Crossing District Lines: The Impact of Urban–Suburban Desegregation Programs on Educational Attainments

Abstract

Interdistrict desegregation programs, which provide opportunities for urban children of color to attend suburban schools, are a potential means of addressing persistent racial inequalities in educational opportunities and outcomes. These voluntary programs offer a test of whether nonresident students can leverage the resources and social capital available at high-performing suburban schools to improve their educational outcomes. In the first impact study of Boston’s long-running program, I find large differences in the adjusted high-school graduation and college enrollment rates of applicants referred to a suburban district, compared with observably similar applicants who were not referred. The college effect is due to enrollment in 4-year institutions and does not vary by gender. Estimates are robust to adjustments for remaining omitted variables bias.

Keywords

desegregation program evaluation urban education educational policy race regression analyses

Long-standing inequalities in educational opportunity and outcomes pose a persistent challenge in the United States, particularly along lines of race and ethnicity. Black and Latino students have lower rates of high-school graduation and college enrollment than do Whites (McFarland et al., 2019). While 36% of non-Hispanic Whites over age 25 have attained a bachelor’s degree, only 22% of Blacks and 16% of Latinos hold this credential (Ryan & Bauman, 2016). The disproportionate number of Black and Latino students exiting the educational pipeline at these critical junctures contributes to the perpetuation of income inequality, particularly given the wage premium associated with a 4-year college degree (Autor, 2014; McFarland et al., 2019). The racial/ethnic gap in educational attainments contributes to a pronounced disparity in annual median income, with full-time Black and Latino workers aged 25–34 years earning only 75% of White workers in the same age bracket (de Brey et al., 2019).

One strategy to address these racial/ethnic disparities in educational outcomes is cross-district desegregation programs, in which urban students of color have the opportunity to attend public schools in neighboring suburban communities.¹ Patterns of residential segregation by race and income have concentrated minority children in high-poverty public schools, with adverse effects on their educational achievement and completion rates (Reardon et al., 2017, 2019). Interdistrict desegregation programs offer a test of whether the resources and social capital available at advantaged suburban schools can be leveraged by nonresident minority students and their families to improve their own outcomes.²

I study the voluntary program in Boston, one of the oldest in the country, through which limited numbers of minority students gain admission to suburban public schools.³ I compare outcomes for program applicants who received a referral to a suburban district to those of applicants who did not, using student-level covariates to adjust for nonrandom selection. I find large, positive effects on high-school graduation and immediate college enrollment. The causal interpretation of these effects requires an assumption that referral is as good as randomly assigned, conditional on observed characteristics, which is highly implausible given the program’s selection procedures. Therefore, I make use of students’ eighth-grade standardized test scores, comparing the stability of the treatment coefficients with and without these scores in the model, and quantify the degree of omitted variables bias that would be necessary to invalidate the inference. I also test the sensitivity of my estimates to attrition and missing referral data and find that the results are robust to a variety of specifications and strategies for missing-data imputation.

My work is most similar to that of Bergman (2016), who reported large effects on college enrollment for an interdistrict transfer program in East Palo Alto, California. That program differs from the Boston-based intervention studied here on several important dimensions. First, it enrolls about 22% of incoming kindergartners from the sending urban district, compared with only 3% for the Boston program. Next, receiving districts are mandated by court order to participate in the East Palo Alto program by accepting a fixed number of applicants each year, and students are admitted via lotteries. In Boston, participation by suburban districts is entirely voluntary, and these districts exercise some discretion in which urban applicants they admit. Finally, the transfer program in East Palo Alto admits only students entering Grades K–2, and six of the seven receiving districts serve only Grades K–8 (meaning that those participants must exit the program for high school). In Boston, all receiving districts offer the opportunity for participants to continue on through high-school graduation, meaning that the intended treatment extends through the measurement of the first outcome of interest.

I begin by reviewing the literature on the potential effects of suburban enrollment for minority urban students and the limited evidence from interdistrict desegregation programs. I present two sets of analyses to estimate the difference in outcomes for students who applied or participated in the program versus comparable students who did not. I then discuss the threats posed by selection bias and missing data to my estimates.

Background and Context

Boston’s interdistrict integration program, known as the Metropolitan Council for Educational Opportunity (METCO), affords a rare opportunity to study the effects of a dramatic shift in school context. (The METCO program also exists in Springfield but is administered and structured separately; the present study focuses entirely on Boston.) In Table 1, I present descriptive statistics for first-time ninth-grade students enrolled in the Boston Public Schools, Boston charter schools, and the METCO receiving districts in 2007–2008 and 2008–2009.⁴ The demographic differences between the two Boston settings and the suburban districts are stark. Two thirds of the students in Boston schools qualify for free or reduced-price lunch, my only available measure of family income, compared with only 5% of suburban residents. The receiving districts are 85% White, not counting METCO participants, while less than 15% of students enrolled in Boston’s district and charter schools are. Finally, there are very large differences in average academic proficiency, as measured by eighth-grade scores on the Massachusetts Comprehensive Assessment System (MCAS): suburban students perform .75 to 1.0 standard deviations higher on average than either of the Boston groups in both math and English language arts (ELA).

Table 1

Enrollment Figures and Mean Test Scores for First-Time Ninth-Grade Students in the Fall of 2007 and 2008, by Enrollment Group

Variable	BPS	Boston charter	Receiving districts
Students	8,682	771	15,448
Schools	36	8	30
Low-income	0.69	0.62	0.05
Female	0.49	0.59	0.51
ELL	0.14	0.00	0.01
IEP	0.19	0.15	0.14
Immigrant	0.08	0.00	0.01
Asian	0.09	0.02	0.08
Black	0.40	0.61	0.02
Latino	0.35	0.23	0.03
White	0.14	0.12	0.85
MCAS 8th-grade math	–0.39	–0.14	0.62
MCAS 8th-grade ELA	–0.50	–0.25	0.57

Note. MCAS scores standardized within grade and year, using statewide mean and SD. BPS = Boston Public Schools; ELL = English language learners; IEP = Individualized Education Plan; MCAS = Massachusetts Comprehensive Assessment System; ELA = English language arts.

Potential Mechanisms

METCO students, most of whom enter the program in Grades K–2, are moving into well-resourced suburban schools filled with highly advantaged, high-performing students. Possible mechanisms through which the program might positively impact participants’ later outcomes include exposure to a richer curriculum taught by highly qualified teachers, a college-going culture and engagement with schoolwork among peers, and better access to information and support during the college applications process. However, other school and nonschool factors could work against the program’s effectiveness: the potential tracking of METCO students into less demanding classes, the burden of a long daily commute that makes extracurricular participation and homework completion difficult, and the challenges of social isolation in largely White school contexts.

Urban public schools with large concentrations of minority and low-income students employ less experienced, less effective teachers on average and experience higher rates of turnover (Boyd et al., 2011; Loeb et al., 2005; Steele et al., 2015). Students in these schools may have little access to Advanced Placement and other college-preparatory courses (Orfield et al., 2008; Rumberger & Palardy, 2005; Theokas & Saaris, 2013). In contrast, Boston students who participate in METCO gain access to the extensive resources of suburban high schools, such as rich curricular offerings, smaller classes, and highly qualified teachers. Importantly, these resources also include peer groups of highly advantaged, college-bound students, with relatively fewer children with academic or behavioral challenges (see Harris, 2010; Jencks & Mayer, 1990; Lazear, 2001; Sacerdote, 2011).

While information about the college admissions process and financial aid can be difficult to obtain in large urban high schools with few guidance counselors (Chapman, 2014), METCO students experience no such informational constraints. In a qualitative study of the program’s alumni, Eaton (2001) documented the importance of information sharing among METCO parents and students, suburban friends and their families, and school staff. Through these mechanisms, METCO participants were able to navigate the application processes for college and financial aid and hit important milestones, such as preparing for and taking the SAT.

However, other factors may mitigate the benefits of a program like METCO. Teachers and peers in majority-White schools may harbor lower expectations for Black and Latino students and doubts about their intellectual abilities (Diamond, 2006; Eaton, 2001). Research in majority-White schools has revealed the disproportionate placement of Black and Latino students in lower academic tracks, which lessens their exposure to high-quality teachers and college-oriented peers (Chapman, 2014; Ispa-Landa, 2013; Ispa-Landa & Conwell, 2015; O’Connor et al., 2011). State data provide suggestive evidence in this regard: In the spring semester of 2012, only 8% of METCO high-school seniors in this study’s sample were enrolled in at least one Advanced Placement course, compared with 41% of their counterparts in Boston Public Schools (BPS) (author calculations). These students might have been encouraged to take more challenging courses and achieve at higher levels had they remained in Boston schools (Warikoo & Carter, 2009).

Other disadvantages can discourage the academic success of minority students in majority-White suburban schools (Diamond, 2006; Lewis & Diamond, 2016). Their families have lower average levels of income to put toward educational supports like computers, calculators, and books that suburban teachers may utilize heavily. In the case of METCO, many participants endure bus rides of an hour or more to schools that are quite far from Boston. Such long commutes leave less time for homework, adversely impact sleep, and can impede participation in extracurricular activities (Holland, 2012).

Finally, there are very large achievement gaps between METCO high schoolers and the resident students in the program’s suburban districts: .90 SD on the 10th-grade math MCAS assessment and .75 SD on the ELA test. Whatever gave rise to these differences, they may be sizable enough to demoralize METCO students and negatively affect their effort and aspirations. If this occurs, then the achievement and attainment of METCO students may be lower than they would have been in a Boston high school. Using data from the National Longitudinal Study of Adolescent Health, Owens (2010) reported that students from neighborhoods with low average levels of socioeconomic status (SES) have lower odds of educational attainment if they attend schools with higher proportions of White and high-SES students.

Evidence From Interdistrict Programs

The only extant causal evidence linking interdistrict integration programs with high-school and college outcomes comes from the Ravenswood/Tinsley program in Palo Alto, California. This program includes sending and receiving districts with large student test-score differentials, similar to METCO, but differs from it as described above. Bergman (2016) found that Ravenswood/Tinsley lottery winners were 10 percentage points more likely to enroll in college, with effects concentrated for males and driven by enrollment in 2-year public colleges.

Early research from Project Concern, a program based in Hartford that randomly assigned most but not all urban applicants to suburban schools beginning in 1966, also found more robust effects for boys on high-school graduation and college enrollment (Crain, 1992). In the St. Louis program, which does not use random assignment, Black students attending suburban schools graduated from high school and enrolled in college at rates approximately twice those of similar students in the city’s public schools (Wells & Crain, 1997).

Finally, the limited previous reporting on METCO’s program outcomes has compared the average performance of students active in the program to those in the Boston Public Schools and the state of Massachusetts as a whole. These results are simple differences, unadjusted for any demographic dissimilarities in the students comprising the two groups. These calculations also include only those students enrolled in the program as ninth graders, without accounting for program attrition prior to that.

Eaton and Chirichigno (2011) reported 4-year high-school graduation rates for 2006–2009 for the state, METCO students (including both Boston and Springfield participants), Boston students, and the two subgroups of Black and Latino students in Boston. The METCO average, which ranged from 92.1% to 94.8% over the 4 years reported, was consistently the highest across these groups. The average for “Boston students” ranged from 58% to 61%, though it is unclear whether this group includes just BPS students or students enrolled in Boston’s charter schools as well.

In 2013, the Executive Office of Education (EOE) produced a report for the state legislature that included data on the high-school and college outcomes of METCO students. Over the period from 2006 to 2011, the METCO average 4-year high-school graduation rate was always above 94% and consistently equaled or exceeded the average rate for resident students in receiving districts. In comparison, the BPS rates for these years ranged between 58% and 64%. The state also reported that 89% of METCO students who graduated from high school in 2010 enrolled in college the following fall, compared with 68% of BPS graduates and 91% of resident students in receiving districts. Among that year’s METCO graduates, 21% enrolled in 2-year community colleges and 68% in 4-year institutions.

METCO Selection

METCO is one of several options available for K–12 students residing in Boston: 73% are enrolled in BPS, 12% in charter schools, and another 11% in private schools (BPS Communications Office, 2016). METCO enrolls approximately 3,100 students per year, or about 4% of the total. Thirty-three public school districts near Boston voluntarily participate in the program, which is administered by a nonprofit known as METCO, Inc. The list of METCO suburban districts has remained virtually unchanged since the 1970s.

The METCO selection process during the period covered by this study is lengthy and complex. All students residing in Boston are eligible, except those with out-of-district special education placements; in practice, nearly all applicants are non-White. After completing a paper application and submitting ancillary materials like proof of residence, health records, and their cumulative schooling and MCAS histories, students are eligible for referral to one of the suburban districts.

Every year, each of the 33 participating suburban districts provides the METCO central office with an estimate of the number of slots by grade and gender it will have available for new METCO students in the coming school year. Most district requests are for younger students; approximately 70% of students enter the program in grades kindergarten through two. These requests come in as early as November and as late as May. The placement staff then forwards completed packets to each district on a rolling basis. A student’s file goes to only one district at a time and is considered pending until the district makes a decision. There is no set timetable for districts to act on each application, given that they each have their own selection processes, usually involving multiple steps. Approximately 73% of students who are referred go on to enroll in the program; the others do not receive an enrollment offer from the district or choose to decline the offer once it is made.⁵

METCO, Inc., has a selection algorithm that is intended to determine the order in which applications are referred to districts. Siblings of enrolled students should receive first priority, unless the receiving district judges that the parents have been uncooperative or that the enrolled child is not experiencing success in the program. Once siblings have been accounted for, the METCO placement staff should pull folders for the remaining spots based on the order in which families registered for the program. However, applicants who have already been referred to one or more districts, but were declined, may receive priority. (Parents may also decline a placement offered by a particular district but run the risk of not receiving another referral that year.) Also, the child’s race/ethnicity is a factor, since METCO, Inc., is trying to make its student demographics resemble those of the Boston Public Schools. The placement staff employs an annual referral target of 30% to 40% Latinos and Asians, so a student in one of these underrepresented groups can be referred ahead of a Black child who applied earlier.

According to this intended algorithm, then, METCO assigns referrals separately within four categories of students: (a) students of any race or ethnicity with a sibling already in a METCO district, (b) Black, (c) Latino, and (d) Asian. The first of these categories receives priority over the other three. Within each category, referrals are to be given in the order families applied, with older applications referred first.

However, evidence from the METCO application records indicates that METCO placement staff did not adhere strictly to this algorithm in assigning referrals.⁶ Within each of the applicant risk sets created by the above categories, there existed no sharp cutoff date of application such that students who applied before that date received a referral and those applying after it did not. In Table 2, I show counts of actual referrals and of students who should have been referred had the algorithm been strictly followed. The table is for a subsample of 1,181 Black and Latino applicants to kindergarten from 2004 to 2010.⁷ The vast majority of the kindergarten applicants in this group received referrals. However, the referral decisions on 300 applications were not in accordance with selection rules: 150 students were not referred who should have been, given their dates of registration, and another 150 were referred in their stead. This means that 25% of the total applicants were referred or not referred using some decision rule apart from the algorithm, a fact that has major implications for the inferences that can be drawn from the sample of METCO applicants, as I explain below.⁸

Table 2

Referral Statistics for METCO, Inc., Black and Latino Applicants to Kindergarten, 2004–2010, by Whether Students Were Actually Referred to Suburban Districts and Whether They Should Have Been Referred According to the Program’s Selection Algorithm

Referral by algorithm	Not referred (actual)	Referred (actual)
Not referred by algorithm	41 (21%)	150 (15%)
Referred by algorithm	150 (78%)	840 (85%)
Total	191 (100%)	990 (100%)

Note. The counts in the table do not include 51 Black and Latino students who enrolled in METCO as kindergartners but could not be matched to the METCO, Inc., application records. METCO = Metropolitan Council for Educational Opportunity.

Analyses of METCO Impacts

To estimate the impact of the program on on-time high-school graduation and immediate college enrollment, I use two sets of analyses, each with its own advantages and drawbacks. First, I compare several groups of METCO applicants, classified according to their final status in the selection process, with nonapplicants who attended a public high school in Boston. While these groups differ on pretreatment characteristics, the relative average performance of demographically similar high school students in METCO, BPS, and charter schools is a question of key interest to stakeholders. Moreover, the separate consideration of active METCO participants, those who left the program prior to ninth grade, and applicants who never enrolled gives some sense of the likely biases due to program selection and attrition. I fit simple unadjusted linear probability models, presented here to facilitate comparison to previous work; models that include demographic covariates, thereby controlling for observed differences between the groups; and models that adjust for eighth grade test scores to compare students of similar academic proficiency. Including eighth-grade scores controls away some of the program’s effect, given that most participants entered the program well before eighth grade, so this approach is quite conservative.

