Operationally Conditioned Transit Injury Risk: A National Machine Learning Analysis of Safety, Mobility, and Socioeconomic Exposure.

Open Access
Article
Conference Proceedings
Authors: Timothy Kangogo, Oladotun Adeola, Barbara Asamoah, Anthony Dontoh, Judith Mwakalonge, Gurcan Comert, Varghese Vaidyan, Vijay Bendigeri
Abstract

Transit injury risk may vary with operating conditions and community context, but the usefulness of these factors for safety screening must be assessed against agencies' prior injury experience. We examined whether operations, socioeconomic conditions and ambient mobility improve injury prediction beyond ridership exposure and history, while distinguishing prediction from moderation of associations. National agency-mode-month records linked transit safety and service data to headquarters-county context. The broad sample comprised 84,721 observations from 2019–2025. Poisson regression, negative binomial regression and gradient-boosted trees used identical samples within each comparison, with 2019–2023 training, 2024 validation, retrospective evaluation in January–June 2025 and a reserved July–December 2025 evaluation after model settings were fixed. History reduced validation Poisson deviance by 37.5% beyond exposure, mode and calendar controls. A smoothed historical-rate benchmark remained competitive, with close calibration of reserved-period national totals but variation across subgroups. Socioeconomic variables improved boosted validation deviance by 0.0219; gains in both 2025 evaluations were positive but uncertain (0.0279 and 0.0177; 95% confidence intervals spanning zero). Headquarters-county mobility showed no reliable incremental improvement in either available period; earlier exploratory tests did not establish mobility–vulnerability moderation. A secondary bus analysis found no clear difference between collision and non-collision service-hour associations: the ratio of rate ratios for 10% greater hours was 0.985 (95% confidence interval 0.936–1.035). These findings support historical benchmarking and subgroup calibration before local safety review. They neither identify causal service effects nor establish the absence of mobility-related vulnerability

Keywords: Transit Safety, Injury Prediction, Historical Benchmarks, Exposure, Calibration, Socioeconomic Context, Ambient Mobility

DOI: 10.54941/ahfe1008464

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← Human Factors and Propulsion Failures in U.S. General Aviation Accidents: What a Decade of NTSB Findings Implies for Electric Training AircraftHuman-Centered Creative Support for VE Idea Realization through Contradiction Analysis →
View all articles in Human Factors in Design, Engineering, and Computing →