Limits of explainable variation in road crashes

Monash University Accident Research Centre - Technical Report No. 001/98

Authors:  M. Shtifelman, M. Cameron, S. Newstead & G. Avar

Abstract

Recent research has illustrated and evaluated the dependency of the monthly variation in serious casualty crashes in Victoria on a number of major factors.  These factors include the RBT "booze bus" and speed camera programs plus their supporting publicity, and economic, social and demographic factors.

Part of the variation in accident counts can be explained by pure randomness because by their nature accident counts are subject to random variation (chance).

This study aimed to investigate the theoretical maximum proportion of explainable variation in road trauma trends and to help make more definite conclusions about the explanatory factors that influence the number of fatal and serious crashes.

Simulation studies were used to examine how the estimated unexplained or random variation component of a regression model varies with random variation introduced into the original process.  It is commonly assumed that road crashes occur independently according to a Poisson distribution, hence the simulation studies focused on Poisson random variation.

The optimal fit of the model and the likely variation of the optimal fit was found under a number of specified frameworks.  This will allow understanding of when models of road trauma have been under-fitted (i.e. too few explanatory factors have been included to adequately describe the observed series), or in some cases over-fitted (i.e. too many explanatory factors have been included).