Hiring and Labor Market Discrimination

What the evidence says

The best-replicated result in this literature is the resume-audit callback gap by race. Bertrand and Mullainathan's correspondence study found resumes with White-sounding names received callbacks 50 percent more often than identical resumes with African American-sounding names (10.08 percent vs. 6.70 percent, a 3.35-percentage-point gap), an advantage roughly equivalent to eight additional years of experience, and uniform across the occupation and industry categories tested. Quillian's meta-analysis of decades of correspondence studies puts the modern estimate at White applicants receiving 36 percent more callbacks than equally qualified Black applicants (95% CI 25–47%) and 24 percent more than Latino applicants (95% CI 15–33%), and finds no change in the level of discrimination against Black applicants over roughly 25 years, with only a statistically insignificant (p = 0.099) hint of decline against Latino applicants. Where a racial gap has narrowed, it has narrowed unevenly: the Black–white wage gap fell from 30–42 percentage points in 1940 to 12–15 percent by 1980–1990, with the sharpest decline concentrated in 1965–1975, immediately after the Civil Rights Act — but it did not close, and Black men were still measured losing 12–15 percent of earnings to discrimination in 1980 and 1990. Discrimination is not confined to the callback stage: when employers do respond to Black candidates, the jobs offered carry a listed salary $3,071 lower on the low end and are 56.1 percent as likely to be a high-prestige (analyst/managerial) role; an elite-university credential raises callbacks for everyone (84 percent higher), but a Black candidate with an elite credential only matches the callback rate of a White candidate from a less-selective school.

Gender discrimination in hiring does not run in one consistent direction — it flips with the sex-composition of the occupation. Matched-pair correspondence tests found women rejected for interview 46 percent of the time in engineering (male-dominated) versus 23 percent for men, while men were rejected 59 percent of the time for secretarial roles (female-dominated) versus 16 percent for women, and similarly in computer-analyst programming (47 percent vs. 12 percent) and chartered accountancy (60 percent vs. 17 percent). Motherhood carries an independent penalty distinct from sex-typing: mothers received hiring recommendations 47 percent of the time versus 84 percent for otherwise-identical non-mothers, a 37-point gap, while fathers saw a smaller 11-point premium (73 percent vs. 62 percent). That gender identity itself, not just qualifications, drives some of this is supported by the finding that blind orchestra auditions increased the probability a woman advanced past the preliminary round by 50 percent. These hiring-stage findings sit alongside persistent aggregate gaps: women's labor force participation rate was 12.5 percentage points lower in 2013, the annual earnings gap was about 22 percent (18 percent weekly), and women's full-time earnings rose from 59 percent to 71 percent of men's between 1981 and 1995.

Beyond race and gender, several other traits produce callback penalties of comparable size, per a meta-analysis of correspondence studies: disability status reduces positive callback response by about 41 percent, older age by about 31 percent, and lower physical attractiveness by about 37 percent — magnitudes the same source describes as "at least equally problematic" as the 29 percent reduction found for ethnic minority status. Ethnicity discrimination outside the Black/white binary is documented but with wide variance in precision: in a Turkey-based correspondence test, Turkish applicants received employer responses 3.8 times as often as identically qualified Kurdish applicants (6.5 percent vs. 1.7 percent), with the gap widening at the request-for-information stage (10x) and narrowing sharply by the interview-invitation stage (1.17x, based on very few cases). Criminal-record penalties compound with race: audit studies find ex-convicts face reduced callbacks generally, with a larger reduction for Black than white applicants, and Black ex-inmates show slower post-release wage growth than white ex-inmates. Credential type also matters independent of a candidate's demographic traits: a for-profit online bachelor's degree reduced callbacks 22 percent versus a public non-selective degree, and a for-profit health certificate reduced callbacks 57 percent versus a public one.

Two mechanisms recur across this literature as structuring forces behind who gets hired. First, informal referral networks: living on the same residential block as another worker increases the probability of working together by over 33 percent versus a nearby block, and referral availability alone raises co-employment probability by 6.9 percentage points, with referrals significantly more likely between demographically similar pairs. Second, employer-side "cultural fit": in one study of professional-services hiring, more than half of evaluators ranked fit as their single most important interview criterion, ahead of analytical thinking and communication, and firms formally mandated fit assessment in hiring policy; the same study found that similarity to the evaluator produced more favorable competence ratings independent of the actual quality or quantity of information available. A separate high-quality data domain — professional team sports, where performance and pay are both precisely measured — shows racial salary discrimination declining to statistical insignificance overall by the mid-1990s in the NBA (from an 11–25 percent Black salary shortfall in the mid-1980s), with no significant premium ever found in MLB and only a small, mostly insignificant 1–4 percent premium in the NFL; but positional segregation (Black players overrepresented in some positions, underrepresented in others) and higher career-exit rates for Black players persisted.

A methodological fault line runs under much of this evidence. Heckman's review of the audit method notes that in Chicago and Denver, 86 percent of matched hiring-pair tests showed symmetric treatment (both testers hired or both rejected), and argues audit-method validity "is critically dependent on unverified assumptions" about the equality of unobserved productivity across race/gender groups, such that audits "can find discrimination when none exists and can disguise it when it does." This critique does not overturn the large, statistically precise callback gaps reported elsewhere in this evidence base, but it is a direct challenge to how those gaps should be interpreted, and this page carries both without resolving them. Separately, a body of self-report survey evidence (ethnic minority status associated with an odds ratio of 5.72 for reporting personal discrimination) measures perceived discrimination rather than an audited hiring outcome, and should not be read as equivalent evidence to the correspondence-study numbers above.

Where the evidence splits

Deming's large for-profit-credential correspondence experiment explicitly states it found "no consistent evidence of lower callback rates for racial minorities," contrasting this directly with Bertrand and Mullainathan's finding of a large, robust racial callback gap. The claims here do not state a design difference that would explain the discrepancy — Deming's applicant pool was built around credential-type manipulation (business and health-support job postings) rather than the general labor market Bertrand and Mullainathan sampled, but the source does not itself attribute the null race result to that difference, so the split is unexplained by the sources on this page.

On gender specifically, Bertrand and Mullainathan (per Azmat's synthesis) detected no evidence of lower callback rates for women, while Riach's matched-pair studies and Azmat's own occupation-based correspondence test found large, statistically significant, occupation-dependent gender gaps. Here the sources do offer an explanation: Azmat notes gender discrimination was not the primary focus of the Bertrand and Mullainathan study (race was), and separately states that "compared to the regression approach, the experimental approach tends to find far more limited evidence of discrimination against women in the marketplace" — identifying a difference in study focus and identification strategy (regression-based vs. experimental estimates) as the likely source of the disagreement.

On age, a German study found selection rates declining significantly with applicant age and older workers rated lower on learning capacity and flexibility, while a Norwegian study of similar design found managers did not rate older workers lower on any capability, with point estimates similar across age groups. The claims attribute this split to country: the same review states hiring discrimination against older applicants is "more prominent in Europe than in the United States" generally, though it does not explain why Germany and Norway themselves differ.

Thresholds and dose-response

In the NBA, a significant white salary premium of about 20 percent appears specifically at the 75th percentile of the salary distribution and above, holding performance constant; no significant premium was found at the 10th, 25th, or 50th percentiles.

For for-profit health certificates, the 57 percent callback penalty applies only when the job posting does not explicitly require a postsecondary certificate; when the position also requires a valid occupational license, the penalty disappears and no significant difference by institution type is found.

Gender discrimination in restaurant hiring scales with restaurant price tier: in low-priced restaurants, men received job offers women did not 29 percent of the time; in high-priced restaurants, that gap widened to 43 percent of tests, with women showing roughly a 40 percent lower probability of being interviewed and a 50 percent lower probability of receiving an offer.

Among Black males, skin tone sets a gradient within the racial category itself: dark-skinned Black males had 52 percent lower odds of working than light-skinned Black males from the same neighborhoods, after controlling for education, age, and criminal record.

Open empirical questions

• Whether race-blind (anonymized) resume or interview screening reduces the racial callback gap the way blind auditions reduced the gender gap in orchestras — no equivalent race-blind test appears in these sources.
• What magnitude of bias current algorithmic/AI hiring tools produce relative to human recruiters — the review literature flags this as a top concern but reports almost no empirically tested effect size.
• Why greater female presence on a hiring committee was associated with a lower probability of hiring women in the Spanish judiciary case — the mechanism is not examined in these sources.
• Whether specific policy interventions (enforcement, quotas, structured-interview mandates, diversity training) causally reduce measured hiring discrimination at scale — the disability-attitude findings here are qualitative and not tied to a quantified hiring-outcome effect.