Function 3 · fair lending and adverse action · real HMDA data

Are some applicants denied more often than comparable applicants?

In New Jersey's 215,881 decided mortgage applications for 2024 and 2025, Black applicants were denied at 24.0% against 10.7% for white non Hispanic applicants. Holding income, loan size, loan to value, debt to income, loan type and purpose, the underwriting system and the lender fixed, the odds ratio falls from 2.63 to 1.76. The gap narrows and does not close, and the public file has no credit score or underwriting result, so what remains is a question to ask of a lender's own data, not a finding.

Applications with a decision
215,881
646 lenders
Denial rate, all applicants
14.4%
Black applicants, adjusted odds ratio
1.76
1.67 to 1.85
Comparisons surviving Benjamini and Hochberg
7 of 8
Adjusted, at q below 0.05

Denial disparities against the reference group, before and after adjustment

0.50.7511.52Odds of denial against the reference group (log scale; 1 is parity)Race and ethnicity, against white non Hispanic applicantsBlackNot providedHispanic or LatinoAsianOther or jointSex, against male applicantsNot providedFemaleJoint
  • Unadjusted odds ratio
  • Adjusted odds ratio
Black applicants were denied at 24.0% against 10.7% for white non Hispanic applicants, a ratio of 2.24, or an odds ratio of 2.63. With income, loan size, loan to value, debt to income, loan type and purpose, the underwriting system and the lender held fixed, the odds ratio is 1.76 (1.67 to 1.85). Adjustment narrows the gap and does not close it; the public file has no credit score or underwriting result, so what remains is a question for a lender's own data, not a finding.

Source: real:hmda, New Jersey HMDA 2024 and 2025, first lien owner occupied home purchase and refinance applications on 1 to 4 unit homes that reached a decision, as of 2026-08-31.

Denial rate by county, unadjusted and adjusted for the application mix

10.2%18.5%State: 14.4%
Unadjusted, Essex County denies the most (18.5%) and Hunterdon the least (10.2%) against 14.4% for the state. Adjusted for each county's applications and lenders, Hudson still denies 1.31 times the denials its mix predicts. County is a map, never a model input: it is on the proxy list.

Source: real:hmda, New Jersey HMDA 2024 and 2025, first lien owner occupied home purchase and refinance applications on 1 to 4 unit homes that reached a decision, as of 2026-08-31.

Before adjustment

DimensionGroupApplicationsDenialsDenial rateRatio to reference95% low95% highBH q value
race and ethnicityBlack18,5504,44824.0%2.242.172.310.0000
race and ethnicityNot provided35,1837,26720.7%1.931.881.980.0000
race and ethnicityHispanic or Latino28,9715,01017.3%1.621.571.670.0000
race and ethnicityAsian28,1683,16011.2%1.051.011.090.0172
race and ethnicityWhite non Hispanic (reference)95,39710,21410.7%not applicablenot applicablenot applicablenot applicable
race and ethnicityOther or joint9,6129519.9%0.920.870.980.0172
sexNot provided14,0143,86627.6%1.741.691.800.0000
sexFemale45,4627,25816.0%1.010.981.040.5053
sexMale (reference)72,17811,41815.8%not applicablenot applicablenot applicablenot applicable
sexJoint84,2278,50810.1%0.640.620.660.0000

Source: real:hmda, New Jersey HMDA 2024 and 2025, first lien owner occupied home purchase and refinance applications on 1 to 4 unit homes that reached a decision, as of 2026-08-31.

After adjustment

Asian applicants are denied about as often as white non Hispanic applicants before adjustment (a rate ratio of 1.05) and carry an adjusted odds ratio of 1.42: on the observed controls their files are stronger, so equal files are denied more often. Adjustment can widen a gap as well as narrow one.

DimensionGroupOdds ratio95% low95% highBH q valueFinding at q < 0.05
race and ethnicityBlack1.761.671.850.0000higher odds of denial
race and ethnicityHispanic or Latino1.421.361.490.0000higher odds of denial
race and ethnicityAsian1.421.341.490.0000higher odds of denial
race and ethnicityNot provided1.401.331.470.0000higher odds of denial
race and ethnicityOther or joint1.091.001.190.0539not distinguishable from the reference
sexNot provided0.910.850.980.0136lower odds of denial
sexFemale0.880.840.910.0000lower odds of denial
sexJoint0.760.730.790.0000lower odds of denial

Source: real:hmda, New Jersey HMDA 2024 and 2025, first lien owner occupied home purchase and refinance applications on 1 to 4 unit homes that reached a decision, as of 2026-08-31.

The reasons lenders gave

Debt to income ratio is the most common first reason, on 31.6% of denials. Credit history comes first on 21.0% of denials to Black applicants against 18.3% for white non Hispanic applicants, and credit history is exactly what the public file leaves out.

Reason for denialDenialsShareShare, White non HispanicShare, BlackShare, Hispanic or LatinoShare, Asian
Debt to income ratio9,82031.6%30.9%33.8%36.8%35.3%
Credit history5,15016.6%18.3%21.0%14.4%7.4%
Credit application incomplete4,98016.0%13.6%11.9%11.2%13.5%
Collateral4,93115.9%18.3%15.5%15.2%18.2%
Other2,5308.1%8.3%7.5%9.1%9.7%
Unverifiable information1,6185.2%4.9%4.0%6.1%7.0%
Insufficient cash1,5775.1%4.4%4.9%5.2%7.2%
Employment history4291.4%1.3%1.4%1.9%1.6%
Mortgage insurance denied150.0%0.0%0.0%0.1%0.0%

Source: real:hmda, New Jersey HMDA 2024 and 2025, first lien owner occupied home purchase and refinance applications on 1 to 4 unit homes that reached a decision, as of 2026-08-31.

The proxy list, and the test that holds every model to it

32 protected attributes and proxies no credit or fraud model may use: age, applicant_age, applicant_zip5, birth_year, census_tract, city_pop, county_fips, customer_age, customer_lat, customer_long, date_of_birth_distinct_emails_4w, derived_ethnicity, derived_race, derived_sex, education, ethnicity, first_name, gender, last_name, lat, long, marital_status, marriage, name, name_email_similarity, national_origin, public_assistance_income, race, religion, sex, zip5, zip_code. Every published model's feature list is checked against it in the pipeline and in CI.

ModelFeaturesProtected attributes or proxies among them
aml_triage13none
attrition18none
cure13none
fraud_account27none
fraud_card18none
pd_challenger13none
pd_scorecard8none
relief5none

The scorecard's adverse action reasons come from the same discipline: the characteristics furthest below their best attainable points, never an entry on this list, and none for an approved applicant. Try it in the studio.

Method and limitations

  • HMDA's public file has no credit score and no automated underwriting result. They are the variables a lender's own analysis controls for first, and their absence is the main reason an adjusted gap here is not a finding.
  • Race, ethnicity and sex are HMDA's derived fields as applicants reported them. Applicants who did not report are their own group and are denied more often, which a lender's own review would look into.
  • Lenders with too few applications, or no denials, share one fixed effect; the rest are compared within lender.
  • Harborline Bank's own mortgage applications are generated, with no real HMDA filing behind them, so the method runs on the real New Jersey register, the data an examiner would start from.