The second, preferred approach involves restricting the sample to students within the METCO applicant pool, since students who applied to the program likely differ in important ways from those who did not. Within this sample, I fit the same three models to estimate raw differences, differences that are adjusted for demographic covariates, and differences that are adjusted for demographics and eighth-grade test scores. Parameter estimates of interest from the adjusted models likely include bias from additional omitted variables, which I address in the subsequent discussion.

Approach 1: METCO Applicants Versus Other Boston Resident Students

I begin by examining outcomes for Boston residents who entered ninth grade for the first time in the fall of 2007, 2008, and 2010.⁹ I look at first-time ninth graders due to my focus on on-time high-school graduation, which I define as receiving a diploma within 4 years of ninth-grade entry. I compare seven mutually exclusive groups of students: (a) METCO ninth-grade participants (whom I call Actives); (b) METCO participants from these cohorts who exited the program prior to ninth grade (Attriters); (c) METCO applicants from these cohorts who were offered a seat for Grade 9 or earlier but never enrolled (Decliners); (d) METCO applicants who were referred to a suburban district for Grade 9 or earlier but were never offered a seat (Unsuccessful Referrals); (e) METCO applicants who never received a referral (Nonreferrals); (f) nonapplicants enrolled in BPS as first-time ninth graders (BPS); and (g) nonapplicants in Boston charter schools (Charter).¹⁰

The sample for these analyses includes 14,430 students residing in Boston. I identified these students and compiled their cumulative schooling histories using annual Student Management Information System (SIMS) data files from the Massachusetts Department of Elementary and Secondary Education (DESE), which include data on every K–12 student attending a public school in Massachusetts that year. I used the October enrollment files from 2001, the first year SIMS data were collected, through 2014 to identify students enrolled in METCO, classifying all students who were in the program as of the October enrollment count as full-year participants.

In Table 3, I present descriptive statistics for the seven distinct groups of Boston residents identified above. The first five columns correspond to the different categories of METCO applicants; mean statistics for all applicants are in a summary column for reference. The final two columns disaggregate METCO nonapplicants by the type of high school they attended as first-time ninth graders: BPS or a charter school.

Table 3

Descriptive Statistics for the Sample of First-Time Ninth Graders in 2007, 2008, and 2010, by Whether They Had Ever Applied to METCO (n = 14,430)

	METCO applicants						METCO nonapplicants
Variable	Actives(1)	Attriters(2)	Decliners(3)	Unsuccessful referrals(4)	Nonreferrals(5)	All applicants	BPS high school(6)	Boston charter school(7)
Female	0.55	0.47	0.54	0.50	0.47	0.53	0.49	0.57
Low-income	0.50	0.53	0.62	0.70	0.68	0.56	0.70	0.65
IEP in 9th grade	0.31	0.06	0.16	0.18	0.24	0.26	0.19	0.13
ELL	0.03	0.02	0.06	0.07	0.04	0.04	0.17	0.01
Immigrant	0.01	0.02	0.00	0.01	0.01	0.01	0.09	0
Asian	0.03	0.05	0.14	0.07	0.04	0.04	0.1	0.02
Black	0.82	0.72	0.64	0.63	0.73	0.77	0.38	0.60
Latino	0.13	0.20	0.20	0.28	0.21	0.16	0.36	0.23
White	0.01	0.01	0.02	0.02	0.01	0.01	0.14	0.13
MCAS 8th-grade math	–0.40	–0.23	–0.01	–0.26	–0.52	–0.36	–0.39	–0.09
MCAS 8th-grade ELA	–0.16	–0.11	–0.16	–0.32	–0.48	–0.21	–0.52	–0.23
Proportion with 8th-grade scores	0.95	0.94	0.89	0.93	0.94	0.94	0.82	0.88
K–8 enrollment:
Mean years in METCO	5	3.4	0.1	0	0	3.5	0	0
Mean years in BPS	1.1	2.3	4.1	5.4	5.2	2.3	5.1	3.8
Mean years in charter	0.3	0.3	1.1	1.1	1.1	0.5	0.2	1.9
Count	856	83	81	164	144	1,328	12,064	1,038
Proportion	0.64	0.06	0.06	0.12	0.11	1.00	0.92	0.08

Note. MCAS scores standardized within grade and year, using statewide mean and SD. (1) Actives: students enrolled in METCO as ninth graders; (2) Attriters: students enrolled in METCO who exited prior to ninth grade; (3) Decliners: students who were offered a METCO seat but declined it; (4) Unsuccessful Referrals: students who were referred to a suburban district but did not receive an offer; (5) Nonreferrals: students who did not receive a referral to a suburban district. Nonapplicants are categorized by whether they were enrolled in a Boston Public Schools high school (6) or charter school (7) as ninth graders. METCO = Metropolitan Council for Educational Opportunity; BPS = Boston Public Schools; IEP = Individualized Education Plan; ELL = English language learners; MCAS = Massachusetts Comprehensive Assessment System; ELA = English language arts.

Comparing the demographics of METCO applicants with those of the nonapplicants indicates why simple comparisons of outcomes may be misleading. The applicant group is 77% Black and only 16% Latino, compared with 38% and 36% of the BPS group, respectively. Nonapplicants in both BPS and Boston charter schools are about 13% White, while applicants are virtually all from minority groups. About 55% of students who apply to METCO are low-income, compared with over two thirds of nonapplicants. The table reveals other proportional differences by gender, English language proficiency, and immigration status. Compared with Boston students who never applied to the program, METCO ninth-grade participants in particular are more likely to be Black and to have an Individualized Education Plan (IEP), and less likely to be Latino or classified as English language learners.

Because of the correlation of student demographics with educational attainment, one would expect the applicant group to have different mean outcomes from students who never applied, independent of their program participation. The two groups are also very likely to differ on unobserved characteristics. The multiple factors that might have led some students but not others to apply, such as access to information via social networks, parental engagement and persistence, and the relative valuing of education among family priorities, are unmeasured but probably correlated with the outcomes of interest.

Other important aspects of Table 3 are differences within the METCO applicant pool, which provides a window into the various stages of program selection. Almost two thirds of applicants were active in the program entering ninth grade, and almost 90% had received a referral to a suburban district at some point during Grades K–9. Students who fulfilled all application requirements, but whom METCO, Inc., placement staff never referred to a suburban district (Nonreferrals), are disproportionately male and low-income compared with applicants who were referred. They also had lower eighth-grade MCAS scores, although this measure of academic achievement is endogenous since most METCO applicants are first eligible for the program in Grades K–2. These scores are missing for the 15% of students in the sample who were enrolled in private schools or lived out of state for eighth grade.

Students who received a referral but not an offer of admission (Unsuccessful Referrals) are similar in demographics to the unreferred group but have higher eighth-grade scores, which is additional suggestive evidence of nonrandom selection for referrals by METCO staff. The group that did receive an offer but never enrolled (Decliners) has a noticeably higher proportion of Asian students and the highest mean eighth-grade scores of any group. The 6% of applicants who enrolled but left before ninth grade (Attriters) are fairly similar on demographic characteristics to those who remain (Actives) and have slightly higher average eighth-grade test scores; their exposure to the program before ninth grade is 3.4 years on average compared with 5 years for the Actives.

I examined two main outcomes in all the analyses for this study. The first is HSgrad_i, an indicator for whether student i graduated from a Massachusetts public school within 4 years of initially enrolling in ninth grade. This indicator derived from the October and end-of-year SIMS files through 2014. Using the state’s definition, I treated students who graduated during the summer following their fourth year as on-time graduates (Massachusetts Department of Elementary and Secondary Education [DESE], 2006). Students who did not graduate by the end of the 2013–2014 school year had dropped out, were still enrolled in high school, or had transferred out of the state’s public schools. While many students in the last group probably did graduate on time from high school elsewhere, I coded them all as nongraduates in the main analyses. A sensitivity analysis excluding these students yielded slightly different results for the smaller applicant groups (Attriters, Decliners, Unsuccessful Referrals, and Nonreferrals), so I collapsed these groups into a single category and present results for the Active, Inactive Applicant, Charter, and BPS groups.

