The veil-of-darkness test

Published July 2026 · Data window: MPD stops, 2019–2025, inter-twilight window (4:45 PM to 9:09 PM local time)

The question: Does daylight, when officers can see race before stopping someone, change who gets stopped?

The answer: No. The test finds no daylight pattern: the disparities on this site are not explained by officers reacting to what they can see.

If officers can see race before deciding whom to stop, daylight stops should skew Blacker than after-dark stops at the same clock time. Daylight saving transitions make that comparable: 7:00 PM is bright in June and dark in December. The classic daylight signature is NOT present: at the same clock time, stops after dark were slightly more likely to be of Black people (odds ratio 0.93, 95% CI 0.90 to 0.96). This test probes one specific mechanism, whether daylight visibility changes who is stopped, and in a densely lit city its premise, that race is hard to see after dark, is at its weakest. It does not offset the representation, search, and arrest disparities documented on this site.

0.93

Mantel-Haenszel odds ratio, daylight vs after dark, that a stopped person is Black, pooled across 30-minute clock strata. 1.00 means lighting makes no difference; above 1.00 means daylight stops skew Blacker.

Black share of stops by clock time within the inter-twilight window. Points rest on at least 100 stops; the gap between the lines, inside the same 30-minute stratum, is what the odds ratio pools.

Stops	Daylight Black share (N)	After-dark Black share (N)	MH odds ratio	95% CI	Significant
All stops (clock strata)	77.8% (78,104)	80.2% (49,843)	0.927	0.896 to 0.959	yes
All stops (clock and district strata)	77.8% (78,104)	80.2% (49,843)	1.024	0.987 to 1.063	no
Ticket Only, traffic proxy (clock strata)	68.4% (41,804)	71.8% (25,367)	0.929	0.891 to 0.970	yes
Ticket Only (clock and district strata)	68.4% (41,804)	71.8% (25,367)	1.064	1.016 to 1.114	yes
All stops, transition window (clock and season strata)	78.2% (19,784)	80.1% (24,425)	0.976	0.902 to 1.056	no
All stops, transition window (adding district strata)	78.2% (19,784)	80.1% (24,425)	1.034	0.949 to 1.126	no
Ticket Only, transition window (clock and season strata)	68.7% (10,861)	71.6% (12,624)	0.933	0.849 to 1.026	no
Ticket Only, transition window (adding district strata)	68.7% (10,861)	71.6% (12,624)	1.046	0.944 to 1.159	no

Holding district constant dissolves the pattern: within districts the test cannot tell daylight from after-dark stops apart (odds ratio 1.02, not significant, vs 0.93 citywide). The citywide lighting difference is carried by WHERE policing happens, not by lighting itself: direct evidence that the disparity operates through deployment geography rather than through officers reacting to what they can see.

One signal does survive district stratification: within districts, at the same clock time, Ticket Only stops made in daylight were slightly more likely to be of Black people (odds ratio 1.06, 95% CI 1.02 to 1.11). That is the direction the classic visibility mechanism predicts, small, and confined to the traffic-stop proxy.

Seasonal check: restricted to the 44,209 stops within 30 days of a daylight-saving clock change, where season is held nearly constant, the test still cannot tell daylight from after-dark stops apart (odds ratio 0.98, 95% CI 0.90 to 1.06): the absence of a daylight signature is not a seasonal artifact. The Ticket Only daylight skew cannot be confirmed on the restricted sample (odds ratio 1.05, 95% CI 0.94 to 1.16, not significant on 23,485 stops): the full-year figure carries a seasonal caveat until more transition-window data accumulates.

How this number is built (and where it's soft)

Design (Grogger & Ridgeway 2006). Only stops inside the inter-twilight window (4:45 PM to 9:09 PM local) are compared: the clock range that is daylight in one season and dark in another. Comparisons happen inside 30-minute clock strata, so 5:00 PM is only ever compared with 5:00 PM.
Lighting from astronomy, not guesswork. Sunset and civil dusk are computed per date for DC (NOAA solar algorithm, unit-tested against published times). Stops between sunset and civil dusk are ambiguous lighting: excluded and reported (13,940 stops).
What this test can and cannot say. It probes one mechanism: whether seeing race changes stop decisions. In a dense, artificially lit city the premise that race is invisible after dark is weak, so a null or reversed result here neither confirms nor offsets the representation, search, and arrest disparities on this site, which do not depend on this mechanism.
District-stratified variant. The second estimate stratifies by clock bin AND police district, holding WHERE policing happens constant as well as when. Comparing the two separates lighting effects from deployment geography. Stops with no mappable district (1,100) are excluded from that estimate only, and counted.
Seasonal robustness (transition window). A full-year pool has a known confound: at the same clock time, the daylight arm is mostly summer and the dark arm mostly winter, and who is on the street differs by season for reasons unrelated to visibility. The transition-window rows rerun the estimator on stops within 30 days of a daylight-saving clock change, where season is nearly constant and the clock change itself supplies the lighting contrast. Strata there are clock bin and transition season (spring or fall), plus district in the stratified variant, so a spring stop is never compared with a fall stop.
No vehicle flag. MPD stop rows carry no vehicle/pedestrian split, so the headline runs on all stops; the Ticket Only subset is the closest proxy for the classic traffic-stop design and is reported alongside.
Exclusions (reported, not dropped quietly). Unknown-race stops are excluded upstream; stops outside the window (89,387 in the evening fetch) are outside the design; ambiguous-twilight stops as above.
Grain and window. Stops of persons, both eras pooled (Jul 2019 to 2025). Odds ratio with Robins-Breslow-Greenland variance; significance at the 5% level, and a direction is only ever asserted when significant.