Although several approaches to estimating prevalence and excess risk exist, each relies on behavioral assumptions that are subject to credible objections. In this article, we compare the assumptions of the most widely accepted approach over the past three decades-results from the National Roadside Survey (NRS)-with a recently revived model-based alternative that estimates these population parameters from the observed distribution of fatal motor vehicle crashes. Comparing estimates of prevalence covering the past four decades, we find that when driver non-response rates in NRS are small, estimates of the prevalence of alcohol-involved driving are nearly identical between methods, suggesting that the underlying behavior assumptions of both models approximately hold. For the past two decades, however, as the rate of driver refusal in the NRS has increased substantially, prevalence estimates between methods have diverged. A counterfactual analysis reveals that the estimates for drinking-and-driving from the model-based approach should be taken as at least as valid as those from the NRS. That is troubling as these methods yield markedly different conclusions about the continued effectiveness of existing traffic safety policy: the NRS finds that the prevalence of drinking-and-driving has fallen monotonically over time, while estimates from the model-based approach suggest that prevalence has plateaued at 15% for the past two decades. More unsettling however, is the conclusion that researchers and policy-makers may know very little about the extent of legally-impaired driving or how it has changed over time.
Keywords: Fatality Analysis Reporting System; National Roadside Survey; drinking-and-driving; externality.