Sign the Resolution

Contents | Feedback | Search

DRCNet Home | Join DRCNet

DRCNet Library | Schaffer
Library | Drugs and Driving

** School of Justice Administration, Griffith University, Queensland 4111, Australia*

*** Data Analysis Australia*

Random breath testing (RBT) was introduced in New South Wales on December 17, 1982. Although there have been many studies of the impact of RBT, no comprehensive time-series analysis of accidents has ever been published. In this paper, daily accident data for the period January 1976 to December 1992 are analysed for the impact of RBT, controlling for weather information, road usage/economic indicators, "time" factors, and the .05 legislation introduced in December 1980.

The initial effect of RBT on total fatal accidents was extremely marked, with a drop of 48%, an effect which was however limited to about two months' duration. The long-term effects of RBT were revealed most clearly for single-vehicle night-time accidents, for which the initial impact was a 24% decline that was sustained for nearly five years. In addition, there were an estimated 12% fewer such accidents for every 1,000 drivers tested, an effect which intensified as levels of RBT enforcement were increased from 1987 onwards. As predicted, there was almost no discernible impact of RBT on non-alcohol related accidents. The analyses strongly support the theoretical model of deterrence developed by Homel.

The theory of deterrence through criminal law enforcement has determined the major system of public responsibility for road safety in Australia. It is argued that if punishment for drinking and driving in a jurisdiction is swift, sure, and tough, the rate of occurrence of the offence will be correspondingly low. Historically the police have sought to deter offenders through a policy of targeted enforcement in which arrest rates are maximised, often as a result of intense blitzes at weekends and during holiday periods. In major reviews of the effectiveness of such measures in many jurisdictions, and also of new laws such as the British Road Safety Act 1967 (which introduced breath testing equipment to Britain), Ross (1982) concluded that at best only temporary reductions in road accidents can be achieved. For example, the effects of the British Act dissipated in about four years, as people realised that their chances of getting caught were not as high as they had thought when they were exposed initially to the intense media publicity about the new law. Ross also concluded that measures which simply increased the severity of punishments without also increasing the perceived chances of apprehension do not achieve deterrent effects even in the short term.

On the basis of Ross' research, Homel (1988) and others argued that a system of well publicised and intensively enforced random testing would be likely to prevent most of the decay in deterrent effects documented in the literature. The reason for this is that if any motorist at any time can be breath tested, the potential drinking driver, no matter how skilled he or she believes they are in avoiding detection when over the limit, will have cause to think twice before actually committing the offence. Arguments such as these led on December 17, 1982 to the introduction of random breath testing in New South Wales. RBT had been introduced six years earlier in Victoria, although in a low-key fashion, and was introduced in all other states and territories in 1983 and later years. The focus in this paper is on the NSW experience.

The RBT law in NSW was very extensively advertised and vigorously enforced, with about a million tests in the first year out of a licensed driving population of three million. In later years, police improved on this ratio of one to three. Indeed, RBT in NSW must rank as one of the best enforced and most widely publicised laws ever enacted (Homel 1988). There was an instantaneous 22% decline in total fatal crashes, and a drop of about 36% in alcohol-related fatal crashes, relative to the previous 3 years (Homel, Carseldine and Kearns 1988).

In order to arrive at a more accurate estimate of the causal impact of RBT, it is necessary to control in a time series analysis for other factors which are known to influence accident rates. These include weather details, particularly rainfall; time factors such as the day of the week, the season of the year, and holiday periods; road usage indicators, such as petrol production and driver licences; economic indicators, such as gross domestic product (GDP), expenditure on alcohol, disposable income and unemployment levels; and levels of road safety publicity, particularly related to drink-driving. Homel, Henstridge and McKay (1995) carried out such an analysis for NSW, using daily accident data from January 1 1976 to December 31 1992 (17 years of data). This paper summarises aspects of that analysis.

Fatal accidents or accidents which led to serious injury resulting in the hospitalisation of at least one person involved in the accident, were considered. Such accidents will be referred to as "serious" accidents in this paper. The accidents targeted by RBT legislation are alcohol-related, and ideally numbers of serious alcohol-related accidents should be analysed. Unfortunately information on blood alcohol levels of drivers is often either unreliable or not available, especially for accidents which occurred in the earlier years of the study. A solution is to look at the types of accidents thought to have a high probability of being alcohol-related, such as single-vehicle night-time accidents.

Of most importance (and one of the motivations for doing the analyses on a daily basis) was rainfall. Minimum and maximum daily temperatures were also included as well as a range of other conditions such as whether fog, a thunderstorm or strong winds had occurred on the day. Road usage and economic indicators are summarised in the Table 1.

