Reflections on methods of statistical inference in research on the effect of safety countermeasures

Abstract

Sensible management of traffic safety is predicated on having reasonable expectations about the effect of various safety countermeasures. It is the role of evaluative research to derive such intelligence from empirical data. In spite of decades of research and experience, the safety effect of many countermeasures remains unknown. This sorry state of affairs is largely due to the objective difficulty of conducting conclusive experiments. Recognition of this objective difficulty should lead to the realization that in transport safety, knowledge is accumulated gradually from small, noisy and diverse experiments. The statistical tools used to extract knowledge from data should reflect this aspect of reality. One must therefore question the usefulness of classical tests of significance as a device for scientific progress in this field. It is argued that the unquestioning and all-pervasive use of significance testing in evaluative research on transport safety amounts to a self-inflicted learning disability. In contrast, it is shown that classical Point Estimation. Likelihood-Support and Bayesian methods can all make good use of experimental evidence which comes in small doses. In particular, the likelihood function is an efficient device for the accumulation of objective information and a necessary ingredient for Bayesian decision analysis.