The study in brief · Authored summary
Lighting, visibility and the quality of crash records.
This study examines 6,190 police-recorded crashes in Eslamshahr from March 2022 to June 2024. The outcome distinguishes the 1,805 injury or fatal crashes from property-damage-only crashes. Predictive models were developed on 5,299 records and assessed on 891 later records; adjusted logistic regression examined infrastructure associations using the full cohort. Inadequate nighttime lighting and visibility obstacles showed elevated adjusted odds of casualty involvement. Calibration, road-width checks and predictor-removal tests show why recorded context and data quality matter when interpreting these results.
Selected reported findings
Explore three parts of the assessment.
Open each finding for the reported estimates and their interpretation.
01Model performance and calibration
Reference CatBoost achieved ROC-AUC 0.857 (95% CI 0.825–0.889), average precision 0.725 and Brier score 0.120. Sigmoid recalibration fitted using development data reduced Brier to 0.112. Mean predicted casualty involvement moved from 36.1% to 28.2%, compared with 27.3% observed in the assessment set.
The selected random forest reached ROC-AUC 0.863, average precision 0.752 and Brier 0.112. The 39-configuration comparison reused the previously inspected test set; it was not a new blind or external validation.
02Infrastructure associations
Relative to daylight, inadequately lit nighttime conditions were associated with higher casualty-involvement odds: OR 2.78 (95% CI 1.66–4.64). Visibility obstacles, compared with no obstacles, had OR 1.89 (95% CI 1.05–3.38).
These adjusted associations remained elevated across the reported width and influence checks. The findings identify lighting and visibility as priorities for field verification, not established effects of infrastructure interventions.
03Record quality and available predictors
Exactly 12 m appeared in 64.3% of recorded road widths. This heaping materially affected geometric estimates; it does not establish that every 12 m record is erroneous.
Removing crash-type and at-fault-vehicle information reduced CatBoost ROC-AUC to 0.532. Stronger performance with those fields therefore cannot be interpreted as evidence of a prospective infrastructure-only prediction tool.
Methods and scope
Different questions, different model roles.
Predictive classifiers used 5,299 development records from 1401–1402 and 891 later records from spring 1403 (March–June 2024). Their assessment covers discrimination, probability calibration and predictor availability. Separately, unpenalized adjusted logistic regression used all 6,190 records to estimate infrastructure associations. Width and influence checks examine sensitivity to recorded geometry and influential observations.
The study concerns injury/fatal versus property-damage-only outcomes among recorded crashes. It does not estimate crash occurrence risk or demonstrate causal effects of infrastructure changes.
Sources and access
Read the final published article.
This independently written summary draws on the final article's methods and results (Sections 2–3; Tables 4–6 and A8–A9). Read the publisher's article page or final PDF for complete definitions, analyses and limitations. The article is published under CC BY 4.0.