The postsecondary outcome is College_1sem_i, an indicator for whether student i enrolled in college during the fall immediately following graduation from high school, regardless of whether the student graduated high school “on time.” I derived this variable from a National Student Clearinghouse (NSC) data set with college enrollment records for all Massachusetts high school graduates. This data set includes the same unique State-Assigned Student Identifiers (SASIDs) as the SIMS files, enabling accurate merging of the two. The NSC file contains enrollment records by semester for each institution a student attended, including over 3,400 colleges and universities covering 96% of U.S. students. Two ancillary outcomes come from these same data: Fouryear_i, which takes a value of 1 if the student first enrolled in a 4-year college or university and zero otherwise, and Twoyear_i, which is coded similarly for immediate enrollment in a 2-year institution.

My main model, here with HSgrad as the outcome, is

H S g r a d_{i} = β_{0} + γ^{'} E n r o l l_{G r 9, i} + β X_{i} + ε_{i},

(1)

where Enroll_Gr9,i is a categorical variable indicating to which enrollment group student i belonged as a first-time ninth grader, with BPS nonapplicants serving as the reference category. The vector X_i includes the following demographic covariates: gender, race/ethnicity, and indicators for whether the student had an IEP, was classified as an English language learner or a recent immigrant, and was eligible for free or reduced-price meals. To save space, I do not include these parameter estimates in the following tables.

All dependent variables are dichotomous. Ordinary least squares (OLS) and logistic regression models return similar results as long as the sample means are not near the extremes of zero or one. In Table 4, I display means on the outcomes of interest in the samples I use for my analyses. For all outcomes except Twoyear, the sample means are located in the middle of the range of probabilities. The mean for Twoyear is only 0.10 to 0.15, depending on the sample. However, even in this case, the results from logistic regression models are very similar to OLS in sign, statistical significance, and magnitude.¹¹ To simplify the interpretation of regression coefficients, I report OLS estimates in the tables of results. The full set of logistic results is available upon request.

Table 4

Proportions of Students Attaining Intermediate Outcomes, by Analytic Sample and Whether Their Records Include Baseline Test Scores

Outcome	Boston sample		METCO applicants
Outcome	All students	Students with 8th-grade scores	All students	Students with 8th-grade scores
Main outcomes:
On-time high-school graduation	0.58	0.63	0.80	0.81
College matriculation	0.41	0.45	0.63	0.64
Other outcomes:
Four-year institution	0.32	0.35	0.56	0.58
Two-year institution	0.10	0.10	0.15	0.15

Note. METCO = Metropolitan Council for Educational Opportunity.

Findings in the Full Sample

In Tables 5 and 6, I report results from Equation 1 using on-time high-school graduation and immediate college matriculation as the dependent variables, respectively. The first two columns of each table correspond to models fit within the full sample. The results in the first column, from a model without covariates, are most directly comparable to previous empirical work on METCO. I find a simple difference of about 33 percentage points in on-time graduation between active METCO participants and BPS ninth-grade nonapplicants (Table 5); the unadjusted graduation rates of these groups are .90 and .57, respectively. For college enrollment (Table 6), the difference is 29 percentage points between unadjusted means for the two groups of .72 and .43. The high-school graduation results are in line with previous estimates cited above. My estimates of college enrollment rates include all students, not just high-school graduates, in the denominator and are therefore lower than those in the state’s report (Executive Office of Education, 2013), and the difference between the college enrollment rates of METCO Actives and BPS nonapplicants is larger (29 vs. 20 percentage points).

Table 5

Results From OLS Regressions Predicting On-Time High-School Graduation Rates of Boston Public-School Students Who Entered Ninth Grade for the First Time in 2007, 2008, and 2010 (Standard Errors in Parentheses)

	Full sample		Subsample with 8th-grade scores
Predictor	Unadjusted	Adjusted	Adjusted (no scores)	Adjusted (with scores)
Actives	0.33** (0.017)	0.35** (0.017)	0.32** (0.017)	0.26** (0.017)
Inactive applicants	0.085** (0.023)	0.083** (0.022)	0.055* (0.022)	0.021 (0.022)
Charter nonapplicants	0.073** (0.016)	0.052** (0.015)	0.045** (0.016)	–0.0001 (0.015)
Eighth-grade math score				0.12** (0.0064)
Eighth-grade ELA score				0.067** (0.0061)
(Adjusted) mean for Never applied: BPS HS	0.57** (0.0044)	0.58** (0.010)	0.62** (0.011)	0.67** (0.011)
Demographic controls	No	Yes	Yes	Yes
N	14,430	14,430	12,042	12,042
R ²	.027	.116	.103	.175

Note. The reference category is students who never applied to METCO and enrolled in Boston Public Schools (BPS) for ninth grade; the intercept is here relabeled as the (adjusted) mean for this group to facilitate interpretation. The adjusted models include dummies for Asians, Latinos, Whites, and other race/ethnicity (with Blacks as the reference category), and free or reduced-price lunch status, immigrant status, and IEP status entering ninth grade. These parameter estimates are omitted to save space. OLS = ordinary least squares; ELA = English language arts; BPS = Boston Public Schools; HS = high school; METCO = Metropolitan Council for Educational Opportunity; IEP = Individualized Education Plan.

p < .05. **p < .01.

Table 6

Results From OLS Regressions Predicting Immediate College Matriculation Rates of Boston Public-School Students Who Entered Ninth Grade for the First Time in 2007, 2008, and 2010 (Standard Errors in Parentheses)

	Full sample		Subsample with 8th-grade scores
Predictor	Unadjusted	Adjusted	Adjusted (no scores)	Adjusted (with scores)
Actives	0.29** (0.018)	0.32** (0.017)	0.30** (0.018)	0.25** (0.017)
Inactive applicants	0.14** (0.024)	0.15** (0.023)	0.12** (0.024)	0.084** (0.023)
Charter nonapplicants	0.23** (0.016)	0.21** (0.016)	0.21** (0.017)	0.17** (0.016)
Eighth-grade math score				0.12** (0.0068)
Eighth-grade ELA score				0.066** (0.0066)
Adjusted mean for never applied: BPS HS	0.43** (0.0047)	0.43** (0.011)	0.46** (0.012)	0.51** (0.011)
Demographic controls	No	Yes	Yes	Yes
N	13,130	13,130	11,156	11,156
R ²	.034	.133	.125	.191

Note. The reference category is Black students who never applied to METCO and enrolled in Boston Public Schools (BPS) for ninth grade. The adjusted models include dummies for Asians, Latinos, Whites, other race/ethnicity, free or reduced-price lunch status, immigrant status, and IEP status entering ninth grade; these parameter estimates are omitted to save space. OLS = ordinary least squares; ELA = English language arts; BPS = Boston Public Schools; HS = high school; METCO = Metropolitan Council for Educational Opportunity; IEP = Individualized Education Plan.

p < .05. **p < .01.

I find somewhat smaller differences between METCO participants and students in Boston charter schools, a comparison missing from previous work. The METCO on-time graduation rate is about 25 percentage points higher than nonapplicants in Boston charter schools (Table 5), but the corresponding difference in immediate college enrollment is only six percentage points (Table 6). The latter is still statistically significant at the .05 level. Given that almost 20% of METCO applicants who are not active in the program as ninth graders are enrolled in charter schools, this much smaller difference in outcomes is important in gauging the METCO effect.

The results discussed thus far derive from a model with no statistical controls for observed differences between these groups. However, in the second column of the tables, I include student demographic covariates and obtain very similar estimates. The adjusted high-school graduation rate of METCO Actives is 35 percentage points higher than similar BPS nonapplicants and 30 points higher than similar Boston charter school nonapplicants. For college enrollment, the adjusted METCO-BPS and METCO-charter differences are 32 and 11 points, respectively.

These large differences in college matriculation are due almost entirely to enrollment in 4-year, not 2-year, institutions. In results not shown, I find that METCO graduates proceed immediately to 4-year colleges and universities at an adjusted rate that is 30 percentage points higher than BPS graduates and 20 points higher than charter school graduates. The METCO enrollment rate in 2-year institutions, by contrast, is only three percentage points more than BPS and not significantly different from the rate for Boston charter alumni.

