OPTIMAL MODEL FOR ROAD ACCIDENT CASES IN GHANA (2017-2021)

Abstract

Poisson Generalised Linear Models (GLM) which is frequently used to model count response variable like the number of road traffic accident poses a challenge when the model involves too many predictors. This situation can lead to overfitting which is a serious problem in maximum likelihood. The purpose of this study is to explore and develop the optimal model for road accident casualties in Ghana. The study compared the performance of four Poisson Generalised linear models: Poisson regression, Zero Truncated Poisson regression, Negative Binomial Regression and Zero Truncated Negative Binomial Regression in estimating road accident casualties in Ghana with the help of R-software. The optimal model was selected with the help of a selection criteria (Akaike Information Criterion and Bayesian Information Criterion) based on known parameters. The best performing model was compared to its LASSO regularised linear model to determine which of them best select relevant features for estimating road accident casualties. The accident data used for the study was a secondary data on road accident cases compiled by Building and Road Research Institute from 2017 to 2021. The results indicate that, LASSO negative binomial regression model is more suitable than classical negative binomial regression model in estimating road accident casualties in Ghana with many predictors. Again, the type of weather, vehicle involve in an accident, location and age are major determinants of road accident casualties in Ghana. Thus, in building a generalised linear model, one should consider a regularisation technique to improve the model's performance and feature selection especially with many predictors.