NC Traffic-Stop Veil-of-Darkness Racial Bias Test

What this research found

The veil-of-darkness test asks whether the racial mix of drivers police stop changes when darkness makes race harder to see, holding clock time fixed. Applied to 9,490,673 standardized North Carolina traffic stops from 2010 to 2015, the adjusted darkness odds ratio came out at 1.045 — slightly above one rather than below it, so the specific day-night signature the test looks for is absent. A parallel look at what happened after the stop found a much sharper pattern: Black drivers were searched at 2.10 times the White rate, and that gap barely moved when county and year were controlled.

• The fully adjusted darkness odds ratio is 1.045 (95% CI 1.024–1.067, p = 2.1 × 10⁻⁵) — the odds a stopped driver is Black are about 4.5% higher after dark at the same clock time, the opposite direction to the classic profiling prediction of an odds ratio below one.
• Clock time confounds the naive comparison badly. Adding 30-minute time-of-day bins and season shrinks the raw odds ratio from 1.121 to 1.046, absorbing roughly two-thirds of the unconditional day-night gap, and adding year leaves it essentially unchanged.
• Black drivers were searched in 3.674% of stops against 1.748% for White drivers, a disparity ratio of 2.10 (95% CI 2.08–2.12). Adding county and year fixed effects across 8,453,751 stops barely dents it (adjusted odds ratio 2.019), so the gap holds within counties and within years.
• Contraband turned up in 29.639% of searches of Black drivers versus 31.364% for White drivers — more searches yielding fewer hits, the textbook outcome-test signature of a lower evidentiary threshold.
• That hit-rate gap does not survive geographic controls: with county and year fixed effects the odds ratio is 0.986 (95% CI 0.966–1.007, p = 0.19), indicating the pooled deficit largely reflects Black drivers being concentrated in counties with lower overall hit rates rather than a within-county threshold difference.
• The strongest and most control-robust evidence of unequal treatment in these data sits in the decision to search, not in the visibility contrast at the stop or in the marginal productivity of searches.

How it was done

The Stanford Open Policing Project's standardized North Carolina statewide file was filtered from 20,286,645 raw stops to the 9,490,673 recorded between 2010 and 2015. Daily sunset and civil dusk were computed for every one of the 2,191 days at a single statewide reference point, Raleigh, with daylight-saving transitions handled through timezone-aware calculation. The inter-twilight window was set from the period extremes — earliest sunset at 17:00:27 and latest civil dusk at 21:05:08, a band 244.7 minutes wide — leaving 824,776 stops, of which 729,785 involved Black or White drivers. Logistic regressions then modelled the odds a stopped driver was Black on a darkness indicator across four nested specifications, adding 30-minute time bins, season, year, and finally standard errors clustered on the 100 counties. Separately, search and hit rates were compared by race with Wilson intervals, chi-squared and two-proportion tests, and then re-estimated with county and year fixed effects.

Data sources

• Stanford Open Policing Project — standardized North Carolina statewide traffic-stop file, release 2020-04-01 (20.29 million stops, 9,490,673 retained for 2010–2015)
• astral v3.2 — daily sunset and civil dusk for Raleigh, North Carolina across 2,191 days
• Grogger & Ridgeway, Journal of the American Statistical Association 101:878–887 (2006) — the original veil-of-darkness design
• Knowles, Persico & Todd, Journal of Political Economy 109:203–229 (2001) — the outcome test for search decisions