Findings From the Subsample With Eighth-Grade Test Scores

A major analytical challenge is that the student-level covariates available to adjust for group differences are quite limited. While the adjusted model above includes many demographic controls, the differences in outcomes may be attributable to unobserved traits like academic skills, student motivation, and parental education and involvement, rather than to METCO. The nonrandom selection of Boston students into schooling environments implies that the groups of program applicants and nonapplicants very likely differ on these traits.

To probe the degree of selection bias, I fit the original model with two new covariates: the student’s eighth-grade math and ELA MCAS scores. Doing so restricts the sample to students with nonmissing eighth-grade MCAS scores, resulting in the exclusion of 16.5% of cases. These students completed eighth grade in a different state or in a private school in Massachusetts.

As mentioned above, eighth-grade scores are endogenous. Most families decide to apply to METCO long before their child’s eighth-grade year, and the majority of Actives have been enrolled in the program for several years by then. Therefore, the inclusion of eighth-grade scores effectively controls away some of the true program effect. I return to this issue below, but here include these scores to assess differences on outcomes for students of similar academic ability when they were entering high school. In Tables 5 and 6, I fit the original model without eighth-grade test scores in the smaller subsample and obtain slightly smaller estimated differences than with the full sample.

In models shown in column 4, the reduction in the estimates with the addition of eighth-grade scores is quite modest at about 5 to 10 percentage points. METCO participants graduate high school within 4 years at a rate 26 percentage points higher than nonapplicants with similar eighth-grade scores. For immediate college enrollment, the adjusted METCO-BPS difference is comparable at 25 points, while the adjusted METCO-charter difference shrinks to 8 percentage points. The main finding is robust to the inclusion of controls for prior academic performance, even though the controls are endogenous: METCO high-school participants have higher high-school graduation and college enrollment rates than similar students in BPS and Boston charter high schools.

Approach 2: Referred Versus Unreferred METCO Applicants

My preferred identification strategy is to compare those applicants who were successful in securing a METCO referral to those who were not.¹² While problems remain, this strategy represents a major advance over previous estimates, including the ones I report above. The available evidence indicates that the applicant population differs in many ways from Boston students who never applied to METCO. Limiting the analysis to applicants yields treatment and control groups who are presumably more similar on unobservable characteristics correlated with the outcomes of interest (e.g., parental motivation, persistence, access to information, and prioritization of their children’s education). In the end, the results from both samples are quite consistent.

My goal is to estimate the total effect of a METCO referral on on-time high-school graduation and immediate college enrollment. In the sample of METCO applicants, the identifying assumption underlying this estimation for outcome Y for student i at time t is

E [Y_{0 i t} | Y_{i (t - s)}, X_{i}] = E [Y_{0 i t} | Y_{i (t - s)}, X_{i}, R e f_{i}],

(2)

where Ref_i takes a value of 1 for applicants referred in their first year of METCO eligibility and 0 otherwise, X_i is a vector of student-level demographic covariates, and Y_i_(t-s) is the student’s unobserved propensity to graduate from high school on time as it existed when the student first applied to METCO (t-s). If this ideal conditioning variable existed such that the equality in Equation 2 held, then METCO referral would be as good as randomly assigned, conditional on the student’s pretreatment propensity to graduate and the other X_i covariates.

However, instead of Y_i(t-s), I have students’ eighth-grade MCAS scores in math and ELA. These variables are useful in that they are quite likely to be positively correlated with the conditioning variable. The drawback is that, as noted above, this measure of student achievement is taken many years after most students are actually first eligible for METCO, so it is a function of both the unobserved propensity and program referral:

M C A S 8 t h_{i} = θ Y_{i (t - s)} + π R e f_{i} + η_{i} .

(3)

I assume that part of the total effect of a METCO referral occurs before high school and that it is positively correlated with eighth-grade scores.¹³ Therefore, the inclusion of eighth-grade scores in my analytic models effectively controls away some of the program effect, so that my estimates likely understate the true effect.

I fit the following model using eighth-grade scores:

Y_{i} = β M C A S 8 t h_{i} + δ R e f_{i} + γ X_{i} + a_{g t} + ϵ_{i},

(4)

where a_gt represents fixed effects for grade-by-year of eligibility, to compare applicants who applied at the same time to the same grade. By substitution,

Y_{i} = β {θ Y_{i (t - s)} + π R e f_{i} + η_{i}} + δ R e f_{i} + γ X_{i} + a_{g t} + ϵ_{i} .

(5)

Y_{i} = β θ Y_{i (t - s)} + (β π + δ) R e f_{i} + γ X_{i} + a_{g t} + (ϵ_{i} + β η_{i}) .

(6)

In this model, the “true” effect of referral is $(β π + δ) .$ The estimate of the effect of a referral from Equation 4, $\hat{δ}$ , will be biased downward. The bias depends on both β and π, with the magnitude of the bias contingent on the size of δ. However, in results from a companion analysis, I obtained estimates of .1 SD or less for π, the effect of a referral on the standardized test scores of METCO applicants. So the bias here is likely to be small.

To estimate Equation 4, I use data on applicants in three birth cohorts of students, most of whom entered ninth grade for the first time in the fall of 2007, 2008, and 2010. The sample also includes students who later joined these cohorts for a variety of reasons, including retention in grade, skipping a grade, and transferring from districts that used different birthdate cutoffs for kindergarten entry. Restricting the sample to only those in the three birth cohorts does not change the results I present in the next section. I treat students who transferred out of Massachusetts public schools before 12th grade as missing on high-school graduation and college enrollment outcomes. In a sensitivity analysis reported in a subsequent section, I establish Lee bounds around my main estimates to account for this program attrition.

Descriptive statistics on this analytic sample appear in Table 7. I have a total of 2,010 students in the sample after excluding applicants who were first eligible for 10th grade or later.¹⁴ Approximately 83% received a METCO referral in the first year they were eligible for the program.

Table 7

Student Demographics and Mean Test Scores for METCO Applicants in Three Birth Cohorts, by Whether They Received a Referral to a Suburban District in the First Year They Were Eligible for METCO Placement

Variable	Referred	Unreferred	Total
Low-income	0.53	0.72	0.56
Female	0.52	0.43	0.51
ELL	0.06	0.05	0.06
IEP	0.06	0.14	0.08
Immigrant	0.04	0.03	0.03
Asian	0.04	0.05	0.04
Black	0.77	0.74	0.76
Latino	0.16	0.18	0.16
White	0.02	0.02	0.02
MCAS 8th-grade math	–0.35	–0.56	–0.38
MCAS 8th-grade ELA	–0.19	–0.54	–0.25
N	1,674	336	2,010

Note. MCAS scores standardized within grade and year, using statewide mean and SD. METCO = Metropolitan Council for Educational Opportunity; ELL = English language learners; IEP = Individualized Education Plan; MCAS = Massachusetts Comprehensive Assessment System; ELA = English language arts.

Dissimilarities in the demographic composition of the referred and unreferred groups are obvious in Table 7, consistent with nonrandom selection for referrals. Compared with applicants who received a referral, the “Unreferred” group has much larger proportions of low-income students, male students, and students who already had an IEP at the time they were first eligible for METCO. The marked difference on the income variable is indicative of deviations from the METCO, Inc., placement algorithm, which does not include family income. Unreferred applicants also scored .2 SD lower on average on the eighth-grade math MCAS and .32 SD lower in ELA, although this is not a baseline measure.

I find large, positive differences in the proportion of referred students who graduated high school on time and who enrolled in college the fall after their graduation, compared with students who did not receive a referral. In Table 8, the unadjusted model in the first column contains only the Referred predictor and fixed effects for grade-by-year of eligibility. The probabilities of on-time high-school graduation and immediate college enrollment are both 20 percentage points higher for referred students, and these differences persist after including demographic covariates in the model (second column).

