Substantial gaps across major race groups
U.S. adults ages 25–54, 2024
2024 ACS PUMS, weighted estimates and 95% intervals. Includes institutionalized adults and armed forces.
About 75 seconds. Start with levels, rather than an explanation. Each dot is a weighted employment-to-population ratio. Current employment includes people at work or temporarily absent with a job, including the armed forces. The sample includes institutional group quarters, including prisons. This is deliberate: the question is opportunity among all adults 25–54, not only those living outside institutions. The conventional civilian noninstitutional comparison is in the backup. In that sample the raw Black–White male employment gap is 9.9 percentage points, versus 13.4 here. It remains 8.5 points after the main controls, versus 11.4 here.
The gender comparison is important. Black men have markedly lower employment than White men; the corresponding difference among women is much smaller. Hispanic men’s employment is close to White men’s, while Hispanic women’s employment is lower. Asian and White adults do not uniformly have the same ordering. Avoid implying that every non-White group has the same experience.
Hispanic is an ethnicity, not a race. Hispanic adults of any race form one group; White, Black and Asian here mean non-Hispanic adults reporting those single-race categories. The four displayed groups do not exhaust the population. Other categories remain in the analysis and are shown in the backup. Men/women refer to the recorded ACS sex variable. Source: https://www2.census.gov/programs-surveys/acs/data/pums/2024/1-Year/csv_pus.zip and the 2024 PUMS data dictionary. These are 1-year estimates, not pooled 5-year estimates. 2025 microdata were not yet released when this draft was prepared on October 30, 2026.
Substantial gaps across major race groups
U.S. adults ages 25–54 with positive labor earnings, 2024
2024 ACS PUMS. Survey-weighted means, in 2024 dollars.
About 60 seconds. These are ordinary arithmetic means of annual earnings, weighted using ACS person weights. The earnings reference period is the past twelve months. The sample consists of people with positive total labor earnings who report working during that period. Annual earnings combine hourly pay, weeks and hours of work. People with nonpositive total labor earnings remain in the employment sample but are excluded from this earnings sample.
Black men’s mean earnings are about $60,719 versus $95,626 for White men, a 36.5% difference. Black women’s mean earnings are about $52,706 versus $67,092 for White women, a 21.4% difference. These descriptive means differ from the historical quantile earnings gaps and the skills study’s hourly-wage regressions that follow. The backup includes full-time/full-year and earnings-per-reported-hour sensitivity results. Annual earnings differences are not entirely wage-rate differences.
Source: 2024 ACS 1-year PUMS. PERNP multiplied by ADJINC/1,000,000. Person weights PWGTP. The analysis script and full group estimates accompany the deck.
The same gaps were much larger historically
Black–White earnings gaps, 1940–1980
Bayer, Charles & Derenoncourt (2025), Handbook of Labor Economics , Figure 4.
About 60 seconds. The Black–White earnings gap narrowed substantially in the middle of the twentieth century. This is Figure 4 from the 2025 Handbook chapter, reproduced directly. The two panels show men and working women; red is the median, blue the 90th percentile. Moving toward zero means a smaller gap. The major convergence occurred before 1980. The gains at the median were especially pronounced among working women.
The Census sample is U.S.-born, non-Hispanic Black and White adults ages 25–54. The gaps are conditional quantile differences in log annual earnings, controlling for six five-year age bins. The men’s sample includes nonworkers; the women’s sample includes working wage/salary earners, as described in Appendix B. The axis reproduces the authors’ log-unit scale, not percentage differences or dollars. The figures are not spliced to our contemporary ACS means. The sample restriction for women matters because labor-force participation changed substantially.
The chapter’s later figures document much less progress and reversals after 1980, especially for men. Do not extrapolate these historical declines through the present. This slide establishes the substantial earlier convergence that motivated the debate over skills and discrimination.
Source: Patrick Bayer, Kerwin Kofi Charles and Ellora Derenoncourt (2025), Racial inequality in the labor market, Handbook of Labor Economics, volume 6, pp. 159–228, Figure 4, p. 176. https://doi.org/10.1016/bs.heslab.2025.08.001. Author PDF: https://sites.duke.edu/patrickbayer/files/2025/01/Bayer_Charles_Derenoncourt_Handbook_of_Labor.pdf. Extracted without redrawing or estimating plotted values by scripts/extract_talk_figures.py.
A long decline in explicit prejudice
Three measures of expressed racial inclusion
Marriage and president: Gallup, U.S. adults, 1958–2021 and 1958–2019.
