The US Supreme Court overturned affirmative action in 2023 1 ; however, higher-education institutions continue to make admissions decisions that affect the racial diversity of their student cohorts 2 , 3 , 4 . Therefore, it is important to know whether racial diversity in an educational cohort is associated with higher or lower student cohort salaries at graduation. Learning theory argues that racial diversity promotes student learning, which should increase salaries 5 , 6 , 7 , 8 , 9 . However, well-documented racial wage discrimination indicates that higher racial diversity should decrease salaries 10 , 11 , 12 , 13 , 14 . As highlighted in the recent Supreme Court decision, there is no empirical evidence on racial diversity’s association with student cohort salaries. Here, to address this gap, we compile two unique and comprehensive datasets: 2,964 Master of Business Administration cohorts across 141 business schools over 29 years and 3,386 Juris Doctor cohorts across 200 law schools over 21 years. In both datasets, we find that higher cohort racial diversity is associated with higher cohort median salaries at graduation across numerous model specifications and after controlling for student quality, universities and years. The key implication is that policies to increase or leverage racial diversity (for example, affirmative action and diversity, equity and inclusion programmes) enhance human capital and benefit society.

Main

The US Supreme Court recently made race-based affirmative action illegal in higher-education admissions decisions. Although the court still allows “considering an applicant’s discussion of how race affected his or her life,” 1 this new approach has led many universities to have fewer minority students in recent matriculating classes 2 , 3 , 4 . The economic impact of this change in racial diversity is unclear because the association between student cohort racial diversity and student cohort salary has not been addressed in previous research. A student cohort (used interchangeably with ‘cohort,’ in this paper) is defined as a group of students who start and finish an education programme together and share the curricular, co-curricular and extracurricular learning experiences offered by the programme.

The Supreme Court decision overturning affirmative action relies on at least three key reasons. “First, the interests that [Harvard and UNC] view as compelling cannot be subjected to meaningful judicial review” because “it is unclear how courts are supposed to measure any of these goals” (page 6 in ref. 1 ). These interests include training future leaders, preparing graduates for a pluralistic society, enhanced education through diversity, generating new knowledge from diverse perspectives, enhancing cross-racial understanding and breaking stereotypes (also see pages 307, 308 and 332 in ref. 15 ). “Second, [Harvard’s and UNC’s] admissions programs fail to articulate a meaningful connection between the means they employ and the goals they pursue” (page 6 in ref. 1 ). Third, “[Harvard’s and UNC’s] admissions programs also lack a ‘logical end point’” (page 7 in ref. 1 ). Our research addresses these three concerns by specifying a measurable input (racial diversity percentage of student cohort at matriculation), a clearly connected, measurable earnings outcome (that is, cohort median salary at graduation), and a means for assessing when an economic goal of racial diversity has been achieved (that is, when increases in cohort racial diversity are no longer associated with increases in cohort median starting salary).

Justices opposing affirmative action have not been persuaded by current evidence on diversity’s educational benefits. One reason could be that while educational benefits are grounded in cohort-wide arguments (that is, a diverse student cohort should provide educational benefits across the entire cohort owing to learning) 5 , 6 , 7 , 8 , 9 , empirical research focuses on self-selected individuals’ experiences and outcomes 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , which are often subjective (Supplementary Information section 1 and Supplementary Table 1). Moreover, these individual-level studies drawing on students from a handful of universities in a few selected years cannot rule out alternative explanations related to universities (for example, curriculum, applicant pool, prestige, size, location) or specific years (for example, economic growth, inflation, unemployment). Therefore, we currently do not know whether a more racially diverse student cohort will have a higher or lower median starting salary, ceteris paribus, than a less racially diverse cohort.

Our investigation informs the overall impact of multiple competing forces that are likely to be operating simultaneously. In fact, some previous research argues that increasing racial diversity will lower student cohort salary owing to the consistent and persistent wage gap reflecting discrimination against racial minorities 10 , 11 , 12 , 14 . For a survey of recent Juris Doctor (JD) graduates that reveal a racial wage gap, see ref. 13 ; also see ref. 24 for a survey report on wage gaps for graduating Masters in Business Administration (MBA) students. Racial diversity’s potentially offsetting benefit is grounded in substantial previous research based on learning theory, which argues and demonstrates that racially diverse peers undertaking a shared learning experience (inside and outside classrooms) improve students’ cognitive and interpersonal skills, critical thinking and creativity, through consideration of different lived experiences and perspectives 25 , 26 , 27 , 28 , 29 , 30 , 31 (Supplementary Information section 1). While these studies argue in favour of diversity’s learning benefits and find empirical support, we present additional plausible mechanisms for the positive impact of racial diversity on salaries in Supplementary Information section 4. Prominent education leaders, including presidents of 62 research universities, assert that the “fundamental rationale for student diversity in higher education [is] its educational value” 32 , 33 . Importantly, these educational benefits are “universal in that all students learn from [diversity], not just minority students” 34 . Moreover, interacting with diverse classmates prepares students to enter and lead an increasingly diverse workforce, and exposure to diverse students’ cultures leads individuals to increase civic engagement and racial understanding 5 , 7 , 8 , 9 , 18 , essentially enhancing the “ability of students to live and work effectively in a diverse society” 35 . A critical aspect of learning theory is that these educational benefits result from interactions among diverse students in an educational cohort as opposed to mere exposure 36 , 37 . Thus, these benefits are more likely to accrue in education settings with cohort-based programmes (that is, a class that starts and completes its education together and experiences a common curriculum) and, if so, should be captured in the labour market earnings of these graduates 38 , 39 . Overall, we seek to answer the question of whether higher racial diversity in cohort-based programmes at higher-education institutions is associated with lower, neutral or higher median cohort salaries at graduation.

Empirical approach

We compile data on cohort-wide median starting salary for student cohorts with varying racial diversity, both across multiple business and law schools and within each business school and law school over time, along with control variables. We operationalize cohort racial diversity in two ways: (1) narrowly using the percentage of minority students within each focal student cohort and (2) broadly using the percentages of minority students within the focal cohort as well as other cohorts in the programme during the same years as the focal cohort. These data are “the sort of evidence the courts appear to be requiring if they are to support race-sensitive admissions policies” (page 512 in ref. 30 ). Previous researchers may not have attempted this approach because of the paucity of well-defined cohorts in their datasets.

Notably, we have substantial variation in our two key variables (cohort racial diversity and cohort median starting salary) across schools and within schools over time for both business schools and law schools (Extended Data Table 1). Similarly, in Extended Data Table 2, we observe high frequencies of year-to-year increases and decreases for our focal variables (cohort racial diversity and cohort median starting salary) for both business schools and law schools. In Extended Data Table 3, we report correlations between our variables. We observe meaningful and statistically significant correlations between cohort median starting salary and both narrow and broad cohort racial diversity in MBA cohorts and JD cohorts.

Our focal variables (that is, narrow cohort racial diversity and cohort median starting salary) are reported in our dataset for 72% of the MBA cohort population in 141 business schools between 1993 and 2021 and 81% of the JD cohort population in 200 law schools between 1999 and 2019 (Extended Data Table 4). The unobserved cohorts appear to be missing at random, as categorizations of each focal variable as observed or unobserved are not significantly associated with any of the other observed variables. To the best of our knowledge, our data are the most comprehensive compiled so far for investigating this topic, because no extant study uses such a broad-based number of cohorts in two specific programmes to conduct time-series panel-data analyses.

Findings

Model-free evidence

We present descriptive evidence on the association between cohort racial diversity and cohort median starting salary. We standardize salary and racial diversity by calculating the number of standard deviations away from the mean of all cohorts across universities in each year. We then obtain the medians of these standardized salaries and racial diversities across all cohorts in all years. We classify each cohort in our sample as high (that is, above median) or low (that is, below median) on standardized salary and standardized racial diversity. In Table 1, we report frequencies for each of the four cells. For business schools, out of 2,964 cohorts, 966 with high narrow cohort racial diversity obtained high salary and 933 with low narrow cohort racial diversity obtained low salary. For law schools, out of 3,386 cohorts, 1,128 with high narrow cohort racial diversity obtained high salary and 1,134 with low narrow cohort racial diversity obtained low salary. The results are similar for broad cohort racial diversity. These results show that higher racial diversity is significantly associated with higher salaries (P < 0.01).

In Supplementary Fig. 1, we present scatterplots of cohort median starting salary versus cohort racial diversity (with a Lowess smoother) for MBA and JD cohorts for both narrow and broad cohort racial diversity. These scatterplots reveal an overall positive relationship across the range of our data.

Model results

Table 2 presents results from estimating equation (1) (see Methods) for both narrow and broad measures of cohort racial diversity.

We observe positive associations between cohort racial diversity and cohort median starting salary while including control variables, university fixed effects and year fixed effects. These positive associations are statistically significant for both business and law schools and for both narrow and broad measures of cohort racial diversity (two-tailed P < 0.05 using narrow cohort racial diversity and P < 0.01 using broad cohort racial diversity for MBA cohorts; P < 0.01 using either narrow or broad cohort racial diversity for JD cohorts).