Table 8

Results From High-Density Fixed-Effects Regressions Predicting On-Time High-School Graduation of METCO Applicants in Three Birth Cohorts (Standard Errors in Parentheses)

	Full sample		Subsample with 8th-grade scores
Predictor	Unadjusted	Adjusted	Adjusted (No scores)	Adjusted (With scores)
A. On-time high-school graduation
Referred	0.20** (0.028)	0.18** (0.028)	0.15** (0.029)	0.13** (0.029)
Eighth-grade math score				0.050** (0.015)
Eighth-grade ELA score				0.063** (0.015)
Demographic controls	No	Yes	Yes	Yes
N	1,550	1,550	1,399	1,399
R²	.121	.143	.145	.185
B. Immediate college matriculation
Referred	0.20** (0.034)	0.17** (0.034)	0.16** (0.036)	0.13** (0.036)
Eighth-grade math score				0.074** (0.019)
Eighth-grade ELA score				0.061** (0.019)
Demographic controls	No	Yes	Yes	Yes
N	1,547	1,547	1,397	1,397
R²	.078	.114	.119	.157

Note. All models include fixed effects for grade-by-year of original eligibility. Adjusted models include dummies for Asians, Latinos, Whites, other race/ethnicity, free or reduced-price lunch status, Individualized Education Plan (IEP), English language learners (ELL), and immigrant status; these parameter estimates are omitted to save space. METCO = Metropolitan Council for Educational Opportunity; ELA = English language arts.

*p < .05. **p < .01.

I again find in results not shown that matriculation into 4-year, not 2-year, colleges and universities is driving the college enrollment difference. In fact, referred and unreferred students go on to enroll in 2-year programs at rates that are not statistically different. But the probability of attending a 4-year institution is 15 percentage points higher for the referred group, after controlling for student demographics.

These differences are quite robust to the inclusion of eighth-grade test scores as covariates. As before, I fit the original model within the group of applicants with nonmissing eighth-grade MCAS scores and obtain slightly smaller adjusted differences in outcomes. Adding the scores themselves has little effect on the parameter estimate for Referred. Although including eighth-grade scores controls away part of the METCO effect, referred METCO applicants still graduated from high school on time and immediately went on to college at rates 13 percentage points higher than did unreferred applicants with similar levels of academic achievement in eighth grade.

Subgroup and Dosage Effects

I find no evidence of differential effects by income or gender subgroup. The inclusion of an interaction with family income yields no statistically significant differences in the impact of a METCO referral between low-income students and those with more family advantages.

Following Bergman (2016), I also tested whether impacts differ for boys and girls by interacting gender with receipt of a METCO referral. Unlike Bergman, who found larger college enrollment differences for boys, I find no statistically significant gender differences in impact on any of the outcomes studied here. Indeed, the interaction coefficients indicate a 7-point difference in favor of girls, not boys, for both high-school graduation (p = .22) and college enrollment (p = .28). With eighth-grade scores included as a covariate, the estimates shrink to about two points.

One other question of interest concerns heterogeneity of program exposure. Students who participate in METCO for more of their K–12 careers might be expected to have better outcomes, on average, than those whose tenure in the program is short-lived. Unfortunately, for most students in the sample, their first year of kindergarten eligibility is prior to 2002, the first year of SIMS enrollment data. For the most recent cohort, I investigate dosage by adding a linear term for the number of years of METCO participation, which ranges from 0 to 13, to the model; quadratic and cubic terms were not significant. In this much smaller sample (n = 451), the main effect of a METCO referral is a 5 percentage point difference in the probability of high-school graduation, with an additional point for each year of program enrollment. Receiving a referral is associated with a difference of 10 percentage points in the probability of college enrollment, with an additional 1.6 points per year in the program.

However, these findings on program dosage cannot be interpreted as causal effects. Students who entered the program as 5-year-olds and stayed through 12th grade likely differ on many unobservables from those who entered later, or who entered at the same time but left the program much earlier.

Selection Bias

The nonrandom selection of applicants for referrals means that omitted variables bias may affect the main estimates. To probe the extent of this bias, I included eighth-grade math and ELA test scores in my models and obtained slightly smaller but similar coefficients. While academic achievement is only one of many potential sources of omitted variables bias, eighth-grade achievement scores are powerful predictors of college enrollment (Bowen et al., 2009). Moreover, including them as covariates presumably controls away some of the true METCO effect, given that most students apply to the program well before their eighth-grade year. Yet adding these scores to the model decreases the estimates by only 14% for on-time high-school graduation and 17% for immediate college matriculation.

Oster (2016) points out that coefficient stability is insufficient evidence on its own, unless coefficient movements are scaled by the changes in the R² statistic when controls are included. Using her approach, I calculate a bias-adjusted treatment effect of .102 for high-school graduation and .099 for immediate college enrollment.¹⁵ These estimates are somewhat smaller than those reported in the final column of Table 8, but still constitute very large impacts on two central outcomes targeted by policymakers and linked to students’ later labor-market success.

Any remaining omitted variables are likely correlated with eighth-grade scores. While there is no way to ascertain the actual extent of the resulting bias, I can quantify how large it would have to be to invalidate the inference in each case (Frank et al., 2013). For models that include eighth-grade test scores, fully 57% of the estimated effect of a program referral on high-school graduation would need to be due to bias; the corresponding figure for college enrollment is 47%.¹⁶ Given that these calculations are conditional on eighth-grade scores, it seems highly improbable that the remaining omitted variables bias could be that size.

It is also instructive to compare the effect of a hypothetical omitted variable to the strongest observed effect among the covariates in the model—in this case, students’ eighth-grade ELA test score. The combined impacts of ELA score on high-school graduation and college enrollment are both approximately .03, which is the product of the correlations with the outcome and with the Referred predictor. To invalidate the inference, a remaining omitted variable would need to have an impact of .07 on high-school graduation, which is more than twice as large as that of the strongest observed covariate. For college enrollment, the impact would need to be .05, or about 1.7 times as large.

Missing Data

The data on certain METCO applicants are missing on two key dimensions: high-school and college outcomes, which is due to attrition from the sample before high-school graduation, and METCO application information, due to incomplete record-keeping. If data are missing at random, conditional on the observed variables in my model, then the exclusion of these cases from a regression model does not bias the estimates. However, the assumption that the probability of missingness depends only on observables is very strong and difficult to prove. The possibility that missing cases differ systematically from the “complete cases” means that estimates using only the latter may be biased (Gelman & Hill, 2007). I address the two types of missing data in turn.

Sample Attrition

METCO applicants who exit the Massachusetts public school system, which is the source of my K–12 data, before 12th grade are missing the high-school graduation outcome. Of the 452 students missing on HSgrad, 197, or 44%, left the Massachusetts public schools before ninth grade. This group was too young to drop out, so they either left the state or switched to private schools; many of them presumably went on to graduate from high school on time. The other 56% exited in ninth grade or later as transfers, or their enrollment records are too sparse to determine whether they graduated within 4 years of entering ninth grade.

To gauge the effect of these cases, I calculate Lee bounds for the main estimates I obtained from the sample of METCO applicants. This approach involves trimming from the group that experienced less attrition—here, the referred group—until the proportion of missing observations is equal for both groups (Lee, 2009). In Table 9, I present the unconditional upper and lower Lee bounds for the full sample. For on-time high-school graduation, the Lee bounds are 0.20 to 0.35, a range encompassing only large positive effects. The results are similar for college enrollment.

Table 9

Sensitivity Analyses Calculating Lee Bounds on the Estimates From the Prediction of On-Time High-School Graduation (Panel A) and Immediate College Matriculation (Panel B) of METCO Applicants in Three Birth Cohorts (Standard Errors in Parentheses)

	Full sample		Subsample with 8th-grade scores
Predictor	Lower bound	Upper bound	Lower bound	Upper bound
A. On-time high-school graduation
Referred	0.20** (0.035)	0.35** (0.051)	0.15** (0.038)	0.24** (0.045)
Demographic controls	No	No	No	No
N	1,558	1,558	1,407	1,407
B. Immediate college matriculation
Referred	0.18** (0.039)	0.33** (0.047)	0.13** (0.042)	0.25** (0.046)
Demographic controls	No	No	No	No
N	1,553	1,553	1,403	1,403

Note. Estimates for subsample with eighth-grade scores are bounds that have been tightened using the quintiles of students’ eighth-grade math scores. METCO = Metropolitan Council for Educational Opportunity.

*p < .05. **p < .01.

