Use of a Bayesian Changepoint Model to Estimate Effects of a Graduated Driver's Licensing Program

Michael R. Elliott ； Jean T. Shope

《Journal of Data Science》 1卷1期 (2003/01) Pp. 43-63

https://doi.org/10.6339/JDS.2003.01(1).103

並列摘要

In April 1997 the US state of Michigan implemented a graduated licensing program for novice drivers under the age of 18 that ensures that they gain experience and maturity under conditions of low risk before progressing to more risky driving situations. Since there is no reasonable control group of young Michigan drivers not exposed to graduated licensing during this period, the extent to which observed declines in crash rates can be attributed to graduated licensing versus other unobserved changes in crash reporting or driving behavior is important. We assemble a Bayesian changepoint model to assess the probability that changes in crash rate trends among 16-year Michigan drivers can be plausibly linked to the introduction of graduated licensing and, if it can, to make inference about graduated driver licensing effects that take into account the uncertainty in when these effects began and ended, and whether or not a ”rebound” in crash rates occurs afterward. We show that, while there is a moderate degree of sensitivity to the choice of prior distributions for changepoints and rate slopes in determining the number of changepoints present in the crash trends, inference about whether GDL effects are present and their degree are relatively insensitive to prior choice. This analysis suggests that the decline in crash rates among 16-year old Michigan drivers observed between 1996 and 1998 can be reasonably attributed to graduated licensing for all crashes combined, but that observed changes in single-vehicle and especially nighttime crash rates might have been part of longer-term trends among this age group.

並列關鍵字

Linear splines ； model averaging ； model selection ； prior sensitivity