Complementary Goodness of Fit Procedure for Crash Frequency Models

TL;DR

This paper introduces a new visualization-based goodness of fit procedure for crash frequency models, specifically evaluating three GLM methods with ten years of Hawaii roadway data, enhancing model selection and validation.

Contribution

The paper proposes a complementary goodness of fit procedure for crash models, providing visual assessment and pseudo R2 metrics, and compares its effectiveness across different estimation methods.

Findings

All three models fit crash data well within narrow predicted ranges.

The procedure complements traditional statistics like AIC and BIC.

It diverges from traditional criteria for GLMM-NB models, indicating different model evaluation insights.

Abstract

This paper presents a new procedure for evaluating the goodness of fit of Generalized Linear Models (GLM) estimated with Roadway Departure (RwD) crash frequency data for the State of Hawaii on two-lane two-way (TLTW) state roads. The procedure is analyzed using ten years of RwD crash data (including all severity levels) and roadway characteristics (e.g., traffic, geometry, and inventory databases) that can be aggregated at the section level. The three estimation methods evaluated using the proposed procedure include: Negative Binomial (NB), Zero-Inflated Negative Binomial (ZINB), and Generalized Linear Mixed Model-Negative Binomial (GLMM-NB). The procedure shows that the three methodologies can provide very good fits in terms of the distributions of crashes within narrow ranges of the predicted mean frequency of crashes and in terms of observed vs. predicted average crash frequencies for those data segments. The proposed procedure complements other statistics such as Akaike Information Criterion, Bayesian Information Criterion, and Log-likelihood used for model selection. It is consistent with those statistics for models without random effects, but it diverges for GLMM-NB models. The procedure can aid model selection by providing a clear visualization of the fit of crash frequency models and allowing the computation of a pseudo R2 similar the one used in linear regression. It is recommended to evaluate its use for evaluating the trade-off between the number of random effects in GLMM-NB models and their goodness of fit using more appropriate datasets that do not lead to convergence problems.