**Table 1**

**Road Usage and Economic Indicator Control Variables**

Variable | Available | For | Measurement Units |
---|---|---|---|

Petrol sales | monthly | states | Megalitres |

Petrol production | monthly | Aust. | Megalitres |

Vehicle registrations | annual | states | x 1000 vehicle |

Drivers licences | annual | states | x 1000 licences |

Unemployment rate | monthly | states | Percentage |

GDP figure | quarterly | Aust. | $million, 1984-85 prices |

Household disposable income | quarterly | Aust. | $million, 1989-90 prices |

Private alcohol expenditure | quarterly | Aust. | $million, 1979-80 prices |

The first four variables are indicators of the numbers of kilometres travelled and of the number of vehicles on the road. The last four variables are indicators of the importance of economic conditions on accident numbers. In order to convert the above data from monthly, quarterly or annual to daily data, linear interpolation was used. In addition all series other than the annual data, were seasonally adjusted. The GDP figures are at average 1984-85 prices, household disposable income at average 1989-90 prices and private alcohol expenditure at 1979-80 prices. Thus all variables in dollar terms are at constant prices.

Seasonal terms, day of week and type of day were controlled. Different numbers of accidents occur at different times of the year and on different days of the week. Typically most accidents occur on Friday and Saturday nights. Furthermore differences may occur between public holidays/long weekends, school holiday periods and other times of the year. The variable for the type of day takes account of differences between different times of the year. The day before a public holiday or long weekend was included in the holiday period since it is likely that many people will travel on these days.

In theory the effects of publicity campaigns can be tested by comparing accident rates in periods where such campaigns were run with periods when there was no publicity. In practice this process is not practical since in most states the tendency is to run media campaigns in periods when high rates of accidents occur, for example at Christmas and Easter. Variables related to publicity were therefore not included in the analysis.

Daily counts of accidents were analysed in preference to aggregated data such as monthly counts. Control variables can easily be taken into account in analyses using daily data and it is also possible to test for lag periods of different lengths for the effects of explanatory variables, in particular enforcement statistics. This cannot be done effectively by analysing monthly accident statistics. Provided that accident numbers are analysed (not numbers of fatalities, injured persons or vehicles) each unit being counted is in some sense independent, so a Poisson distribution model is appropriate. Furthermore since the rates of accidents are being modelled, a multiplicative structure is appropriate, meaning that factors affect the rates by proportionally increasing or decreasing them. These two features of the data suggest the use of log-linear models and GLIM-type software - a special case of generalised linear models (Nelder & Wedderburn, 1971). For the current problem there is the additional time series aspect. Generalised linear models do not extend naturally to include auto-regression effects common with time series. They do provide a set of standardised residuals which can be used to check for the presence of autocorrelation but to include such correlation in a theoretical model is not practical.

The approach taken was to use explanatory variables such as the weather and seasonal trends to remove autocorrelation from the residuals. The model fitting was done in a custom modification version of the time series package TSA-32 which implements algorithms identical to that in the program GLIM. The use of TSA-32 meant that it was possible to monitor the time series aspects. In particular residual autocorrelations were reviewed for every fitted model. The precise means of modelling the effect of RBT was only derived after some investigation of alternatives. Ideally the method would be to have a time series model with an autoregressive filter for the RBT input, which would allow for the magnitude and duration of the RBT effect to be estimated directly. However when using daily data this method is numerically unstable since the parameter for the autoregressive process would be very close to unity, the boundary of what is allowable. In addition standard time series modelling software for fitting autoregressive transfer function models is not designed for Poisson variables.

The procedure adopted was to generate an appropriately shaped exponentially decaying impulse function with specific time constant and to then fit this in the Poisson regression model. The value of the time constant was changed over a range from one month to ten years and the value which minimised the deviance in the fitted model was used as the estimate of the time constant. In addition to the exponentially decaying impulse, a step function was included in the model corresponding to the time of introduction of RBT. This was included to model the permanent effect of RBT. The effects of other legislation such as the lowering of the legal BAC from .08 to .05 g/100ml in December 1980 were also investigated.

In addition to accident statistics, monthly enforcement statistics were obtained. Since this information is obviously only available after the introduction of RBT, separate analyses were carried out on the post RBT data to determine the effects of levels of enforcement on road accidents. First, models which included terms corresponding to levels of enforcement were fitted. The significance of the terms indicated the importance of the individual measures of enforcement. Different lag times were utilised and an optimum time identified where applicable. Secondly a deviance analysis was carried out to determine the overall importance of the enforcement levels. Such an analysis involves comparing the deviance obtained from the model which incudes the enforcement terms with the deviance for the model which does not. This difference in deviances can be assumed to have a chi-squared distribution with degrees of freedom equal to the difference in degrees of freedom for the two models. Thus there is a means of testing whether enforcement levels have a significant effect on accident numbers.

All the explanatory variables were included in the models initially and then some were excluded until a suitable model was obtained. "Suitable" refers to a model thought to contain sufficient, but no more than necessary, explanatory variables to control for factors other than RBT that may influence accident numbers. Some of the variables are interchangeable in that they are thought to represent the same underlying factor. For example, numbers of licensed drivers and numbers of registered vehicles are highly correlated and thus only one is necessary in the model.

