An interrupted time series analysis of the differential impact of New Jersey’s BAC legislation on driver and passenger crash fatalities

Abstract

Most research, drawing on deterrence and rational choice models of social action, examines the effects of reductions of blood alcohol concentration [BAC] limits to secure drunk driving convictions on the total volume of crash fatalities. This paper extends this work by investigating the impact of New Jersey’s BAC legislation on total and disaggregated crash fatalities. The results from the interrupted times series analyses show that reducing the BAC limit to 0.8 has no effect on total or driver fatalities, but has a negative and lasting effect on passenger fatalities. The implications of these findings for future research are discussed.

Keywords:

• Deterrence
• interrupted times series
• BAC limits

Notes

1. Proponents of the use of rate measures contend that is necessary to deflate raw counts to remove the potentially spurious effects of structured opportunities on the magnitude of an outcome series. Within the present context, the fear is that changes in driving-related fatalities are not affected by changes in BAC limits, but rather by changes in the number of crashes. In contrast, proponents of the use of counts note that ARIMA pre-whitening procedures effectively remove the confounding of unmeasured variables (including the volume of crashes) that accumulate over time (see our discussion ARIMA model building in the body of this paper). Moreover, the ARIMA analysis of raw counts has the advantage of not requiring one to select, a priori, one series in lieu of another to deflate an outcome series.

The above notwithstanding, I suspect that his point of contention is probably less important than it may seem. Based on my understanding of ARIMA interrupted time series models and prior research, I anticipate that the analysis of rates and raw counts will generate similar findings (Granger 1980; Jenkins 1979).

2. The selection of each of the final transfer models is based on the interpretation of model statistics and various diagnostic tests. Chief among them are the Q statistic, which is used to evaluate the overall fit of the model and the parameter estimates (which are both reported in Tables 1 and 2). Coincidentally, for each series, the zero-order transfer function provides the best fit to the data. Hence, we report the results from these models, and not those from the first-order or pulse transfer function equations, in Tables 1 and 2.