How Data Imputation Affects Crash Modeling Results

Adediji, Yemi; Noland, Robert · 2020 · DOAJ

DOI: 10.32866/001c.17386

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study investigates how data imputation for missing Average Annual Daily Traffic (AADT) values affects the results of crash frequency models used to determine Crash Modification Factors (CMFs). CMFs are critical for cost-benefit analyses of road safety treatments, relying on correlations between geometric design attributes and crash frequency. The authors hypothesize that analyst decisions regarding missing data processing, specifically imputation, can lead to incorrect inferences and potentially poor safety decisions. The research focuses exclusively on AADT because traffic volume is highly correlated with crashes, yet data collection is often incomplete compared to static geometric features. The analysis utilizes data from the Highway Safety Information System (HSIS) for North Carolina interstate highways, covering 8,071 segments over 1,944 km from 2009 to 2013. The dataset includes geometric variables such as pavement width, lane count, median width, shoulder width, and sinuosity, alongside crash counts for total crashes and fatal/incapacitating injury crashes. To isolate the effect of missing data, the authors employed negative binomial conditional autoregressive models estimated via Markov Chain Monte Carlo techniques to control for spatial correlation. They first established a baseline model using actual reported AADT data. Then, they randomly removed 10%, 30%, 50%, and 70% of AADT values and imputed them using an ordinary least squares regression approach. Vehicle Kilometers Traveled (VKT) was derived from these AADT values and included in all models. The findings reveal substantial and unpredictable variations in coefficient estimates when AADT is imputed. Contrary to expectations, the largest deviations from the baseline model did not consistently occur at higher missing data rates. For total crashes, the most significant differences appeared at 10% and 30% imputation levels, while for fatal and incapacitating injury crashes, the largest discrepancies occurred at 10% and 70%. Specific geometric variables, such as shoulder width and sinuosity, exhibited extreme percentage differences in their coefficients across different imputation scenarios. However, the coefficient for VKT remained relatively stable, indicating that the primary distortion caused by imputation affects the interpretation of geometric design factors rather than traffic volume itself. The study concludes that imputation decisions can substantially influence modeling results, raising concerns about the reliability of safety treatments based on such models. The authors note that their findings of large variations even at low imputation rates (10%) differ significantly from previous Federal Highway Administration studies, which found minimal bias. While the paper does not propose specific solutions, it emphasizes the need for caution when using imputed data and advocates for improved data collection practices to ensure accurate safety analysis.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• incidence prevalence

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes