Timestamps, Not the Moon, Explain Crime After Dark

Across 116,846 property-offense records over three years, the raw night/day comparison is 1.080 — night looks busier. Remove the 10,168 records stamped at exactly 00:00 and the ratio drops to 0.899: night is 10% quieter. The reversal that defined the earlier article holds in a longer, weather-joined window.

That does not mean darkness reduces crime. It means a dataset where 8.7% of timestamps are exact midnight cannot decide the question on its own.

The shape of the sample makes the raw number look stronger than it is. Across 1,096 days the extract splits into 56,174 daylight and 60,672 night records, with 10,168 of the night records carrying that exact-midnight stamp. Mean daylight runs 12.24 hours a day, so raw counts already lean night simply because more of the calendar is dark. A rate comparison corrects for that; the timestamp handling still decides its sign.

If not the sky, then what? The model’s largest term by a wide margin is the first annual harmonic — a seasonal sinusoid with a coefficient of −0.13. That is roughly ten times the size of the weekend effect (−0.013), seven times the moon coefficient (0.018), and far larger than either weather term. Daily counts in this data are governed more by the time of year than by the phase of the moon or the day’s temperature.

That changes how to read any after-dark headline. A raw night/day split mixes seasonal exposure — longer summer nights, holiday periods, school calendars — with everything else. The environment controls do not rescue the after-dark rule; they show the rule was never really about the environment to begin with.

The negative-binomial model makes the ranking explicit. Across 1,096 days (AIC 6,582.81, dispersion α 0.0196), the annual harmonic dominates: sin1 at −0.130 and sin2 at −0.031 against an intercept of 4.647. Weather and moon coefficients are near zero — temperature 0.0013, precipitation −0.0021, moon illumination 0.0177 — and the weekend term is only −0.013. The season carries roughly ten times the signal of the day-of-week effect.

The full-moon cut agrees. On 38 full days the mean daily count was 109.737 versus 106.499 on all other days, a ratio of 1.030 — a 3% difference on a small set of days, not a scheduling signal.

Routine activity explains the pattern: property crime needs a target, an offender, and absent guardianship, and those arrange around daily and seasonal life far more than around the sky — the routine activity framing. Weather does move offense rates at the margin — econometric work on crime and weather links temperature and precipitation to measurable shifts across U.S. counties — but darkness here stays a measurement question, and the defensible input is the property’s own hourly pattern.