Answers are grounded in this paper and cite their sources.
To determine at what times of day crashes at signalized intersections are most likely to be severe and which intersections warrant field investigation, using methods suited to a small number of sites, where site-level covariates are fixed within each intersection and the effective sample size for their effects is the number of intersections rather than the number of crashes. We analyzed 1,966 crashes at 14 signalized intersections in Erzurum, Türkiye, over ten years. Crash-level hypothesis testing used a mixed-effects logistic regression with a periodic (sine-cosine) hour-of-day term, validated with a 2,000-replicate cluster bootstrap. Intersection-level prioritization used empirical Bayes rankings from two independent models. The two tasks were kept separate. Crash severity varied strongly with hour of day (likelihood-ratio test, p < 0.001) and peaked at about 03:30 (95% bootstrap interval 02:20–05:00). Given that a crash occurred, the estimated probability that it was severe was 34% at 03:00 and 17% at 14:00 (ratio 2.0; 95% bootstrap CI 1.3–2.8). Once hour of day was modeled flexibly, a day/night indicator added no information (p = 0.96), so the apparent night effect is better described as a post-midnight effect than as an effect of darkness. Single-vehicle crashes had higher odds of severity (odds ratio 1.34), but this association was less certain under the cluster bootstrap. Intersection 12 was flagged by both empirical Bayes rankings as a priority for field investigation, and Intersection 2 by the severity-based ranking alone. Split-conformal prediction showed that predictions for unseen intersections are less stable than predictions for new crashes at observed sites. Exposure-adjusted crash rates declined significantly over the study period. Crashes in the late-night (post-midnight) hours, 00:00–05:59, are more likely to be severe than crashes at other times, including other hours of darkness; this describes severity among reported crashes, not exposure-based risk. This points to time-of-day factors such as low traffic volume, fatigue, and impairment, which should be examined with exposure and behavioral data. Separating crash-level testing from intersection-level prioritization, and checking small-cluster inference with a cluster bootstrap, offers a template for other single-city injury-surveillance programs.
No takes yet. Share an insight, caveat, or question.
Atalay et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: