How Data Imputation Affects Crash Modeling Results

Abstract

Crash modification factors (CMF) are used to determine safety treatments for highways. These are based on estimates from correlations of road geometry attributes with crash frequency. These estimates are derived from models using datasets that may have missing data and analysts may choose how to compensate for this. Using data from North Carolina, we examine changes in inferences associated with missing data by comparing models with full data sets with models where we omit vehicle kilometers traveled and impute the missing data. Results were mixed, showing notable changes for some variables. This could potentially lead to bad decisions in practice on how to improve road safety.