Abstract
Mortgage denial rates computed from the Home Mortgage Disclosure Act (HMDA) record are widely used as measures of lender behaviour. This paper argues that a lender-level denial rate is not a measure of institutional strictness but a conditional observation, bounded in four systematic ways. Using 2025 HMDA records — with the primary analyses restricted to FHA forward loans and a supplemental cross-program analysis of lenders carrying both FHA and conventional books — I document that (i) in 151 metropolitan areas where both compete, lenders operating in 40 or more metros denied 23.6% of applications against 16.7% for narrower-footprint lenders, a median within-metro gap of 10.2 points, while narrow-footprint lenders also occupy both extremes of the national distribution (1.8% and 78.7%); (ii) the share of denials citing collateral falls monotonically with loan size, from 25.9% below $100,000 to 14.1% at $300–400,000, and drops out of the three most-cited reasons above $400,000, so that "property-driven" denial is a small-loan rather than a high-value phenomenon; (iii) an attempt to estimate whether denial behaviour is program-invariant is not identified on the available sample — Pearson 0.054 against Spearman 0.543, a 95% interval of [−0.79, +0.83], and 0.835 after a single deletion — so the failure is reported in place of the estimate; and (iv) the record cannot observe exclusions that occur before an application exists, a boundary illustrated by the New York metropolitan area, which recorded 7,144 decisioned FHA applications in 2025 — 31st of 319 covered metros — while remaining the largest housing market in the country. None of these results identifies a causal mechanism. Together they suggest a reading rule: the same denial rate is not the same estimand, the same program, the same applicant universe, or the same observability boundary. All data, receipts and code-level definitions are published under CC BY 4.0.