About 60 seconds. The decline in explicit exclusion is visible in several domains. The panels all show the positively coded response: approval of interracial marriage; willingness to vote for a qualified Black presidential nominee of one’s own party; and the view that Black and White children should attend the same schools. The first two series concern all U.S. adults; the school-integration series concerns White adults. Keep the populations distinct when speaking.
Gallup marriage approval rose from 4% in 1958 to 94% in 2021. For the presidential nominee, the last observation in 1958 is 38% and the 2019 reading is 96%. NORC’s historical compilation reports support for integrated schools rising from 30% in 1942 to 90% in 1984. Its series ends in 1984; no later trend is implied. The graph retains all reported NORC surveys, including multiple readings in some years. The Gallup graphs retain the latest survey in each year. These are expressed attitudes, not a measure of employer behavior.
Question wording changes: early marriage questions refer to White/non-White marriage or use historical racial terminology; later questions ask about Black–White marriage. Earlier presidential questions ask about a well-qualified man and use historical terminology. School questions ask about the same versus separate schools, not support for a specific desegregation policy. Use the reported approval percentages rather than one minus opposition because the polls allow no opinion.
Sources: https://news.gallup.com/file/poll/354641/210910InterracialMarriage.pdf; https://news.gallup.com/file/poll/257495/190509PresidentialCandidates.pdf, section C; Tom W. Smith and Paul B. Sheatsley, American Attitudes toward Race Relations, Table 1, https://gss.norc.org/content/dam/gss/get-documentation/pdf/reports/social-change-reports/SC21%20American%20Attitudes%20towards%20Race%20Relations.pdf. Transcriptions and plotting code: research/interracial_marriage_gallup.csv, research/gallup_black_president.csv, research/norc_school_integration.csv, scripts/plot_explicit_attitudes.R.
Large share of gaps accounted for by measures of skill
Share of young men’s Black–White wage gap accounted for
Neal & Johnson (1996), Table 1. NLSY79; hourly wages in 1990–1991, ages 26–29.
About 75 seconds. Years in school capture only part of what people learn. In the same sample of young men, years of schooling account for about one-fifth of the Black–White log wage gap. Using achievement measured in adolescence accounts for about seven-tenths. This is a concrete example of richer measures of pre-market skills explaining substantially more than schooling alone.
The graphic uses Neal and Johnson (1996), Table 1, columns 1–3. The Black coefficient is −0.244 with age alone, −0.196 with age and years of schooling, and −0.072 with age, AFQT and AFQT squared. The plotted shares are 100 × (1 − adjusted coefficient / age-only coefficient): 19.67% and 70.49%. They are shares of a log wage gap, not percentages of variance explained or estimates of the effect of an education intervention. The schooling and test-score models are alternatives, not cumulative controls. The paper’s young workers were ages 26–29 when wages were observed; the sample has 1,593 men. AFQT was measured at ages 18 or younger and combines verbal and quantitative achievement tests.
For women in the same table, coefficients are −0.185, −0.155 and +0.035, so the corresponding shares are about 16% and more than 100%; the last coefficient is not significantly different from zero. Employment selection complicates that comparison. Fryer’s Handbook chapter extends the analysis to older ages and a later cohort. The portion accounted for varies across samples and outcomes, and adding controls need not monotonically shrink a gap. This is evidence of the importance of measured preparation, not a claim that remaining gaps identify discrimination or that every measure is an innate trait.
Source: Neal and Johnson, The Role of Premarket Factors in Black–White Wage Differences, Journal of Political Economy 104(5), 869–895 (1996), Table 1, p. 875, https://bpb-us-w2.wpmucdn.com/voices.uchicago.edu/dist/6/3381/files/2022/01/JPE1996.pdf. Coefficients and standard errors transcribed to research/neal_johnson_table1.csv. The ACS covariate results remain available in the separate backup deck.
The declining significance of discrimination?
“Some scholars argue that, whereas discrimination was once perhaps the dominant factor affecting Black individuals’ labor market outcomes, its impact relative to that of racial skill differences has waned dramatically over the past fifty-plus years and may even have faded into insignificance today.
By this view, the sharp closing in Black-White labor market gaps and the persistent disparities that remain are due principally to the dramatic but incomplete closing of the difference in skill between Black and White workers (Neal and Johnson, 1996; Loury, 1998; Fryer Jr, 2011).”
Bayer, Charles & Derenoncourt (2025)Handbook of Labor Economics , p. 164
About 45 seconds. Begin with the first paragraph; advance once to reveal the second paragraph, beginning “By this view.” This quotation from the most recent Handbook chapter concisely states the view that motivates the next part of the talk. It describes a position in the literature; it is not Bayer, Charles and Derenoncourt’s endorsement of the claim that discrimination is insignificant. Preserve “Some scholars argue” and “By this view” when introducing it.
The chapter cites Neal and Johnson (1996), Loury (1998), and Fryer Jr (2011). Fryer’s chapter is titled Racial Inequality in the 21st Century: The Declining Significance of Discrimination. William Julius Wilson’s earlier The Declining Significance of Race (1978) emphasizes class and changing economic institutions; it is related background, not a claim that discrimination literally disappeared.
Source: Bayer, Charles and Derenoncourt (2025), Racial inequality in the labor market, Handbook of Labor Economics, volume 6, p. 164, https://doi.org/10.1016/bs.heslab.2025.08.001. The full passage was supplied by Evan and verified against the published chapter. Fryer: https://doi.org/10.1016/S0169-7218(11)02408-7.
But there has been less progress since 1980
Black–White earnings gaps, 1980–2022
Bayer, Charles & Derenoncourt (2025), Handbook of Labor Economics , Figures 9a and 10a.
About 45 seconds. These are the corresponding earnings level-gap panels for men and women in the later period. As on the preceding slide, red is the median and blue is the 90th percentile. Among men, the median gap widened through 2010 and then recovered, leaving it almost unchanged from 1980 to 2022. The 90th-percentile gap remained wider. Among women, both gaps shifted toward favoring White women and were about 20 log points in 2022.
The panels reproduce Figures 9a and 10a, retaining their plotted marks, original axes and legends. Only the original “(a) Level Gaps” panel headings are replaced with “(a) Men” and “(b) Women” to match the preceding slide. The authors use different y-axis ranges in these two panels. The source’s log-unit scale is preserved. The series end in 2022.
The post-1980 samples include nonworkers and both U.S.-born and immigrant non-Hispanic Black and White adults ages 25–54. The preceding historical figure covers U.S.-born men and working women, so the charts are kept separate rather than spliced into continuous series. In particular, the apparent difference in women’s gaps at 1980 across the slides reflects the differing samples. The chapter emphasizes changing selection into employment for women, especially increased employment among White women with high earnings potential. The earnings quantiles here also differ from the opening ACS arithmetic means among positive earners.
For men, Table 3 reports a −0.018 change in the median Black–White log earnings gap and −0.096 at the 90th percentile over 1980–2022. These are 1.8 and 9.6 log points of widening, respectively. The chapter documents modest improvements in Black men’s relative earnings ranks that were offset by changes in the overall earnings distribution. Stagnation in earnings level gaps does not mean no change in relative position.
Source: Bayer, Charles and Derenoncourt (2025), Racial inequality in the labor market, Handbook of Labor Economics, volume 6, Figure 9, p. 193, Figure 10, p. 194, and discussion pp. 192–198. https://doi.org/10.1016/bs.heslab.2025.08.001. Panels extracted and arranged by scripts/extract_talk_figures.py without reconstructing plotted values.
Experimental evidence of contemporary discrimination
Kline, Rose & Walters (2022). Example résumés from the supplied research presentation, p. 8.
About 60–75 seconds. Introduce the technology before the firm-level findings. We build fictitious résumés and submit applications to real vacancies. Names convey different racial signals. Names and qualifications are randomly assigned, so the groups of applicants are comparable by design. These two examples are not identical documents with only the names changed: résumé features and qualifications also vary independently by design.
The platform makes a correspondence experiment possible at scale: automated sampling of vacancies, construction and submission of applications, and recording employer responses. The published study submitted 83,643 applications to entry-level jobs at 108 large employers, generally eight applications per vacancy, four with distinctively Black names and four with distinctively White names. The outcome is employer contact, not an eventual job offer. Randomization lets us compare how employers respond to otherwise comparable groups of applicants.
Source: Kline, Rose and Walters (2022), Systemic Discrimination Among Large U.S. Employers, QJE, https://ekrose.github.io/files/randres.pdf. Résumé examples reproduced from supplied Operation_Characterize (4).pdf, PDF page 8, using scripts/extract_talk_figures.py.
White names receive more callbacks
Kline, Rose & Walters (2022). Main-effects figure from the supplied research presentation, p. 9.
About 45 seconds. Start with the average effects before looking across firms. The original six-bar graph compares callback rates by assigned racial signal, gender and age. White names receive callbacks at a rate of 25.1%, versus 23.0% for Black names, a difference of about 2.1 percentage points. Average gender and age gaps are much smaller.
The brackets reproduce the original estimates: Black minus White −0.0205 (standard error 0.0017); female minus male 0.00064 (0.003); over 40 minus under 40 −0.0059 (0.003). The plotted bar heights are rounded group means, so subtracting their displayed labels need not exactly reproduce the bracket estimates. These are callbacks within 30 days. The next slide shows how the racial gap differs across employers.
Source: supplied Operation_Characterize (4).pdf, PDF page 9. Original chart reproduced intact, with its title removed, by scripts/extract_talk_figures.py. Study: Kline, Rose and Walters (2022), Systemic Discrimination Among Large U.S. Employers, QJE, https://ekrose.github.io/files/randres.pdf.
23 firms with significant black-white gaps reliably detected
83,000+ applications · 108 employers · Randomly assigned racial signals
Kline, Rose & Walters (2022), QJE, Table IX. Hiring observed in 2019–2021. Bars: posterior means; whiskers: 5th–95th percentiles.
About 60 seconds. With the design established, turn to differences across employers. Random assignment lets us test differential treatment in hiring directly. More than 83,000 applications provide enough information to detect patterns within individual employers.
The graphic uses the published Table IX data, ordered by posterior mean gap, in the style of the supplied presentation’s “23 firms” slide. Color identifies federal contractors. The firm at the top has an empirical Bayes posterior mean White–Black contact gap of 8.35 percentage points. The horizontal whiskers are the posterior fifth and ninety-fifth percentiles, a central 90% interval; they are not 95% confidence intervals.
The 23 firms have q-values below .05. This controls the estimated false discovery rate for the selected collection; it is not automatically a statement of 95% posterior probability for every individual company. The paper estimates at most about one false discovery among the 23 under the benchmark procedure. Federal contractor status is descriptive and is not randomly assigned. The experiment measures employer contact, not job offers, wages, or the share of all national inequality caused by discrimination. The 2024 report-card study builds on the same experiment.
Source: https://ekrose.github.io/files/randres.pdf, Table IX and surrounding discussion. Visual reference: supplied Operation_Characterize (4).pdf, PDF page 14. Published table values take precedence over the presentation’s older ordering and estimates.
These patterns are systemic
If hiring were race-neutral:
Less than
chance of evidence
Even rarer than a fatal
Map and reported p < 10⁻⁸: supplied research presentation, pp. 15–16. Lightning benchmark: about 1 in 15 million per year, U.S. average.
About 75 seconds. Move from the across-firm picture to a pattern across jobs within one federal contractor. The supplied map reports 38 jobs with gaps favoring White applicants and seven with gaps favoring Black applicants. The point is that the evidence is not confined to a single isolated location.
The supplied presentation reports p < 10^-8 for its no-discrimination test. Under that null and the test’s assumptions, evidence at least this extreme has probability less than one in 100 million. Preserve the memorable lightning comparison while keeping the conditioning correct: this is smaller than the average annual U.S. lightning-death risk, approximately one in 15 million. It is not a posterior probability of no discrimination. The map’s reported test is also distinct from the firm-level Wald tests in Table IX; do not assign its p-value to the table’s first row or infer it solely from the 38-versus-seven sign count.
The source map is reproduced intact from PDF page 15; page 16 supplies the original lightning analogy. Its underlying image is only 700×500 pixels, so it will be less crisp than the newly drawn Table IX graphic. Exact geographic coordinates are not reconstructed from the image.
Lightning calculation: NOAA’s published 2013–2022 average of 22 deaths divided by its 332 million population benchmark equals about 1 in 15.1 million annually. This is a population average, not a personalized risk. Source: https://www.noaa.gov/jetstream/lightning/frequently-asked-questions. Map: supplied Operation_Characterize (4).pdf, pages 15–16.
Attitudes on remedies for inequality are changing
White people benefit fromAt least a fair amount · Registered voters · Pew
New laws are needed to reduceU.S. adults · Gallup
Pew Research Center (June 2024); Gallup (August 2025). Separate questions and populations.
About 75 seconds. The defensible point is declining recognition of a specific structural advantage and declining support for a specific remedy relative to 2020. Pew reports 59%, 55%, and 51% of registered voters in 2020, 2022 and 2024 recognizing White societal advantages at least a fair amount. Gallup reports 61% in 2020 and 46% in 2025 saying new civil-rights laws are needed to reduce discrimination against Black people.
