). Focusing on the two most selective schools allows us to examine where affirmative action policies might have the largest impact on assignments. The schools also attract different applicant pools due to their geographic location, allowing us to examine whether race-neutral policies perform better in some situations. Payton’s central location is appealing to students living in a number of predominantly low-income black and Hispanic neighborhoods, whereas Northside is far from most predominantly black neighborhoods and would naturally draw more of its poor and minority students from surrounding middle-class neighborhoods. We first note the percentage of underrepresented minority students that each school would have if admissions were based solely on the “composite score” that CPS currently uses. At both schools, minorityrepresentationcanbeincreasedsubstantiallyfromitslevelunderapurelyscore-basedadmis- 5 sions scheme with only a slight decrease in average composite scores, but score declines become much steeperasminorityrepresentationisfurtherincreased. Sincethenewpolicyisbasedonsocioeconomic factors, we also report how policies affect the number of low-income students qualifying for free- and reduced-price lunch at each school and consider Pareto frontiers relevant when both minority- and low-income representation are concerns. Section4examinesChicago’ssocioeconomic-basedaffirmativeactionplanandvariantsthatreserve more or fewer slots for students from low-SES areas. Our primary focus is on the relative efficiency of the CPS plan as a race-neutral method for increasing minority representation. Reserving a small number of slots for students from lower-socioeconomic tiers is nearly as efficient as racial quotas if one only wants to slightly boost minority enrollment. But the CPS policy is much less efficient when one tries to use it to keep minority representation anywhere close to its former level. There is also a limit to what can be accomplished with these policies ignoring efficiency concerns: minority representation at Payton would decrease even if one were to allocate 100% of seats to the tier-specific quotas. The CPS policy is more effective at Payton than Northside, because there are relatively fewer high-scoring applicants in the lowest SES areas who apply to Northside. SES-based affirmative action plans are intended to also provide integration on other dimensions including low-income representation. In theory, such a benefit may offset these plans’ decreased efficiency as a tool for racial integration. But in practice we find that Chicago’s plan is actually less efficient than racial quotas as a means for increasing the number of low-income students at Payton and Northside. We also examine within-school heterogeneity in scores and within-school racial gaps. While the majority of admitted students have 99th percentile scores under either plan, there are more lower- scoring students and their scores are lower under the CPS plan. The shift to race-neutral affirmative action could increase or decrease within-school racial gaps. One effect that tends to narrow the gap is that some of the lower-tier slots go to whites and Asians, which brings down their average scores. But a second opposite effect is that some of the highest scoring blacks and Hispanics (including some low- income students qualifying for free lunch) fail to gain admission which also brings down the minority average. WefindthatthesecondeffectisstrongerthanthefirstandthattheSES-basedplansincrease the within-school racial gap in composite scores. Section 5 explores how much of the inefficiency of Chicago’s SES-based system is driven by a suboptimal implementation of the general idea as opposed to being inherent to any race-neutral policy. ThecurrentCPSsocioeconomicindexisanunweightedaverageofsixtract-levelcharacteristics. 6 Intuitively, an index-based rule can only be nearly as efficient as a race-conscious rule if the index is highly correlated with the minority status among students near the acceptance/rejection margin. The CPS index does not have this property. We identify alternative policies (involving SES-related “bonus points”) that would be somewhat more efficient. Using additional census tract characteristics to predict minority status, chosen by a cross-validated LASSO procedure, does not result in a significant improvement. We also examine approaches inspired by rules like Texas’s top 10% rule, implemented based on Chicago’s 77 neighborhood areas or census tracts. We conceptualize these rules as a much more granularversionofChicago’stier-basedadmissions. Insteadofallocatingslotstofourtiers,thepolicies assign slots to much smaller entities like the census tract or neighborhood area. A policy using census tracts would be much less efficient than the CPS policy. Intuitively, the policy is highly inefficient when used for programs serving extreme high-achievers because they are outliers: there will be many cases in which geographic unit A has two candidates who are much stronger than the top candidate from geographic unit B, even when units A and B serve students of similar socioeconomic status. An implementation based on neighborhood areas is much better than the tract-based implementation and provides another way to improve the current CPS policy. Our biggest-picture conclusion, however, is that all feasible place-based policies appear to be substantially less efficient than race-based policies. This paper is most closely related to the innovative work of Fryer, Loury, and Yuret (2008), which developed a model of “color-blind” vs. “sighted” affirmative action to show that the reduction in the quality of the student body is an efficiency cost of color-blindness. For university admissions, they considerathoughtexperimentintheCollegeandBeyonddatasetwhereeachofseveralcollegesadmits a student body half as large as their current one from an applicant pool consisting of their current students. We build on their analysis and expand on their empirical work in several dimensions: we examine examine more realistic changes to admissions policies, an approach made possible