Robustness

We assess robustness of our findings as reported in Table 2 to several alternative model specifications. First, using racial diversity (narrow cohort and broad cohort measures) as the focal variable and then adding fixed effects and other control variables, we find that racial diversity on its own explains 5–8% of the cohort median starting salary while university fixed effects account for about 80% of the variation (Extended Data Table 5). Second, although Hausman tests reject random effects in favour of fixed effects, we examine how alternative random-effect specifications impact the racial diversity parameter. Extended Data Table 6 shows that racial diversity’s parameters for the narrow cohort as well as the broad cohort are consistently positive and significant (P < 0.01 for 12 of 16 specifications, P < 0.05 for two specifications, P < 0.10 for one specification, and directionally consistent for the remaining specification). Third, we examine the possibility that common shocks interact with school-specific characteristics resulting in unique outcomes for each cohort by allowing each school to have a separate time trend. Extended Data Table 7 shows, for both narrow cohort and broad cohort racial diversity in business schools and law schools, that associations with cohort median starting salary are significantly positive (P < 0.01). Fourth, we consider the possibility that certain autocorrelated omitted variables including students’ salary expectations are driving our results. Therefore, in our model, we include lagged median starting salary that “controls for omitted variables” (p. 156 in ref. 40 ). In the absence of a state-dependent process, including lagged values of median starting salary renders the “coefficient of the exogenous variable [that is, cohort racial diversity in our case] towards zero” resulting in a conservative parameter estimate (p. 357 in ref. 41 ). Extended Data Table 8 reports that the parameters for narrow cohort and broad cohort racial diversity remain positive and significant (P < 0.01 for 3 of 4 specifications and P < 0.05 for the remaining specification). Fifth, we examine whether racial diversity has a nonlinear association with starting salary. We use threshold interactions at discrete levels of cohort racial diversity ranging from >5% to >35% in 5% intervals. For business schools, none of these interactions is significant. There is a direction change of the interaction terms from positive to negative when cohort diversity exceeds 30%, but these parameters are not quite statistically significant. For law schools, we find the interaction terms to be negative and statistically significant at racial diversity levels higher than 35% (P < 0.10; Extended Data Table 9). Thus, when racial diversity exceeds certain thresholds, an increase in diversity is likely to be associated with a smaller increase in median salary compared with that when racial diversity is lower. Importantly, the overall effect sizes in the range of our data remain positive. Sixth, instead of using racial diversity as a raw percentage, we use an alternative symmetric operationalization of diversity using the complement of a Herfindahl concentration index 21 , 22 , 23 . This diversity index increases as the proportions of majority and minority converge, decreases as they diverge, and is zero when either of the two proportions reaches 100% (Extended Data Table 10). The resulting parameter of this racial diversity index, for the narrow cohort as well as the broad cohort, is positive and significant for both MBA and JD cohorts. Seventh, we examine the impact of outliers (cohorts with extremely high or low racial diversity) by winsorizing racial diversity at the 5% level; that is, all data points below the 5th percentile are replaced with the value at the 5th percentile, and all data points above the 95th percentile are replaced with the value at the 95th percentile. Again, the racial diversity parameters remain positive and significant (P < 0.01 and P < 0.05 for the broad and narrow diversity measures, respectively, in MBA and P < 0.05 for both measures in JD) (Supplementary Table 6). Eighth, we assess robustness across universities by conducting two analyses. To understand patterns across schools, we estimate equation (2) (see Methods) by considering interactions of racial diversity with schools in (1) prestigious, (2) large or (3) urban universities. In Table 3, parameters for both narrow and broad cohort racial diversity are consistently positive and significant across three model specifications for both MBA and JD programmes. This means that racial diversity’s positive parameter is not driven by specific types of school and that for a school of average prestige, size and urban location, racial diversity’s parameter is positive.