I follow the approach recommended by Lee to tighten the bounds further, using the quintile of eighth-grade math MCAS scores as a covariate. Within each quintile, bounds are calculated separately, and then a weighted average of the results is generated. The resulting upper and lower bounds around the high-school graduation effect are 0.15 and 0.24, with similar bounds for college enrollment. Had sample attrition not occurred, the differences on the outcomes would still likely have been positive and of nearly the same magnitude.

Missing METCO Application Data

The second type of missing data is information on METCO referrals for some children who completed all application requirements for the program. For 208 applicants, their paper and electronic records indicate that they were referred to a suburban district, but omit the date when this occurred. DESE did not begin electronic data collection of student information until the 2001–2002 school year, or up to 3 years after the children in this sample started kindergarten. So I was unable to verify the complete enrollment histories of these students in SIMS.

Because my main models include fixed effects for the grade and year of initial METCO eligibility, to account for differences in the age of applicants and the time they chose to apply to the program, these students are excluded from the analytic sample. I tested the sensitivity of my results to this exclusion by dropping the fixed effects and refitting Equation 4 with and without eighth-grade test scores. For both on-time high-school graduation and immediate college enrollment, the results were unaffected.

However, another 691 applicants who would otherwise be in the sample are missing on whether they received a referral in their first year of eligibility or not, which is the treatment indicator. For this group, I could not locate their paper application folders, which in many cases dated back to the late 1990s, and METCO’s electronic records contained no information on whether they ever received a referral or not. Many such students later showed up as active METCO participants in the SIMS data, in which case I inferred that they had been referred without a record being made of it. But for most of these applicants, I have no way of determining whether they received a referral or not.

I conducted three sensitivity analyses to explore the effect of this type of missing data. I constructed three versions of the Referred predictor: one with unknown students coded as referred, a second with them coded as unreferred, and a final predictor with three categories: referred, unreferred, and missing. The results of models using these three versions appear in Table 10.

Table 10

Sensitivity Analyses Imputing Missing Referral Data in the Prediction of On-Time High-School Graduation (Panel A) and Immediate College Matriculation (Panel B) of METCO Applicants in Three Birth Cohorts (Standard Errors in Parentheses)

	Full sample			Subsample with 8th-grade scores
Predictor	Missing as referred	Missing as unreferred	Missing as separate group	Missing as referred	Missing as unreferred	Missing as separate group
A. On-time high-school graduation
Referred	0.17** (0.028)	0.14** (0.025)	0.18** (0.028)	0.13** (0.029)	0.10** (0.025)	0.13** (0.029)
Referred—missing			0.11* (0.047)			0.11* (0.049)
Eighth-grade math score				0.051** (0.015)	0.052** (0.015)	0.051** (0.015)
Eighth-grade ELA score				0.061** (0.015)	0.061** (0.015)	0.061** (0.015)
N	1,674	1,674	1,674	1,503	1,503	1,503
R²	.141	.139	.142	.183	.180	.183
B. Immediate college matriculation
Referred	0.17** (0.035)	0.13** (0.030)	0.17** (0.035)	0.13** (0.036)	0.092** (0.031)	0.13** (0.036)
Referred—missing			0.13* (0.058)			0.13* (0.061)
Eighth-grade math score				0.068** (0.019)	0.070** (0.019)	0.068** (0.019)
Eighth-grade ELA score				0.068** (0.018)	0.069** (0.018)	0.068** (0.018)
N	1,668	1,668	1,668	1,499	1,499	1,499
R²	.106	.103	.106	.152	.150	.152

Note. All models include fixed effects for grade-by-year of original eligibility. Student demographic controls include dummies for Asians, Latinos, Whites, other race/ethnicity, free or reduced-price lunch status, Individualized Education Plan (IEP), English language learners (ELL), and immigrant status; these parameter estimates are omitted to save space. METCO = Metropolitan Council for Educational Opportunity; ELA = English language arts.

*p < .05. **p < .01.

For both outcomes, the number of restored cases appears to be too small to exert much influence on the estimated effect. Unfortunately, most of the students of unknown referral status are missing other data as well: one or both of the outcome variables, and/or the year and grade they were initially eligible for METCO. In the high-school graduation results shown in Panel A, only 124 cases have been restored to the sample, or about 18% of the students for whom referral status is unknown. For college enrollment in Panel B, 121 cases have been restored. I do obtain very similar estimated effects regardless of how these cases are handled.

Discussion

This study offers the first estimates of differences on intermediate educational outcomes for participants in interdistrict integration programs that (a) adjust for observed dissimilarities in group demographics and (b) compare participants to similar students enrolled in local charter schools as well as district-run public schools.

I find that Boston students enrolled in METCO at the beginning of ninth-grade graduate from high school within 4 years at a rate approximately 30 percentage points higher than demographically similar students attending BPS and Boston charter schools. The difference in college matriculation is similar for METCO versus BPS students, but a more modest 11 percentage points when METCO participants are compared with similar students attending charter high schools.

This study is also the first to estimate differences between students who were offered a METCO referral and those who were not. These children are arguably more equal in expectation than to Boston children who did not apply to the program. My estimates here approximate intent-to-treat effects, since students who received a referral comprise the intended target population of the METCO intervention even though many did not actually enroll.

Within the restricted sample of applicants, I again find pronounced differences on intermediate outcomes. Students who received a METCO referral had an adjusted high-school graduation rate approximately 18 percentage points higher than students who did not, while their college enrollment rate was about 17 points higher. Unlike Bergman (2016), I find that the latter difference is due primarily to enrollment in 4-year, not 2-year, institutions, and that the magnitude of this difference is statistically equivalent for girls and boys.

In a study of METCO alumni, almost all expressed the belief that the program’s benefits were worth the challenges that participation entailed, and the results here offer some empirical support for this stance (Eaton, 2001). Students who enter ninth grade in METCO are far more likely to graduate high school on-time than their peers in Boston’s public high schools, and they are much more likely to proceed immediately to a 4-year college or university; this is true after controlling for demographics and even an endogenous measure of academic achievement. While I do not have data on college completion, METCO students are passing more successfully through the educational pipeline up to that point.

These results suggest that the benefits of attending advantaged suburban schools outweigh the potential negative aspects of METCO participation. These benefits include the culture of attainment and college-going in suburban schools, the access METCO students have to information about college planning and financial assistance, and the connections they develop that lead to educational and internship opportunities. Another possible explanation for my findings relates to the presence of fewer students with academic challenges in suburban schools. If these schools have relatively few low-performing students and relatively more resources, staff can more effectively support those students in fulfilling graduation requirements. In fact, each receiving district has a METCO director who monitors the academic progress of participants and connects them with tutoring and other supports when necessary. In urban settings where there are more students at risk of not graduating, the available resources are spread far more thinly. Further research is necessary to distinguish between these potential factors and identify those most critical to the success of METCO students at the high-school level.

The large positive impacts here are extremely unlikely to be the product of selection bias alone and carry implications for policymakers looking to address persistent racial/ethnic gaps in educational attainments. The replication of programs like METCO to other metropolitan areas in Massachusetts and around the nation appears to be a promising strategy in boosting the intermediate educational outcomes of urban students of color. While cross-district programs are limited in their reach (METCO only enrolls about 4% of the students in Boston), the results of this study indicate they confer a valuable opportunity on those students who are selected to participate.

Of course, to draw fully credible causal inferences about cross-district transfer programs, two prerequisites must be in place: accurate recordkeeping of students’ placement histories and school performance, and rigid implementation of a selection algorithm that would enable at least the referral decision to be treated as randomly assigned, conditional on the variables in the algorithm. While the discretion exercised by participating suburban districts in admitting applicants has probably contributed to the Boston program’s longevity, it is less clear why the referral of eligible applicants to these districts need be subject to the discretion of placement staff. In 2019, METCO switched to a randomized applicant lottery to allocate referrals to suburban districts. This change will finally allow the application of quasi-experimental methods to an impact evaluation, once the students from the first few lotteries progress through high school and beyond. The design of all such programs should feature a selection process that is transparent, strictly followed, and approximates random assignment as closely as possible.

Footnotes

Acknowledgements

The author is grateful to Eric Taylor, Richard Murnane, Paul Reville, John Papay, Carrie Conaway, Daniel Koretz, Jal Mehta, Kevin Lang, Felipe Barrera-Osorio, and three anonymous reviewers for helpful comments and suggestions. Jean McGuire, John Shandorf, Milly Arbaje-Thomas, and Donna Washington provided important insights into METCO program operations. Thanks to METCO, Inc., and the Massachusetts Department of Elementary and Secondary Education for sharing their data.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study received generous financial support from the Harvard University Inequality and Social Policy Program.

