UAP sightings cluster where the seafloor drops fastest.

Recently I posted Phase 1. UAP reports cluster near submarine canyons, even after controlling for population. I tried to challenge it. Maybe the signal is just coastal density the controls didn’t catch?

So I ran four additional analyses, each attacking the population confound differently. Here’s what I found.

The short version

1. The signal survives, but it’s sharper than I thought.

Phase 1 detected canyon cells using a 20 m/km gradient threshold and found odds ratios of 3–5× against population-matched controls. Phase 2 breaks that range into steepness bins, and only the steepest (60+ m/km, which maps to 85% of actual mapped submarine canyons) survives finer-grained population controls. Below that steepness, the signal disappears. Above it: odds ratio 3.90 [1.42–10.83], meaning reports are roughly 4× more likely near steep canyons than expected.

That’s lower than Phase 1’s headline OR of 5.30, because Phase 2 uses more conservative population controls that work at sub-county resolution instead of county level. The effect is real but smaller than it first looked. Maybe closer to something that actually makes sense.

2. It’s not just spatial. It’s also temporal.

Reports near canyons also cluster in time. Not a steady background hum. Episodic bursts. A few reports in the same area within days of each other, then nothing. I found 61 such clusters. The top 5 are all within 10 km of a canyon: three in Puget Sound, two in Southern California. Specific coordinates and dates in the repo.

Figure 2 – 2×2 results panel

What changed from Phase 1?

Phase 1 showed that spatial association is real and survives metro removal and placebo shelf tests. Phase 2 sharpens it.

The five flap episodes

The temporal test found 61 spatio-temporal clusters. Here are the top 5:

These are exact coordinates and date ranges. Checkable against independent records. Full episode map in the repo.

What this doesn’t prove?

The temporal clustering could be social contagion. One person reports something, neighbors look up and report too. The 60+ m/km threshold could be geometric, canyon mouths sit right at the coast where people live, and the controls may not fully capture that. The confidence interval on the odds ratio spans from 1.4 to 10.8, which is almost an order of magnitude. And I can’t control for observer type: fishermen and sailors see different things than suburban residents.

This is a pattern in self-reported data. It measures reporting behavior, not the phenomenon.

What would settle it?

Hydrophones. NOAA passive acoustic arrays sit near Puget Sound, La Jolla, and Monterey. Exactly where the flap episodes concentrate. Underwater, there’s no reporting bias. If anomalous acoustic signatures show up at the same coordinates and dates listed above, the reporting-bias explanation dies.

The flap table gives exact where and when. That’s a testable prediction.

For the technically minded

Full methodology below. Same 42,008 coastal NUFORC reports and 19,977 population-matched controls as Phase 1. All code, data, and intermediate outputs in the repo.

Detailed methodology:

Temporal permutation test

For each report, I find all other reports within 50 km, then count how many fall within ±7 days. The ratio of observed to expected temporal neighbors gives an excess score, normalized for local density. Test statistic: median excess near canyons minus far from canyons. Null: shuffle dates within each calendar year (1,000 iterations) or within each month (200 iterations, stricter).

Within-year: z = 6.18, p < 0.001. Within-month: z = 4.05, p = 0.015.

The signal lives in the tails. Trimming to the 5th–95th percentile reverses the effect (z = −5.32). It’s driven by rare, sharp bursts, not a diffuse background. 10/36 parameter combinations (temporal window × spatial radius × canyon threshold) are significant after FDR correction.

GAM partial dependence

Generalized additive model with 7 covariates: distance to canyon, coast, military bases, population density, ocean depth, port distance, and port count. Thin-plate spline on canyon distance (8 basis functions, AIC-selected). The partial effect spans 2.77 log-odds over 0–300 km, with most of the drop in the first 50 km. GAM beats linear on all metrics (AIC 68,612 vs 68,774, CV AUC 0.675 vs 0.657).

Weighted odds ratios by canyon steepness

Phase 1’s county-matched ORs of 3–5× don’t fully resolve within-county density gradients along canyon coastlines. Importance weighting (1/sampling score) isolates the canyon-specific component at sub-county resolution, with 2,000 bootstrap iterations.

Results: only the 60+ m/km bin (weighted OR 3.90 [1.42–10.83]) excludes 1.0. Lower gradient bins don’t survive weighting. This is a binary threshold, not dose-response.

Phase 1’s county-matched ORs are 2–3× higher across all bins. The difference reflects within-county population gradients. Importance-weighted estimates are the more conservative measure.

Cluster bootstrap

Standard errors assume independence. UAP reports from the same area aren’t independent. Cluster bootstrap (2,000 resamples, 4,057 spatial clusters): β = −0.166, CI [−0.258, −0.074]. The CI is 4.4× wider than naive but still excludes zero.

Per-distance cluster-bootstrapped ORs: 1.21 at 10 km [1.09–1.34], 1.18 at 25 km [1.08–1.29], 1.13 at 50 km [1.06–1.21].

Code, data, and full tables: https://github.com/antoniwedzikowski-rgb/uap-canyon-analysis

Analysis designed by me. Code generated with Claude Code. Writeup edited with AI assistance. I welcome methodological critique.