Next, we allow school-specific racial diversity slopes in the fixed-effect model specification (Supplementary Fig. 2). We use an alternative fixed-effect individual slopes specification by including university-specific trends instead of year fixed effects. The results are similar for both model specifications. The average slopes for business and law schools are positive for narrow cohort diversity (111.44 for business schools and 331.14 for law schools) as well as broad cohort diversity (135.25 for business schools and 334.17 for law schools). For business schools, 40 of 141 narrow cohort racial diversity slopes are significantly positive while only 19 are significantly negative (P < 0.10). Likewise, 39 of 141 broad cohort racial diversity slopes are significantly positive while only 24 are significantly negative (P < 0.10). For law schools, 64 of 196 narrow cohort racial diversity slopes are significantly positive while only 28 are significantly negative (P < 0.10; 4 of the 200 law schools do not have enough annual observations to estimate school-specific slopes). The corresponding numbers for broad cohort racial diversity are 57 of 196 significantly positive with only 18 significantly negative (P < 0.10). Overall, many more business and law schools show a significantly positive racial diversity slope than show a significantly negative slope. Ninth, we assess whether the association between cohort racial diversity and cohort median starting salary in MBA programmes might vary according to proportions of Asian American students and the combination of Black and Hispanic students (Supplementary Table 7). Similar data on racial breakdowns are not available for JD programmes. After extending equation (1) by including an interaction with standardized minority student percentages (equation (3) in Methods), none of the interaction terms is statistically significant and the parameters of racial diversity at the point where minority student percentages are average are also not significant. This means that our results on the parameters for narrow and broad cohort racial diversity are not likely to be driven by a specific race among minorities overall. Tenth, we assess whether the association between cohort racial diversity and cohort median starting salary is robust to using mean instead of median salary. The business school data contain the mean starting salary for MBA cohorts (Supplementary Table 8). The parameter for racial diversity is positive and significant (P < 0.10 for narrow cohort racial diversity and P < 0.01 for broad cohort racial diversity). Mean salary data are not provided for JD cohorts, but racial diversity has a higher parameter value for JDs than for MBAs based on median starting salary (202.82 versus 64.71 in Table 2). Eleventh, we assess whether MBA salary data reported from 1 year earlier and JD salary data reported from 1 year later (compared with our primary results in Table 2) could impact our results. Parameters for racial diversity remain positive and significant with these alternative data alignments (P < 0.01 for MBA students and P < 0.10 for JD students; Supplementary Table 9). Twelfth, we assess whether a measure of racial diversity using equal weights across cohorts who are in the programme at the same time as the narrow cohort might impact our results. Parameters for racial diversity remain positive and significant with this alternative measure (P < 0.01 for MBA and JD students; Supplementary Table 10). Thirteenth, the association between cohort racial diversity and cohort median starting salary might also vary systematically by employer types, if certain employers have a need for or derive higher benefits from employing students educated in a diverse cohort. Law school data enable us to examine associations between racial diversity and median starting salary offered by public-sector and private-sector employers (similar data are not available for business schools). Furthermore, starting in 2012, law school data contain percentages of students in a graduating cohort employed by private law firms of various sizes. These data allow us to extend equation (1) to equation (4) (see Methods). Accordingly, we estimate associations between cohort racial diversity and public- or private-sector median starting salary plus interactions between racial diversity and employment percentages of large and small private-sector firms relative to medium-size private-sector law firms (Supplementary Table 11). We observe a positive association in the public sector. In the private sector overall, there is no significant association; instead, results are contingent on firm size. In particular, there is a significant positive interaction between racial diversity and percentage of cohort graduates employed at large law firms but not at small law firms. Previous courts heard that “the educational benefits that flow from a diverse student body” prepare students “for an increasingly diverse workforce and society”, and create “leaders with legitimacy in the eyes of the citizenry” (p. 307, 330 and 332 in ref. 15 ). Such benefits would arguably be critical for public-sector organizations and large private law firms serving a broader customer base. Indeed, many large firms have affirmed that a workforce trained in a racially diverse environment is an “economic imperative…critical to business success” 42 .

Discussion

Our study provides a broad-based, cross-sectional and longitudinal examination of racial diversity’s association with median starting salary for graduating student cohorts in business schools and law schools. Our results consistently demonstrate that racial diversity of an educational cohort is positively associated with cohort median starting salary at graduation. Although we include female composition of the cohort as a control variable in our study, we encourage focused research on gender diversity’s potential benefits along with examining race–gender interactions 43 and racial framing around affirmative action that has hindered recognition of racial diversity’s benefits 44 .

Our focus on the economic benefit of racial diversity to entire cohorts of students is consistent with previous Supreme Court decisions highlighting the educational benefits of diversity, consistent with learning theory: “The atmosphere of ‘speculation, experiment and creation’—so essential to the quality of higher education—is widely believed to be promoted by a diverse student body” (p. 312 in ref. 45 ). Therefore, “student body diversity is a compelling state interest that can justify the use of race in university admissions” (p. 325 in ref. 15 ) because learning benefits derived from racial diversity apply to all students in a diverse cohort, thus preparing them “for success in the corporate world” irrespective of “whether [students] are minorities or not” 42 .

Moreover, previous Supreme Court decisions found that racial diversity can achieve other notable outcomes, including enhancing national security and, more broadly that our “‘nation’s future depends upon leaders trained through wide exposure’ to the ideas and mores of students as diverse as this Nation of many peoples” (p. 313 in ref. 45 ). This rationale may have been what compelled the most recent court to not overturn race-based admissions programmes at US military academies “in light of the potentially distinct interests that military academies may present” (p. 22 in ref. 1 ). In addition, diversity “promotes ‘cross-racial understanding’, helps to break down racial stereotypes, and ‘enables [students] to better understand persons of different races’” (p. 330 in ref. 15 ). Finally, despite being a minority opinion, justices and others have argued that affirmative action programmes are justified “to remedy disadvantages cast on minorities by past racial prejudice” (p. 325 in ref. 45 ). This remediation argument is advanced through a distributional justice lens, which advocates using affirmative action as a remedy for overcoming current inequities attributable to past racial prejudices 46 .