These series are not a single scale or matched panel. The first is registered voters; the second is all U.S. adults. Gallup’s observations are endpoints; the straight line does not establish a smooth annual decline. “New laws” is narrower than all remedial action, and opinions about new laws do not directly measure support for enforcing existing law. Do not infer that recognition of every form of racism is falling. Gallup’s broader question about widespread racism against Black people remained at 64% in 2021 and 2025, and perceptions of workplace unfairness were at a record or near-record high. Nor does this slide prove that discriminatory employer behavior has increased.
Sources: https://www.pewresearch.org/politics/2024/06/06/racial-attitudes-and-the-2024-election/ and https://news.gallup.com/poll/693893/steady-say-racism-against-black-people-widespread.aspx.
The pre- and post-market dichotomy is misleading
Post-market
Less discrimination
A reinforcing cycle
Less labor-market discrimination
Smaller skill gaps less segregation
About 90 seconds. We can address pre-market and post-market gaps at the same time. Pre-market refers to conditions and investments before labor-market entry; post-market here means treatment after entry, including hiring, pay and advancement. School finance and school quality can improve preparation. Reducing discrimination can increase the return to that preparation and the payoff to earlier-life interventions.
The cycle builds over four clicks. Start with less labor-market discrimination. Click 1 adds higher returns to skills; click 2 adds more investment; click 3 adds smaller skill gaps and less segregation; click 4 closes the loop. This is the complementarity: fairer access raises expected returns to acquiring skills, encouraging investment; smaller skill gaps can then weaken self-confirming statistical discrimination in some models. This is a potential mechanism, not an automatic result or an ACS causal estimate. Schooling alone need not eliminate taste-based discrimination, and some forms of discrimination can increase education as a signal (Lang and Manove, 2011).
Related evidence: Jensen (2012) randomized recruiting access for young women in rural India and found increased schooling or training, providing evidence that expected labor opportunities can affect investment. This is not a direct estimate of the cycle for U.S. racial gaps. Kline, Rose and Walters (2026) document survey respondents reporting that fear of discrimination deterred applications; the backup includes that figure. Use this evidence as motivation for pursuing both policy fronts together.
Sources: Jensen, QJE 127(2), 753–792, https://doi.org/10.1093/qje/qjs002; Lang & Manove, AER 101(4), 1467–1496, https://doi.org/10.1257/aer.101.4.1467; https://ekrose.github.io/files/discrim_beliefs.pdf.
Reducing racial inequality
Early-life
Schools
Employer
Hiring
Labor-market
Wage-setting
Bayer, Charles & Derenoncourt (2025); Jackson, Johnson & Persico (2016); Kline, Rose & Walters (2022, 2024).Evidence and methods
About 60 seconds. Close with the broader agenda. Racial inequality reflects the conditions in which people grow up, how employers allocate opportunities, and the institutions that shape the distribution of work and pay. These are complementary areas for policy. School finance and school reform address access to preparation; discrimination enforcement and better hiring systems address access to opportunities and returns. Wage-setting institutions and worker voice affect the overall distribution of earnings, with consequences for racial gaps even without explicitly race-targeted rules. The point is to work across these areas, with evidence on which interventions improve outcomes.
Jackson, Johnson and Persico (2016), The Effects of School Spending on Educational and Economic Outcomes, uses variation from school-finance reforms and finds that greater spending improved later outcomes, especially for children from low-income families. The published version reports that a 10% increase in per-pupil spending throughout all twelve years increased wages by about 7.7%; the earlier NBER abstract reports 7.25%. Use the published version if giving a number. This historical study does not imply that every marginal dollar in every setting has the same effect.
The audit results document the need and potential targets for enforcement; do not present their firm correlations as causal effects of enforcement or centralized hiring. Make the concluding connection to the future of work: new technologies and hiring practices should expand the ability to recognize and use talent, while retaining accountability for disparate treatment. No independent quantitative claim about AI is made in these slides.
Broad context: Bayer, Charles and Derenoncourt (2025), https://doi.org/10.1016/bs.heslab.2025.08.001, including its discussion of race-specific factors and the overall earnings distribution. Source for school spending: https://cpb-us-e1.wpmucdn.com/sites.northwestern.edu/dist/b/3664/files/2020/10/QJE_resubmit_final_version_stamped.pdf. Published QJE 131(1), 157–218 (2016). Audits: https://ekrose.github.io/files/randres.pdf and https://ekrose.github.io/files/report_cards.pdf.
