On prediction in road safety

Abstract

Prediction is about potential outcomes: what will happen if and what would have happened if. The first question arises when safety targets are set, the second when the effect of an intervention on safety is to be evaluated. There are many ways to predict. For the same data different prediction methods produce different predictions. What targets are set and what estimates of intervention effect are produced will depend on what method of prediction is chosen. Therefore one has to determine what method tends to predict best. To do so empirically one asks what method would have predicted best had it been applied in the past and then one assumes, inductively, that the same would apply in the future. Quantitative measures of prediction quality are suggested and it is shown how these measures of prediction quality allow one to determine which of two prediction methods should be preferred. The suggested approach was applied to two data sets: The time series of motor vehicle accident fatalities in Province A and in Province B. On the basis of this analysis one may draw tentative conclusions for these jurisdictions and the methods tested; one can say what method seems preferable, what is the average size of bias than needs to be corrected and how accurate is the prediction likely to be. Broader conclusions will emerge once many additional methods of prediction are applied to data from many other jurisdictions and pertaining to a variety of circumstances.