Causal inference framework for generalizable safety effect estimates.

Area of Science:

• Traffic Safety Engineering
• Statistical Modeling
• Causal Inference

Background:

• Empirical Bayes (EB) before-after method is common for safety effect estimation.
• Current EB methods primarily estimate Average Treatment Effect for the Treated (ATT).
• Traffic safety research often requires Average Treatment Effect (ATE) for broader applicability.

Purpose of the Study:

• To integrate causal inference with the EB method for generalizable safety effect estimates (Crash Modification Factors - CMFs).
• To develop modifications to the EB method for estimating Average Treatment Effect for the Untreated (ATU) and ATE.
• To assess the accuracy of different regression models in estimating ATE CMFs.

Main Methods:

• Applied causal inference framework to the EB before-after method.
• Developed and tested approaches for estimating ATT, ATU, and ATE.
• Utilized standard negative binomial and mixed effects negative binomial regression models.
• Analyzed a dataset with 'no-treatment' scenarios (random and crash-history based selection).

Main Results:

• The standard EB method was found to estimate ATT, not the desired ATE.
• Modified EB methods were proposed to estimate ATU and ATE.
• ATE CMFs derived using the standard negative binomial regression were the most accurate.
• The study identified potential sources of bias in the EB method.

Conclusions:

• The enhanced EB method with causal inference provides more accurate and generalizable CMFs.
• Standard negative binomial regression is recommended for ATE CMF estimation in this context.
• Further investigation into EB method biases is warranted for robust traffic safety analysis.