Sighting-Report Database Clustering Analysis

Case Database

Noise-report rate 140

Cluster count 3

Cluster tightness 0.55

Detector sensitivity 1.00

Reports total

–

Clark-Evans R

–

Detected clusters

–

False-positive rate

–

NN-distance histogram

Observed nearest-neighbor distances (green) vs the Poisson-random null model (dashed line).

How it works

Every marker is one filed sighting report. Most are independent "noise" — hoaxes, misidentified planes or Venus, duplicate filings — scattered uniformly at random (a Poisson process). A handful come from a real correlated event: many independent witnesses reporting the same actual phenomenon, which places their reports close together in space. The database itself cannot tell them apart by label — only their spatial pattern gives it away.

For every report we compute the distance to its single nearest neighbor. Under a pure Poisson process with report density λ (reports per unit area), the expected mean nearest-neighbor distance is:

r_expected = 1 / (2·√λ)
R = r_observed_mean / r_expected

This is the Clark–Evans statistic. R ≈ 1 means the reports look purely random. R < 1 means reports are packed closer together than random noise predicts — evidence of real clustering. A detector flags any report whose own nearest-neighbor distance falls well below r_expected, then groups adjacent flagged reports into candidate clusters (must contain ≥3 reports to count).

Noise-report rate — how many independent random reports contaminate the database; more noise makes real clusters harder to distinguish by chance alone.
Cluster count — how many genuinely correlated events are seeded into the database.
Cluster tightness — spatial spread of reports around a real event; tighter clusters are easier to detect, looser ones blend into the noise floor.
Detector sensitivity — how far below r_expected a report's nearest-neighbor distance must fall to be flagged; too high and it misses loose clusters, too low and it start flagging noise.

Real-world relevance: this is exactly how researchers triage large citizen-report databases (UFO sightings, disease case reports, crime incidents) — density-based spatial clustering against a random null model is the standard first pass before anyone looks at individual case content.

Unflagged report Flagged (in detected cluster) Ground truth: real event