With few exceptions, model selection in traffic safety studies does not receive as much attention as do the methods implemented to estimate the parameters in those models. In this manuscript, we focus on the modeling step in an intervention study and discuss issues associated with formulation, interpretation, comparison and selection of models for intervention studies. All of the statistical models we consider rely on an over-dispersed Poisson assumption for the crash densities, and are fitted by Bayesian methods. The crash data we use arose from a study by the Iowa Department of Transportation to evaluate the effectiveness of converting roads from four lanes to three lanes. Deviance and the deviance information criterion (DIC) are used for model selection. In the Iowa road diet study, a subset of best models (which fit the data better than others) was then also used to carry out posterior predictive checks to assess model fit.