Every research site publishes its findings. This page publishes the other half: analyses we ran that produced no signal, and claims we withdrew after testing them properly. It exists because a method you only see when it works is not a method you can judge.
A plausible idea: when a lender's pipeline fills up, operational strain shows up as a higher denial rate — which would make volume a leading indicator of door tightness, and a useful thing for a counterparty desk to watch. We tested it on the 2018–2025 panel: within each lender, does the annual denial rate move with the log of annual FHA volume?
| Observations (lender-years) | 77 |
| Lenders | 11 |
| Pooled within-lender correlation | r = -0.208 |
| 95% confidence interval | [-0.428, 0.036] — crosses zero |
| Permutation test (years shuffled within lender) | p = 0.148 |
| Implied slope | -0.26 points per 10% volume increase |
The interval spans zero and the permutation test does not reject chance. There is no detectable within-lender relationship between volume and denial rate in this panel, in either direction. We are not publishing a capacity indicator, because we do not have one.
Per-lender numbers are deliberately not shown. With seven annual observations each, individual correlations of ±0.5 are indistinguishable from noise, and printing them would invite exactly the misreading we would then have to correct. Data: capacity-elasticity-2025.json.
We had published that FHA and conventional denial rates are essentially uncorrelated within the same institution (r = 0.056), and drew the tidy conclusion that “lender quality is product-specific”. Checking it before putting it in a paper killed it: the sample is six lenders, Pearson 0.054 contradicts Spearman 0.543, the 95% interval runs from −0.79 to +0.83, and deleting one institution moves Pearson to 0.835 — deleting the outlier creates a relationship rather than weakening one.
The correlation claim was withdrawn and logged. What survives is narrower: at least one large institution behaves very differently across products in the same year (FHA 8.9% vs conventional 47.0%, on tens of thousands of decisions each), which is enough to show program-invariance cannot be assumed and not enough to say how common it is. Corrections log · full leave-one-out table.
If a lender denies more than its applicant mix predicts, one story is that it is simply cautious and prices accordingly; another is that borrowers who get through a hard door pay more for the privilege. Both are testable. For 56 lenders with enough purchase volume to have a stable median, we compared the observed/expected denial ratio against the lender's median rate spread over APOR on the loans it did originate.
| Lenders compared | 56 |
| Pearson / Spearman | -0.179 / -0.249 |
| 95% confidence interval | [-0.422, 0.088] — crosses zero |
| Leave-one-out range | -0.302 to -0.12 (stable, unlike the cross-program case) |
| Permutation test | p = 0.195 |
No detectable relationship in either direction. The sign leans slightly negative — stricter doors, if anything, marginally cheaper — but the interval is consistent with anything from a moderate negative association to none at all, and shuffling the pairs reproduces this correlation about one time in five.
The extremes make the point better than the coefficient does. The softest door in the comparison denies 1.7% of purchase applications and carries one of the highest median spreads in the set; a lender denying nearly twice its expected rate sits close to the market on price. Strictness and price are not two faces of the same policy, at least not at the level a public record can see. Data: strictness-vs-pricing-2025.json.
This one has a consequence we are carrying publicly: it was the analysis we intended to offer as a Q1 2027 layer for Counterparty Screen subscribers. It produced nothing, so it will not be sold, and the roadmap has been amended rather than quietly reworded.
Two reasons, both practical. First, a null result is information: if you were about to build the same capacity indicator, this page saves you the work, and if you think our test was wrong, everything you need to redo it is linked. Second, it is the only honest way to read our positive findings. The door effect · What Denial Rates Cannot See (SSRN 7423798, doi:10.2139/ssrn.7423798) and the footprint gap survived the same treatment that killed the two analyses above; that is what makes them worth reading.
If an analysis here is later revived by better data or a better design, it will be published as a new result with its own date — not quietly moved off this page. Related: how we bind ourselves before measuring · how our own tables reconcile.