Statistical modelling and analysis of accident data

While, David Thomas (1991) Statistical modelling and analysis of accident data. University of Southampton, Doctoral Thesis.

Record type: Thesis (Doctoral)

Abstract

This project investigates different methods of modelling reported accident data in order to identify and quantify the effect of variables influencing accident rates. Strong evidence of differential reporting across the population is identified.

The full reported accident distribution is modelled directly using the traditional accident modelling probability distributions, namely the Poisson and Negative Binomial distributions. The male and female populations arc split into non-overlapping sub-populations with respect to a single variable at a time and the reported accident distribution from each sub-population is modelled directly using the traditional accident modelling probability distributions.

It is found that the traditional models fail to adequately lit the data. A variety of alternative models is investigated and generalised models are devised which take into account accident misreporting. It is shown that the best description of the data is given by a model formed from generalising a Negative Binomial model to include the misreporting of two accidents as one accident. An alternative approach to avoid drawing inaccurate conclusions from the study is to focus the analysis on a large, more reliable subset of the data. This is achieved through the definition of a non-superficial accident distribution.

Regression models are investigated using GLIM and logistic regression modelling techniques. Different degrees of accident seriousness are investigated which result in the identification of different significant influencing factors. This provides further evidence that sections of the reported accident data are so unreliable that they lead to misleading conclusions.

Text

309530.pdf - Version of Record

Available under License University of Southampton Thesis Licence.

Download (16MB)