To run a "demerit point" program, routinely available information about drivers is used to identify those who are most likely to have an accident in the near future. On the basis of a four-year recordfor a large sample of Ontario drivers, several tools for the identification of such drivers were examined in order to ascertain how they performed. Each driver is thought to have an expected number of accidents, 'm'. In a group of drivers with common traits (such as age, gender, record of convictions and accidents) the ms have a mean E(m) and a variance VAR(m). Estimates of E(m) and VAR(m) for all combinations of traits can be obtained within the framework of a multivariate statistical model. The same estimates can then be used to judge how well a model identifies drivers who have a large m. In such a multivariate model it is important to use data about previous accidents and convictions. However, the accuracy with which the 'm' of a driver can be estimated is not improved much by distinguishing between offence type or between accidents as being "at fault" or "not at fault". Without much loss in estimation accuracy, a weight maybe attached: (1) to a conviction, and (2) to an accident. Model performance is described in tangible terms: how many accidents are recorded by the drivers identified by a model, what proportion of identified drivers are "false positives", how many drivers with high 'm' remain unidentified. It is concluded that by using a multivariate statistical model a substantially better result can be obtained than by using a demerit point scheme in which points are assigned to offenses on the basis of their perceived seriousness. However, even when the best model is used to identify a large group of drivers, many will be falsepositives.
Samenvatting