by having data on the full applicant pool; we compare a variety of race-neutral policies; and we examine additional outcomes including within-school achievement gaps and effects on the representation of low-income students. Cestau, Epple, and Sieg (2015) develop an econometric model of the referral process for taking the admissions tests for selective elementary schools. They report that profiling by race and income together with affirmative action based on free lunch status can achieve 80% of level of black enrollment as a race-based affirmative action plan. Corcoran and Baker-Smith (2014) study admissions policies at New York’s exam schools, focusing on a descriptive account of the decision to apply. Though their main interest is not in affirmative action, they simulate top 10% rules based 7 on 7th grade scores and find that it leads to an increase in black and Hispanic representation at the schools. Our model is also related to several other papers on affirmative action. Chan and Eyster (2003) show that if an admissions office is prohibited from using affirmative action, they can still increase minority representation by placing less weight on applicant qualifications. In a study motivated by Chicago’s assignment mechanism, Dur, Pathak, and S¨onmez (2016) characterize the best precedence, or order in which affirmative-action and non-affirmative action slots at schools are processed, for applicants from particular tiers. In our simulations, we use their results to focus on the best tier-blind precedence for applicants from the most disadvantaged tier. Epple, Romano, and Sieg (2008) develop an equilibrium model of affirmative action and tuition policies to show that a ban on affirmative action leads to a decline in minority students at top-tier colleges. Ray and Sethi (2010) show that any score-maximizing race-neutral affirmative action plan must be a non-monotone function of past performance. 2 A Model of Elite Schools and Affirmative Action 2.1 Rationalizing elite schools We begin with a simple model in which it is efficient for a school system to create an elite school when students benefit from a curriculum tailored to their ability or preparation. Formally, suppose that a school system serves a continuum of students with types θ continuously distributed on R. The type θ can be thought of as a composite of the student’s ability and/or preparation. The school system operatestwoschoolswithcurriculac ,c ∈ R. Supposethattheexpectedoutcomeforatypeθ-student 1 2 if assigned to school i depends on the student’s type and on the match between the students’ type and the curriculum as follows: V (θ) = h(θ)−k(θ−c )2, (1) i i where h(θ) governs the direct relation between student type and expected outcome, the curriculum c is described by the type of the student it serves best, and k indexes the relative importance of i student-curriculum matching. An assignment policy specifies which student types are assigned to each school and the curriculum taught in each school.5 The optimal assignment policy maximizes the 5Whilecurriculaareachoicevariableinthismodel,analternativeinterpretationisthatonlyschoolassignmentsare achoicevariableandafterschoolassignmentsaremade,classroomdynamicsforceteacherstoadoptthecurriculumbest suited to the mean student in the school. 8 sum of students’ expected outcomes. Giventhisobjective,aggregateoutcomesarehighestifstudentsaregroupedaccordingtoθandeach school’s curriculum is matched to the set of students it serves. The following proposition formalizes the rationale for elite schools. Proposition 1 If students’ outcomes are given by (1), the optimal assignment policy is defined by a cutoff θˆ, where students with θ > θˆ are assigned to school 1 and students with θ < θˆ are assigned to school 2. The curricula are c = E(θ|θ > θˆ) and c = E(θ|θ < θˆ). Furthermore, the optimal θˆ is the 1 2 solution to E(θ|θ < θˆ)+E(θ|θ > θˆ) θˆ= . 2 In any assignment scheme, the curricula c and c are the mean θ’s of the students assigned to each 1 2 school. Given any curricula with c > c , optimal student assignment implies that each student with 1 2 θ > c1+c2 is assigned to school 1, which corresponds to the condition on θˆin the proposition. 2 The top panel of Figure 1 provides a simple illustration of a situation in which an elite school is optimal. The curve represents the density f(θ) of the type distribution, which continues substantially further to the right.6 The vertical line indicates the cutoff θˆ with students to the right of this line being assigned to school 1. The optimal curricula c and c are the conditional means of θ in each 2 1 school. The cutoff θˆis equidistant between c and c . 2 1 2.2 Optimal affirmative action with diversity benefits Continuingourmodeldevelopment,supposethatstudentoutcomesdependonbothcurriculummatch- ing and classroom diversity. Admissions arrangements at elite schools often reflect both of these con- cerns. For example, a Blue Ribbon Commission evaluating Chicago’s policy felt that it was important for the “programs to maintain their academic strength and excellent record of achievement, but also believes that diversity is an important part of [their] historical success” (BRC 2011). Suppose that the school system serves two populations indexed by j = {o,u}, where o denotes overrepresented and u denote underrepresented at the elite school absent affirmative action. Let m be the fraction of students who belong to the population u. Let F and F be the distributions of types o u within each population. Due to some source of disadvantage, we assume that the type distribution is 6The figure was generated by solving a model with k=1, d=1, and θ having a standard exponential distribution. 9 Optimal Curriculum Design and Student Assignment Base Model 0 0.5c 1 1.5 ^ 2 2.5c 2 θ 1 ̂ Optimal Affirmative Action With ADiversity Benefit 0 0.5 1 ϴ^1.5 ϴ^ ϴ^2 2.5 3 u o Second Best Proxy-Based Affirmative Action z=u z=o z=u z=o ^ ^ ^ 0 0.5 1 ϴ1.5 ϴ ϴ 2 2.5 3 u o Figure 1: Optimal admissions policies (a) without a diversity benefit; (b) with a diversity benefit; and (c) when only an imperfect proxy is available. 10