Although our findings are correlational, it is important to note that a cohort’s racial diversity (explanatory variable) is determined at matriculation, whereas the same cohort’s median salary (dependent variable) is determined at graduation, that is, 2 years after matriculation for full-time MBA student cohorts and 3 years after matriculation for full-time JD student cohorts. Indeed, a Granger test on the temporal sequence between incoming cohort racial diversity and graduating cohort median starting salary rejects the null hypothesis that incoming cohort racial diversity does not Granger-cause graduating cohort median starting salary (P < 0.001). By contrast, the null hypothesis that the graduating cohort median starting salary does not Granger-cause the incoming cohort racial diversity cannot be rejected (P > 0.29 for business schools; P > 0.36 for law schools). Furthermore, an Arellano–Bover/Blundell–Bond linear dynamic panel-data generalized method of moments regression reveals that higher racial diversity of an incoming cohort is significantly associated with increases in median starting salary of the graduating cohort over time 47 , 48 , 49 (Supplementary Table 12).

We compile summary statistics of parameter estimates for racial diversity across all 72 of our model specifications (Table 4; also see Supplementary Table 13 for details). Here we find that for the narrow cohort operationalization of racial diversity, the mean salary increase is US$125.24 for MBA graduates and US$297.58 for JD graduates. For the broad cohort operationalization of racial diversity, the mean salary increase is higher: US$164.94 for MBA graduates and US$343.77 for JD graduates. The effect sizes for narrow cohorts mean that in an MBA or JD cohort of 100 students, one additional minority student is associated with an approximate starting salary increase across the entire cohort of US$12,524 or US$29,758 (also see Supplementary Table 8, where we find that increases are larger using cohort mean than cohort median starting salaries). Moreover, such increases in starting salaries are likely to persist and compound over a lifetime of earnings 50 . “For example, assuming a 5 per cent pay increase each year over a 40-year career, a twenty-five-year-old employee that starts at $50 000 would earn $634 198 less than an employee starting at $55 000 by the time s/he reached the age of 65” (p. 371 in ref. 50 ).

Next, we discuss our study’s implications across three realms, assuming that our cumulative analyses are sufficient to infer at least partial causality. However, we are limited in making hard causal claims owing to the observational nature of our data. In Supplementary Information section 4, we discuss additional implications for researchers.

Implications for public policy, universities, other organizations and society

Legislators should recognize the educational and economic value of racial diversity by bypassing the vicissitudes of the judicial system through legislation to legalize affirmative action and affirm and expand civil rights. They should also enshrine the First Amendment rights of university administrators to admit students who best promote learning for all students.

Simultaneously, courts should revisit affirmative action on the basis of our evidence and also because previous courts, business executives, military leaders and university leaders have consistently affirmed its benefits, which only the most recent Supreme Court has rejected. As part of this reconsideration, courts should use our metrics (for example, racial diversity percentage, median starting salary) to scrutinize affirmative action programmes.

University administrators (including undergraduate administrators) should implement programmes to take full advantage of the educational benefits that flow from racial diversity. Even though our data are on two graduate, professional programmes, the learning and salary benefits of racial diversity seem theoretically likely to accrue to students in other programmes. According to learning theory, interactions with diverse individuals, not simply exposure to diverse individuals, are key to deriving learning benefits. These benefits could be achieved through curricular activities (for example, student sections, study groups, projects, teaching assistantships, international study courses), co-curricular activities (for example, academic-based clubs, leadership training, mentorship, case competitions, business plan competitions, centres, guest speakers, workshops, seminars, research assistantships) and extracurricular activities (for example, social clubs, athletic clubs, community service) offered at any educational institution. Moreover, university administrators should continue to admit racially diverse student cohorts, especially since the recent Supreme Court decision allows universities to consider “an applicant’s discussion of how race affected his or her life” (p. 39 in ref. 1 ).

More specifically, professional school administrators should consider filing lawsuits to secure the legal right to reinstate affirmative action. Such a lawsuit could be based in part on this study’s metrics and results. Beyond this, evidence on graduates’ ability to operate successfully in a pluralistic society would reinforce diversity as a compelling state interest. These lawsuits could result in carving out exceptions for MBA and JD programmes (and other cohort-based programmes) similar to the carveout for US military academies. Managers outside universities should also respond to our evidence for positive economic value of racial diversity by sponsoring and promoting additional opportunities for interactions among diverse employees. This could happen through formal work routines (for example, rotating teams, assignments) and properly planned informal activities (for example, social gatherings, sports activities, social events). Managers should also retain and strengthen diversity programmes for employee recruitment, training and retention. More broadly, society should recognize that the Supreme Court decision outlawing affirmative action is likely to lower economic benefits derived from education investments. Aside from the loss of socially enriching experiences and remediation of past racial injustices, economic harm will ensue for students educated in less-diverse cohorts and collectively for society overall.

In conclusion, we hope that our study’s results increase the probability that (1) universities will continue to prioritize racial diversity in their student cohorts by considering how race affected applicants’ lives, as still permitted in the recent Supreme Court decision, (2) a future Supreme Court will revisit the role of affirmative action in higher education and potentially reallow it just as the latest court decision still allows it at US military academies, and (3) organizations will maintain employee diversity programmes by highlighting the economic benefits that accrue from racially diverse groups.

Methods

Data source

We start by providing details on our data sources. We manually compiled data from private (US News Academic Insights) and public (US News print issues) data sources that report student information used in ranking full-time MBA and JD programmes. US News surveys all business and law schools to obtain information on their MBA and JD programmes.

The key advantages of using US News data are their reliance on comprehensive yearly cohort-level information on specific cohort-based programmes, including demographics and salaries at graduation, their reliability, their high response rates, and their consistency over decades. Given the importance of rankings in these publications, few schools can afford to not divulge their information. Moreover, information submitted by schools is not only verified by US News but also vetted by public entities, competing schools and media. Errors are spotlighted and reported to law enforcement for investigation and potential prosecution, leading to a high degree of reliability. For example, see the New York Times story ‘Former Temple U. Dean found guilty of faking data for national rankings’ 51 . To directly examine the reliability of US News as a data source, we benchmark its data with National Science Foundation (NSF) National Survey of College Graduates data for overlapping variables and years on the number of college graduates. Specifically, we compare the number of undergraduates graduating in all US universities obtained by the US News survey during the years 2018–2021. The National Survey of College Graduates conducted by NSF estimates the cumulative number of undergraduates in the USA during the same years. Thus, we can benchmark the US News reported number of undergraduates graduating during the years 2018–2021 with the NSF survey estimated increase in undergraduates during 2018–2021. In Supplementary Table 2, we report these numbers, broken down by gender and majority versus minority races. These show that US News data and NSF survey data are reasonably similar. Moreover, our data do not use any sampling and include all available information from the business schools and law schools responding to the US News surveys during the observation period. In Extended Data Table 4, we report that the proportion of the cohort population in our data is as high as 66–87% and 66–95% for business schools and law schools, respectively. Next we provide details on how we collect and align cohort data.

Data collection

Most of our data are obtained from US News print issues frequently entitled ‘Best Business Schools’ and ‘Best Law Schools’ from public libraries or the researchers’ private collections. Starting in 2004, US News published more comprehensive reports that could be purchased as a separate annual publication devoted to America’s Best Graduate Schools. These annual reports contain yearly information about all US MBA- and JD-granting universities. We obtained all available editions of these annual reports from 2004 to 2019, when they were discontinued. We then merged our data with archival information available from the US News Academic Insights database, which is available to institutional subscribers. More details on the database are available at .

We carefully track the title year (that is, the year mentioned in the title), publication year (that is, the year of issue) and survey year (that is, the year when data are collected) for each piece of information to align all data pertaining to a specific MBA or JD cohort. For example, the Best Business Schools 2020 (title year) and Best Law Schools 2020 (title year) were published in 2019 (publication year) and used survey results from 2018 (survey year) that included the demographic, average or median standardized test score, and average or median undergraduate GPA for the 2018 incoming full-time class (that is, 2018 MBA and JD cohorts) and median cohort salary for 2018 MBA graduates (that is, the 2016 MBA cohort) and 2017 JD graduates (that is, the 2014 JD cohort) (Supplementary Table 3). We then align data by cohort such that all information belongs to the same cohort. Next, we provide a description of our variables and their timelines.

Variables and operationalization

Our focal variables are cohort racial diversity and cohort median starting salary for each MBA and JD cohort. For racial diversity, US News reports the percentage of US domestic minority race students entering MBA and JD programmes at each school during the fall term of the survey year. These minority students identify themselves as African American, Alaska Native, American Indian, Asian American, Native Hawaiian or Pacific Islander, or Hispanic/Latino.

In Supplementary Table 4, we describe our approach to operationalizing cohort diversity based on the percentage of minority students in the MBA and JD cohort entering in a given year. A narrow view of cohort diversity reflects the racial diversity students experience within their own cohort of students. Essentially, this means that 100% of the weight is placed on the current cohort and 0% is placed on classes above and/or below. An alternative broader approach is to weigh the racial diversity of each cohort equally with those cohorts that interact with it during each year in their programme. For example, MBA students in their first year interact with their own cohort plus students in the cohort 1 year above; MBA students in their second year interact with their own cohort plus students in the cohort 1 year below. This means 2 years of interactions with their own cohort plus 1 year each of interactions with cohorts above and below. The approach is the same for JD students across their 3 years in law school. On the basis of our experience and understanding of MBA and JD programmes, MBA students and JD students have more interactions with their own cohort than surrounding cohorts. Therefore, as an initial estimate, we assign 1/2 weight to cohorts 1 year away from the focal cohort and 1/4 weight to cohorts 2 years away from the focal cohorts. Only JD students have cohorts 2 years away from their focal cohort as they are in 3-year programmes while MBA students are in 2-year programmes. As narrow cohort diversity and broad cohort diversity with declining weights most closely align with learning theory based on student interactions, we use these two measures in our primary analyses. We also use equal-weighted broad cohort diversity for robustness checks.

For MBA programmes, US News also reports information on mean starting salary at graduation for each student cohort. For JD programmes, US News reports median starting salary for full-time students separately for public-sector and private-sector employment. To arrive at an overall median salary measure, we use a weighted average of these public- and private-sector median salaries according to the percentage of graduates working in each sector. Before 2014, US News did not report per cent breakdown for full-time JD students’ employment in public and private sectors, so we determined weights on the basis of percentage breakdown for all JD students (full-time and part-time).

Other control variables reflect the gender composition (female student percentage) and student quality of a cohort based on average GMAT and average undergraduate GPA for MBA cohorts; median LSAT and median undergraduate GPA for JD cohorts. In addition, we also collected information on business and law schools’ parent universities’ prestige, student body size and location. For prestige, we use the inverse rank of each school’s parent university as reported in 2024 US News Best National University Rankings. The lowest ranked parent university of a business school or law school in our data is 415. Therefore, we subtracted the US News rank of every university from 416 to arrive at the inverse rank measure. For example, University of Chicago was ranked 12 in the 2024 US News Best National University Rankings, so its inverse rank provides a prestige score of 404 (that is, 416 – 12). For size, we include total student enrolment as reported by the parent university in 2024 US News Best National University Rankings. For location, we include the city population where the parent university is located using 2020 US Census data to reflect its ‘urban’ level. We standardize these variables, so they are on the same scale.

In addition, US News started reporting racial breakdown information for business schools (but not law schools) starting with the 2023 publication (that is, 2022 incoming cohort). We collect information for the proportions of Asian American, Black and Hispanic students for the 2022 incoming cohort.

Overall, our data comprise MBA cohorts in 141 business schools between 1993 and 2021 and JD cohorts in 200 law schools between 1999 and 2019. We have complementary operationalizations of narrow and broad cohort racial diversity for both MBA and JD cohorts, median starting salary for MBA cohorts and weighted median starting salary for JD cohorts. In addition, we have GMAT (for MBA cohorts) or LSAT (for JD cohorts), undergraduate GPA, and percentage female composition for these cohorts We also have measures of university prestige, size, location and racial composition for each school (based on 2022 cohort). Supplementary Table 5 provides a description of our variables and timelines.

Descriptive statistics

We summarize our data in Extended Data Table 1. Notably, we have substantial variation in our two key variables (racial diversity and median starting salary) in both datasets. Unsurprisingly, these variations are larger across schools than within schools. However, the variations within schools are not trivial, as the decision to accept an admission offer is idiosyncratic at the individual student level. Also, in Extended Data Table 2, we find substantial year-to-year increases and decreases in our focal variables, narrow and broad cohort racial diversity, and cohort median starting salary. We report correlations between the variables in Extended Data Table 3. We observe meaningful correlations between starting salary and racial diversity in business schools and law schools.

Models

Our cohort-wide panel data provide cross-sectional and longitudinal variation that enable us to identify racial diversity’s association with graduates’ salaries after accounting for university and year fixed effects. The university and year fixed effects capture multiple factors that could affect both diversity and other outcomes, including university curricula, facilities, instruction quality and year-specific events (for example, affirmative action cases, stock market changes and so on). The use of panel data with university and year fixed effects to capture both inter-university cross-sectional differences and intra-university time-series dynamics leads to high model fit stemming from “several advantages over cross-sectional or time-series data”… [including] “(i) more accurate inference of model parameters, (ii) greater capacity for capturing the complexity of human behavior…, [and] (iii) simplifying computation and statistical inference” (p. 3 in ref. 52 ).