Notes

Author

ANN MANTIL is a postdoctoral research associate at Brown University. She studies charter schools, desegregation and school choice programs, and the equity impacts of educational policies and interventions.

References

Altonji

J. G.

Elder

T. E.

Taber

C. R.

(2005). Selection on observed and unobserved variables: Assessing the effectiveness of Catholic schools. Journal of Political Economy, 113(1), 151–184.

Autor

D. H.

(2014). Skill, education, and the rise of earnings inequality among the “other 99 percent.” Science, 344(6186), 843–851.

Bergman

P. L. S.

(2016, October). The effects of school integration: Evidence from a randomized desegregation program (CESifo Working Paper Series No. 6119). https://ssrn.com/abstract=2871152

Bifulco

Cobb

C. D.

Bell

(2009). Can interdistrict choice boost student achievement? The case of Connecticut’s interdistrict magnet school program. Educational Evaluation and Policy Analysis, 31(4), 323–345.

Bowen

W. G.

Chingos

M. M.

McPherson

M. S.

(2009). Crossing the finish line. Princeton University Press.

Boyd

Lankford

Loeb

Ronfeldt

Wyckoff

(2011). The effect of school neighborhoods on teachers’ career decisions. In Duncan

G. J.

Murnane

R. J.

(Eds.), Whither opportunity? Rising inequality, schools, and children’s life chances (pp. 377–395). Russell Sage Foundation.

BPS Communications Office. (2016). Boston Public Schools at a glance: 2016-17. https://www.bostonpublicschools.org/cms/lib/MA01906464/Centricity/Domain/238/BPS%20at%20a%20Glance%202016-17_online.pdf

Bruhn

(2020). The consequences of sorting for understanding school quality [Unpublished manuscript].

Chapman

T. K.

(2014). Is integration a dream deferred? Students of color in majority white suburban schools. Journal of Negro Education, 83(3), 311–326.

10.

Crain

R. L.

(1992). Finding niches: Desegregated students sixteen years later. Institute for Urban and Minority Education, Columbia University.

11.

de Brey

Musu

McFarland

Wilkinson-Flicker

Diliberti

Zhang

Branstetter

Wang

(2019). Status and trends in the education of racial and ethnic groups 2018 (NCES 2019-038). National Center for Education Statistics, U.S. Department of Education.

12.

Diamond

J. B.

(2006). Still separate and unequal: Examining race, opportunity, and school achievement in “integrated” suburbs. Journal of Negro Education, 75(3), 495–505.

13.

Eaton

S. E.

(2001). The other Boston busing story: What’s won and what’s lost across the boundary line. Yale University Press.

14.

Eaton

S. E.

Chirichigno

(2011). METCO merits more: The history and status of METCO (White Paper No. 74). Pioneer Institute.

15.

Executive Office of Education. (2013). METCO report to the Massachusetts State Legislature.

16.

Frank

K. A.

Maroulis

S. J.

Duong

M. Q.

Kelcey

B. M.

(2013). What would it take to change an inference? Using Rubin’s causal model to interpret the robustness of causal inferences. Educational Evaluation and Policy Analysis, 35(4), 437–460.

17.

Gelman

Hill

(2007). Data analysis using regression and multilevel/hierarchical models. Cambridge University Press.

18.

Harris

D. N.

(2010). How do school peers influence student educational outcomes? Theory and evidence from economics and other social sciences. Teachers College Record, 112(4), 1163–1197.

19.

Holland

(2012). Only here for the day: The social integration of minority students at a majority white high school. Sociology of Education, 85(2), 101–120.

20.

Ispa-Landa

(2013). Gender, race, and justifications for group exclusion: Urban Black students bussed to affluent suburban schools. Sociology of Education, 86(3), 218–233.

21.

Ispa-Landa

Conwell

(2015). “Once you go to a White school, you kind of adapt”: Black adolescents and the racial classification of schools. Sociology of Education, 88(1), 1–19.

22.

Jencks

Mayer

(1990). The social consequences of growing up in a poor neighborhood. In Lynn Jr.

L. E.

McGeary

M. G. H.

(Eds.), Inner-city poverty in the United States (pp. 111–186). The National Academies Press.

23.

Lazear

(2001). Educational production. Quarterly Journal of Economics, 116(3), 777–803.

24.

Lee

D. S.

(2009). Training, wages, and sample selection: Estimating sharp bounds on treatment effects. The Review of Economic Studies, 76(3), 1071–1102.

25.

Lewis

A. E.

Diamond

J. B.

(2016). Despite the best intentions: How racial inequality thrives in good schools. Oxford University Press.

26.

Loeb

Darling-Hammond

Luczak

(2005). How teaching conditions predict teacher turnover in California schools. Peabody Journal of Education, 80, 44–70.

27.

Massachusetts Department of Elementary and Secondary Education. (2006). Calculating Massachusetts public high school graduation rates: Project overview. http://www.doe.mass.edu/infoservices/reports/gradrates/calculating_overview.html

28.

McFarland

Hussar

Zhang

Wang

Hein

Diliberti

Forrest Cataldi

Bullock Mann

Barmer

(2019). The condition of education 2019 (NCES 2019-144). National Center for Education Statistics, U.S. Department of Education.

29.

O’Connor

Mueller

Lewis

R. L.

Rivas-Drake

Rosenberg

(2011). “Being” Black and strategizing for excellence in a racially stratified academic hierarchy. American Educational Research Journal, 48(6), 1232–1257.

30.

Orfield

Frankenberg

Garces

(2008). Statement of American social scientists of research on school desegregation to the U.S. Supreme Court in Parents v. Seattle School District and Meredith v. Jefferson County. The Urban Review, 40(1), 96–136.

31.

Oster

(2016, August). Unobservable selection and coefficient stability: Theory and evidence (NBER Working Paper No. 19054). National Bureau of Economic Research. https://emilyoster.net/s/Unobservable-Selection-and-Coefficient-Stability-Theory-and-Evidence.pdf

32.

Owens

(2010). Neighborhoods and schools as competing and reinforcing contexts for educational attainment. Sociology of Education, 83(4), 287–311.

33.

Reardon

S. F.

Townsend

Fox

(2017). A continuous measure of the joint distribution of race and income among neighborhoods. RSF: The Russell Sage Foundation Journal of the Social Sciences, 3, 34–62.

34.

Reardon

S. F.

Weathers

E. S.

Fahle

E. M.

Jang

Kalogrides

(2019, September). Is separate still unequal? New evidence on school segregation and racial academic achievement gaps (CEPA Working Paper No. 19-06). http://cepa.stanford.edu/wp19-06

35.

Rumberger

Palardy

(2005). Does segregation still matter? The impact of student composition on academic achievement in high school. Teachers College Record, 107(9), 1999–2045.

36.

Ryan

C. L.

Bauman

(2016, March). Educational attainment in the United States: 2015. U.S. Census Bureau.

37.

Sacerdote

(2011). Peer effects in education: How might they work, how big are they, and what do we know so far? In Hanushek

E. A.

Machin

Woessman

(Eds.), Handbook of the economics of education (Vol. 3, pp. 249–277). North Holland.

38.

Steele

J. L.

Pepper

M. J.

Springer

M. G.

Lockwood

J. R.

(2015). The distribution and mobility of effective teachers: Evidence from a large, urban school district. Economics of Education Review, 48, 86–101.

39.

Theokas

Saaris

(2013). Finding America’s missing AP and IB students. The Education Trust. https://edtrust.org/resource/finding-americas-missing-ap-and-ib-students/

40.

Warikoo

Carter

(2009). Cultural explanations for racial and ethnic stratification in academic achievement: A call for a new and improved theory. Review of Educational Research, 79(1), 366–394.

41.

Wells

A. S.

Crain

R. L.

(1997). Stepping over the color line: African-American students in White suburban schools. Yale University Press.

42.

Frank

K. A.

Maroulis

S. J.

Rosenberg

J. M.

(2019). Konfound: Command to quantify robustness of causal inferences. The Stata Journal, 19(3), 523–550.