In general RBT was found to have resulted in a significant initial drop in accident numbers. Estimation of the length of the period that the effect was sustained was extremely difficult. In most instances there was a definite short-term effect and some evidence of a longer term effect. There was a definite decrease in accident numbers when RBT enforcement was at high levels, controlling for other factors.

The initial impact of the introduction of RBT translated into an estimated drop of 19.29% in accident numbers. This lowering of accident rates appeared to last for a 15 month period before reaching 5% of the initial effect. Levels of enforcement impacted significantly on daily accident numbers with an estimated 6.2% fewer accidents per 1000 vehicles stopped. Given that on average 18 accidents occurred per day from the beginning of 1988, this translates to one accident per day or 750 for the period of the study. The average period that the effect of enforcement levels was sustained was about 200 days or 6.5 months.

The introduction of RBT is estimated to have resulted in a 47.98% drop in fatal accidents. However this was a short-term effect of 4.5 months duration, corresponding to 194 fatal accidents prevented. It was not possible to show that enforcement levels had an influence on fatal accident numbers, due in part to a lack of power reflecting low numbers of fatal accidents.

The overall effect of RBT was highly significant and the initial effect was a drop of 24% in such accidents. This lowering of accident numbers was sustained in the long term (the exponential term in the model) and RBT had an ongoing effect as measured by enforcement statistics. It is estimated that 12% fewer accidents occurred per day per 1000 vehicles stopped, corresponding to between 639 and 1413 accidents prevented each year between 1983 and 1992. The period of awareness of RBT enforcement was an average of 15 months.

These were measured by daytime vehicle-vehicle accidents occurring between 9 am and 3 pm on school days. As predicted, the overall impact of RBT was not significant when controlled for other factors (chi-square = 3.99 with 2 df), although the step function was significant on its own (z = 2.88; p = .0039). However, the initial impact on accidents was only .14%, suggesting essentially no effect.

For all three types of accidents the lowering of the blood alcohol limit in December 1980 had a significant negative effect on accident numbers. The length that this effect was sustained was not tested for and would be difficult to ascertain given the introduction of RBT two years later (Homel 1994). As expected the occurrence of rain contributed significantly to accident numbers. Seasonal trends were evident especially in the case of single-vehicle night-time accidents. However fatal accidents were not as affected by the time of the year as other accidents were. Most accidents occurred over weekends, peaking on Saturday nights. Another expected result was that there were significantly more accidents on public holidays than at other times of the year.

The interpretation of the individual economic/road usage variables was problematic since they all contained trends over time which influenced the estimation of their coefficients. In particular the coefficients for GDP and the step function for RBT were found to be correlated. To exclude the GDP figure from the analyses would falsely inflate the RBT effect. Thus GDP was retained in the models and the effect of RBT was underestimated to some extent. Unemployment rates are thought to be a measure of the influence of economic conditions on accident numbers. The analyses reinforced that belief, indicating that significantly fewer accidents occur in periods of relatively high unemployment. For each group of accidents the contribution to the model of petrol sales was evaluated. Since it was not of importance after the effect of the other variables had been taken into account and to include it would reduce the size of the data set, it was excluded from the models. Numbers of drivers' licences and numbers of vehicle registrations are highly correlated and it was unnecessary to include both in the models. Thus driver's licences were used. They were found to have a positive influence on accident numbers.

The analyses confirm previous research which has found that the introduction of RBT in NSW coincided with a marked and permanent decline in accidents. The present analysis is more rigorous than previous studies, both in terms of the range of control factors incorporated in the models and in terms of the sophistication of the statistical methods used.

Depending on the accident measure used, there was clear evidence of a marked initial impact of RBT which decayed to some extent after some months or years. The decay was most marked for the measures that included non-alcohol related accidents. Single vehicle night time accidents, the most accurate measure of alcohol-related accidents, showed less evidence of decay, and in addition the enforcement term indicated an intensification of the impact of RBT, particularly after the end of 1987 when total tests conducted increased markedly. Virtually no impact of RBT could be detected for day time accidents on school days, which constitutes a control series.

The analyses are consistent with Homel's (1988) model of deterrence, which emphasised the instability of deterrent effects and the need for ongoing high visibility enforcement which effectively manages psychological uncertainty so that perceived probabilities of apprehension are maintained at high levels.

This research was supported financially by the Federal Office of Road Safety.

Homel, R. (1988). Policing and punishing the drinking driver: A study of general and specific deterrence. New York: Springer-Verlag.

Homel, R. (1994). Drink-driving law enforcement and the legal blood alcohol limit in New South Wales. Accident Analysis and Prevention, 26, 147-155.

Homel, R., Carseldine, D. & Kearns, I. (1988). Drink-driving countermeasures in Australia. Alcohol, Drugs and Driving, 4(2), 113-144.

Homel, R., Henstridge, D. & McKay, P. (1995). The optimisation of random breath testing. Report to the Federal Office of Road Safety.

Ross, H. L. (1982). Deterring the Drinking Driver: Legal Policy and Social Control. Lexington, MA: Lexington Books.