To identify appropriate control variables, we use an empirical approach by regressing racial diversity on potential control variables from our dataset along with university and year fixed effects. Our goal is to identify variables that can explain variation in cohort racial diversity and then include those variables in our model to address potential omitted variable bias 53 . Among the variables we considered, we found that standardized test scores (average GMAT/median LSAT), average/median undergraduate GPA, and female composition of a cohort have statistically significant associations with cohort racial diversity, although school rank does not. The extant literature on returns to education supports the results of our empirical approach 54 . A square root Lasso procedure reveals that racial diversity and the three control variables (GMAT/LSAT, undergraduate GPA and female composition) provide complementary information and that none of these variables needs to be dropped.

Therefore, our main model includes racial diversity, GMAT/LSAT, undergraduate GPA and female composition of each cohort plus university and year fixed effects to account for unobservables in a cross-sectional, time-series, cohort-level median salary model. These fixed effects potentially “remove the endogeneity bias problem as long as the unobserved product characteristics (i.e., business or law schools) do not vary across time” (p. 667 in ref. 53 ).

Our main model has the following specification:

where Salary i , t is the median salary at graduation for the MBA cohort or JD cohort who started the programme in year t at school i; Racial diversity i , t is the percentage of minority students in the respective cohort; X i,t is a vector of control variables including GMAT/LSAT, GPA, and female composition in the cohort; α i and γ t are university and year fixed effects; ε i , t is idiosyncratic error; and β 1–5 are the model parameters.

Examining systematic variation

Model 1 estimates average association between cohort racial diversity and median starting salary. This association may be nuanced and vary systematically across specific characteristics of universities. Previous evidence on racial diversity’s educational benefits is overwhelmingly based on a few prestigious, large, urban universities 6 , 19 , 21 , 22 , 23 . Therefore, we sequentially add an interaction term to our model specification for each business or law school’s parent university’s prestige, size and location.

where the vector Z represents a business or law school’s parent university’s prestige, size or location.

Some have argued that a higher proportion of Asian American students or lower proportion of Black and Hispanic students among a racially diverse student cohort might drive the association between racial diversity and salary 55 . Breakdowns of diverse races in MBA and JD cohorts are not available for our observation period. Fortunately, US News started reporting such information for business schools (but not law schools) starting with the 2023 publication (the 2022 cohort). To examine whether our results are driven by programmes with a higher percentage of Asian American students or a lower percentage of Black or Hispanic students, we include an interaction between diversity and the standardized percentage of Asian American and Black or Hispanic students in 2022. We do not include the main effect, because the percentages of these minority students over time are not available across the decades of our data and, therefore, are captured by university fixed effects. We estimate the following model for business schools:

where the vectors A and B represent a business school’s percentages of Asian American and Black or Hispanic students, respectively, in their 2022 MBA cohort.

The association between cohort racial diversity and median starting salary may also vary systematically by employer types, if specific employers have a higher need to employ students educated in a diverse programme. Indeed, several amici briefs filed by large, reputable US firms in landmark court cases state this need 42 . Examining this possibility requires salary and employment information by employer type. Fortunately, for law schools, US News collects separate employment information for public-sector and private-sector employment and also for private-sector law firms of varying sizes. Therefore, for law school JD cohorts we estimate:

where Salary i , j , t denotes the median starting salary offered by employer type j to JD graduates in cohort i in year t and the vector E represents the percentage employed by different sizes of employer type j in year t.

The association between racial diversity and cohort median starting salary may also vary across different model specifications, each grounded on specific assumptions. These include using fixed effects or random effects of the cross-sections and/or time series, unique school-specific diversity effects, unobserved omitted variables not captured by the fixed effects, nonlinear racial diversity effects, data outliers and dynamic causality. We use a wide variety of model specifications to examine whether model assumptions might affect the association between racial diversity and starting salary.

Ethics statement

The authors commissioned an independent audit of the results that was conducted by a full-time university statistician who used different statistical software from the one that the authors used. This audit began with reconstructing our datasets from raw US News data and validated our empirical analyses. The raw data used in the study are made publicly available to higher-education institution subscribers by US News (see ‘Data availability’ for further details). US News data are not individually identifiable and therefore analysis would not involve human individuals. The institutional review board recognizes that the analysis of de-identified, publicly available data does not constitute human individuals research as defined at 45 CFR 46.102 and that it does not require institutional review board review.

Statistical software

All analyses were conducted and figures were created using Stata, version StataNow/SE 19.5 for Windows (64-bit x86-64) Revision 13 Aug 2025 Copyright 1985–2025 StataCorp LLC.