Statistical methods to support induced exposure analyses of traffic accident data

When it is possible to identify the drivers involved in two-vehicle accidents as either at fault or innocent, induced exposure methods offer a way to assess the relative accident risk of driver subgroups, even when group-specific measures of exposure are unavailable. A cross tabulation of two-vehicle accidents by group membership of the at-fault and victim drivers forms a contingency table, and statistical methods derived from contingency table analysis can be used to make inferences concerning the variables arising in the induced exposure model. It is shown how the standard contingency table test for independence of row and column classifications provides a test of the assumption that the victims are sampled randomly and how an odds ratio statistic can be used to estimate the ratio of the accident rates between two driver subgroups. This estimator is asymptotically normally distributed, and a formula is given for estimating its standard error. An Empirical Bayes method for identifying sites where one driver subgroup has a significantly higher accident rate than does another is then presented. These procedures are illustrated using several actual accident data sets.