Researchers in developing countries are often faced with the small sample size issue in the crashdata during the crash frequency analysis, due to the incomplete crash and information recordingand reporting systems. Traditional generalized linear model (GLM) for crash frequency modelingis usually based on crash counts aggregated over multiple years, and would have a poorperformance by the limitation of sample size. Treating crash count in each year as a separateobservation could enlarge the sample size, but would create temporal correlation among crashesin different years which could bias the model estimates. In this study, we evaluated theapplication of generalized estimating equation (GEE) procedure to deal with the above dilemmafaced by researchers in developing countries. Using a four-year crash data at exit ramps on afreeway in China, two traditional GLMs based on yearly aggregated and disaggregated data aswell as a GLM with GEE procedure were estimated. The results showed that the models basedon yearly disaggregated data generally performed better than the one with yearly aggregated datadue to the enlarged sample size. The traditional GLM underestimated the standard errors ofcoefficients for explanatory variables which could lead to incorrect inferences. The GEEprocedure with a exchangeable correlation structure successively captured the temporalcorrelation in the longitudinal data of this study and was considered to outperform the traditionalGLMs in estimating the impacts of variables on crash counts.
展开▼
