Evidence from 42.3 million HMDA applications, 2020–2024. Five independent identification strategies — DFL decomposition, within-lender fixed effects, regression discontinuity, difference-in-differences, and partial identification bounds — all reach the same conclusion.
This paper measures how much of the racial gap in mortgage approvals survives once creditworthiness is accounted for, and tests the answer five independent ways. A mortgage is the primary vehicle for wealth accumulation in most American households — who receives credit shapes financial trajectories, neighbourhood composition, and the persistent racial wealth gap. A substantial literature has documented for three decades that Black households face systematically higher denial rates than observably similar White households. But a single regression can always be dismissed as an omitted-variable artifact.
This paper asks three questions: How large is the within-lender racial approval differential — and has it changed over a macroeconomically turbulent five-year period? Does it concentrate at specific institutional rule boundaries? How does it behave under conservative, partial-identification assumptions about unobserved creditworthiness?
The raw Black–White approval gap is 14.95 percentage points. Even after reweighting on every characteristic HMDA observes, 98.6% of this gap remains. The estimated scale: 126,000 fewer annual approvals for Black applicants.
Five independent strategies, five different failure modes if any one were wrong — all point the same direction
The analysis covers a near-universe of mortgage applications filed with HMDA-reporting lenders between January 2020 and December 2024 — 42.3 million records spanning more than 5,500 lending institutions. This five-year window captures extraordinary macroeconomic variation: a pandemic-driven refinancing boom during 2020–2021 (federal funds rate at zero lower bound), followed by the sharpest monetary tightening cycle in four decades (rates rising to 5.5% by 2023, cutting application volume by more than half).
Figure 1 — Sample coverage: 2020–2024, spanning two distinct macro regimes
The 14.95 pp raw gap is not a new finding — but its persistence across a five-year window spanning a pandemic boom and the sharpest credit tightening in 40 years is new. This paper's contribution is documenting the institutional structure of the gap in detail: where it concentrates, how it responds to rule boundaries, and how it behaves when unobservable creditworthiness is bounded.
The headline estimate is stress-tested from five independent directions. Each has different identifying assumptions and different potential failure modes — when they all agree, the result is no longer attributable to any single one.
Within-lender fixed effects reveal that nearly three-quarters of the gap (74.6%) is associated with variation within lending institutions — a share that has grown monotonically from 66.8% in 2020 to 78.3% in 2024. This pattern implies the gap is not primarily a story of Black applicants sorting toward stricter lenders. It is concentrated inside the same institution, often in the same year.
Figure 2 — Share of racial approval gap attributable to within-lender variation, 2020–2024. Growing monotonically from 66.8% to 78.3%.
Above the 80% loan-to-value ratio, private mortgage insurance (PMI) becomes mandatory — a sharp institutional rule that creates a valid regression discontinuity design. Black applicants face an additional 2.0 pp approval differential above this threshold. This effect is concentrated entirely in purchase loans and is entirely absent in refinancings, providing a compelling within-sample test of the PMI mechanism.
Figure 3 — Regression discontinuity at the 80% LTV threshold. Purchase loans (green) show a 2.0 pp jump; refinancings (grey, dashed) show no discontinuity. P-values supported by density tests and exact permutation inference.
State-level approval gaps range from 3.6 to 23.5 percentage points, with regional means spanning 9.3 pp (West) to 16.1 pp (Midwest). County gaps are spatially correlated with HOLC redlining coverage — counties with greater historical D-grade ("redlined") coverage show larger contemporary approval gaps, conditional on observed characteristics.
Figure 4 — Regional mean racial approval gaps. State-level gaps range 3.6–23.5 pp. Larger gaps spatially correlated with historical HOLC redlining coverage.
Critically, the within-lender racial differential does not vary with lender size or market concentration — it is essentially flat across lender-size quartiles and non-monotonic across HHI quartiles. This pattern is more consistent with underwriting practices that are relatively uniform across market structures than with individual taste discrimination, which Becker's (1957) competitive-discipline model predicts competitive markets should eliminate.
The most credible existing estimate — Bhutta, Hizmo, and Ringo (2025) in the Journal of Finance — conditions on actual FICO scores, DTI, liquid assets, and automated underwriting recommendations and finds residual gaps of 1–2 pp. Our estimates using public HMDA data are larger, and we address this directly through partial-identification bounds.
Even when FICO scores, DTI ratios, liquid asset reserves, and employment stability are simultaneously bounded at their maximum plausible values calibrated to SCF 2022 data, the analysis finds that a gap of at least 6.37 to 7.97 percentage points — 44 to 55% of the observed gap — is not captured by these variables in this specification.
Full title: Persistent Racial Disparities in U.S. Mortgage Approval: Evidence from 42 Million Applications, 2020–2024
Venue: Submitted to Journal of Housing Economics, March 2026. JEL codes: G21, J15, G28, R21, R31.
Author: Rajveer Singh Pall (submitted anonymously via journal system)
Data: HMDA public use files (FFIEC), 2020–2024; Survey of Consumer Finances 2022 (partial ID bounds calibration); HOLC Mapping Inequality database (geographic analysis).