Philadelphia Mortgage Data Reveal a Stubborn Racial Denial Gap That Credit Scores Cannot Explain

More than five decades after the United States Congress outlawed discrimination in credit transactions, the numbers coming out of Philadelphia tell a story that refuses to change. A new analysis of more than 90,000 mortgage applications filed in the Philadelphia metropolitan statistical area between 2017 and 2022 finds that Black applicants faced odds of denial roughly two and a half times those of similarly situated White applicants, year after year, across a period that spanned a pandemic, a refinancing boom, and a tightening credit cycle. The study, published in the open-access journal Heliyon, goes further than most previous work by asking a question that has rarely been posed in fair lending research: how strong would an unmeasured factor have to be to wipe out the observed racial gap entirely?

The researchers, Maia Berkane, Diego Alvarez and Nico Tanzi, drew on the public Home Mortgage Disclosure Act database, a federal repository that lenders must populate with details on every application they receive, including the applicant’s race, ethnicity, income, loan amount and neighborhood characteristics. They restricted their sample to the most common borrowing profile: conventional, first-lien loans on one-to-four family homes intended for owner occupancy. After filtering out records with missing covariates, the dataset ranged from roughly 12,700 usable applications in 2017 to more than 18,500 in 2021, a scale that gives the statistical tests considerable power.

The methodological centerpiece is a logistic regression of loan denial on indicators for Black, Hispanic and Asian applicants, adjusted for every relevant variable the public data offers: loan amount, loan-to-income ratio, loan-to-value ratio, income, the minority share of the census tract, the tract’s income relative to the metropolitan area, the age of housing stock and the number of owner-occupied units. Because the public HMDA files omit the FICO credit score and the debt-to-income ratio, both of which lenders weigh heavily and both of which correlate with race, the authors treat these as unmeasured confounders and confront them head-on with a sensitivity analysis borrowed from epidemiology.

That tool, the E-value developed by biostatisticians Peng Ding and Tyler VanderWeele, quantifies the minimum strength of association, on the relative risk scale, that an unmeasured confounder would need with both the exposure and the outcome to fully explain away an observed association. In this setting, the exposure is membership in a minority group and the outcome is mortgage denial. For Black applicants, the E-value stayed above 4.0 in every single year of the study, peaking at 4.78 in 2017 and still standing at 4.45 in 2022. In plain terms, any hidden factor capable of dissolving the disparity would need to be linked to both being Black and being denied a loan by a relative risk of roughly fivefold each, over and above everything already controlled for.

The raw regression results are striking in their stability. The odds ratio for Black applicants ranged from 2.30 in 2021 to 2.67 in 2017, with confidence intervals whose upper bounds hovered near or above three throughout. Translated into more interpretable conditional marginal effects, the model implies that if every White applicant in the 2017 data had instead been Black, the denial probability would have jumped from about 8 percent to about 15.4 percent. By 2022 the corresponding counterfactual shift was from 5.8 percent to 11.0 percent. The average marginal effect for Black applicants in 2022 was 6.6 percentage points, a figure that lands almost exactly on the 7 percent Black-White denial gap reported by Giacoletti and colleagues using national data from 1998 to 2018, despite the difference in era and geography.

The contrast with other minority groups is instructive. Hispanic applicants saw their marginal effect fall from 7.4 percent in 2017 to a statistically insignificant 1.7 percent in 2020 before rebounding to 4 percent in 2022, while Asian applicants peaked near 8.7 percent in 2019 and declined to 2.9 percent by the end of the period. Their E-values, generally between 2.0 and 3.0, indicate gaps that are more vulnerable to explanation by unmeasured factors. The authors are careful to note that this does not mean discrimination against Hispanic or Asian applicants is absent; rather, the evidence for a causal interpretation is weaker, and available confounders may account for more of the observed difference.

The most compelling part of the paper is its attempt to test whether the FICO score, the most obvious candidate for a missing confounder, could plausibly do the explanatory work required. Using published data showing mean FICO scores of 677 for Black Americans and 734 for White Americans in 2021, the authors reconstructed the implied score distributions and calculated that the relative risk of having a score below 677 is about 2.4 for Black versus White Americans. That falls well short of the minimum of 3.40 needed, in combination with an assumed fivefold effect of low scores on denial, to neutralize the observed Black-White odds ratio of 2.3 in 2021. In fact, making the numbers work would require the low-score effect on denial to reach a relative risk of 35, which the authors describe as extremely large and implausible given that the minimum score for a conventional loan is 620.

These findings sit within a long and troubling research tradition. The landmark Boston Fed study by Munnell and McEneaney in the 1990s, which had access to confidential credit files, found Black and Hispanic applicants roughly 60 percent more likely to be denied; a more recent Federal Reserve analysis using automated underwriting data still found a residual 2 percent gap for Black applicants after controlling for credit scores and debt ratios. Studies of historic Homeowners’ Loan Corporation redlining maps have shown that neighborhoods graded as hazardous in the 1930s, typically those with high minority concentrations, still exhibit elevated lending bias today. Philadelphia itself carries a heavy institutional legacy: the Federal Reserve Bank of Philadelphia has documented that the city’s Black-White homeownership gap has barely narrowed in three decades, and recent enforcement actions, including a 2023 settlement in which ESSA Bank & Trust agreed to pay more than $3 million over a federal redlining lawsuit, show the problem is not merely historical.

The authors acknowledge the limits of their approach. Observational data of this kind is subject to selection bias, since applicants are not a random sample of the population, and to measurement error, and the public HMDA files simply do not contain everything a lender sees. The E-value, like the p-value, cannot prove causation on its own; it is a sensitivity metric, and its meaning depends on how strong the known confounders already are. But the authors argue that this is precisely the point: sensitivity analysis has been almost entirely absent from the mortgage discrimination literature, even though it is the central challenge for establishing causality in any observational study, and importing it from epidemiology gives fair lending researchers a way to express how robust their findings are to what the data cannot show.

The implications reach into policy. The authors suggest that the persistence of the Black-White gap, unchanged across six years of wildly different market conditions, points toward disparate impact operating through facially neutral underwriting thresholds on loan-to-value, loan-to-income and credit score criteria that lenders calibrate to minimize default risk but which may impose disproportionate barriers on Black applicants. If that interpretation holds, then closing the gap will require more than enforcing anti-discrimination statutes one lender at a time; it will require scrutinizing the threshold policies themselves, and, crucially, giving regulators and researchers access to the credit variables that have been missing from the public record for nearly fifty years.

Subject of Research: Racial disparities in mortgage lending decisions and sensitivity to unmeasured confounders in the Philadelphia metropolitan area

Article Title: Philadelphia then and now, disparities in mortgage lending and sensitivity to unmeasured confounders

Article References: Berkane, M., Alvarez, D., & Tanzi, N. (2026). Philadelphia then and now, disparities in mortgage lending and sensitivity to unmeasured confounders. Heliyon, 12(15), Article e45520. https://doi.org/10.1016/j.heliyon.2026.e45520

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45520

Keywords: mortgage lending, racial disparities, Philadelphia, HMDA data, E-value, unmeasured confounders, logistic regression, redlining, credit scores, fair lending, sensitivity analysis, homeownership gap