Reanalysis of traffic enforcement data from Victoria

There is an increased interest in the Netherlands in the safety effects of traffic enforcement measures. This regards the intermediate effect of enforcement on behaviour as well as the final effect on safety itself. In addition to these general safety effects, regional differences are also of concern. The current study was meant to identify a feasible and scientifically correct analysis of enforcement data for evaluation purposes. This was done using Australian data. In Victoria, Australia, monthly data of several types of enforcement and campaigns was gathered, together with background data and safety data. These series of monthly data were analysed using Harvey’s structural time-series technique. In cases like these, time-series analysis is superior to ordinary (log-linear) regression analysis, because it takes the developments of the traffic and safety systems into account. In order to learn from the Australian experiences, their data has been reanalysed, also using structural time-series analysis, however, this time applied to all the regional series together. It was assumed that if common trends in the series were represented by common parameters in the joint model, a more stable solution would result from the total set of regional series than from the separate analyses for each series. This study was carried out not to check the results and interpretation of the original study, but to see how such complicated multivariate time-series models, as intended to be used in the Netherlands, would work out on the Australian data. Various models were applied, with various (combinations of) enforcement and campaign variables included. These models were applied to the regional data separately as well as jointly. The model was applied to accident rates (accidents divided by vehicle kilometres). The basic structural model, consisting of trend, drift (Harvey, 1989), and monthly components, was extended by addition of alcohol sales and unemployment figures. To this extended basic model, (combinations of) enforcement variables were added. Chi-square tests were applied to measure the improvement of the model by the addition of enforcement variables. It turned out that in general the residuals are considerable. However, this seems to be more a result of the variation in the individual measurements, caused by the small number of accidents per month, than by the uncertainty in the model as such. It was also found that the monthly trends in accidents are different for different regions. Therefore, the number of parameters in the joint analyses was still considerable, because 11 monthly parameters were added for each series. The results showed that only small (insignificant) effects of campaigns were found and almost no effects of enforcement. This can be explained partly as a result of the conservative way of testing: all effects of enforcement and campaigns that could be explained by other factors as well, were attributed to the other factors. The outcomes of the combined analysis over regions were comparable to those of the separate regions. However, the effects of enforcement or campaigns were still not significant. This research shows that it is important to use rather long time-series to prove effects, especially if the number of accidents per region is low. The major aim of the study was to see if the multivariate time-series models could be applied successfully to several series of traffic safety data as are expected to result from the Dutch experiment. From a technical point of view the outcomes were very promising. The authors hope that their experiences may stimulate others to use (multivariate) structural time-series analysis for similar research problems.