Performance of Multiple Linear Regression (MLR) and Artificial Neural Network (ANN) in Modeling and Prediction of Accident Data

Abstract

The alarming rate of road traffic accident in the country (Nigeria) is
among the most worrisome problems currently facing the nation.
Sadly, Nigeria has earned the unenviable distinction of consistently
leading all the nations of the world in high road traffic accident and
high fatality rate. This study conducts a comprehensive evaluation
of selected expert systems such as multiple linear regression and
artificial neural network for the modelling and prediction of road
accident. The study area is Ugbowo-Lagos Road in Benin City, Edo
State Nigeria. A reconnaissance survey was done first to ascertain
the geometric characteristic of the road which include; the chainage,
the vertical and horizontal curve and the super elevation. Thereafter,
primary data which include road accident data was collected from
Federal Road Safety Office in Benin City. To investigate the
qualities of the primary data, basic preliminary analysis techniques,
namely; outlier detection, homogeneity test, test of normality and
autocorrelation test were conducted while modelling and prediction
of road accident was done with the aid of multiple linear regression
and artificial neural network. From the geometric characteristic of
the road under study, it was observed that for a chainage of 11.5 to
13.0km, the vertical curve was 12.4% while the super elevation was
4.3%. Calculated Cronbach alpha value of 0.900 as observed in the
reliability test revealed that the data are reliable and the computed
goodness of fit statistics of reliability gave a maximum Guttman
coefficient of 88.10% which further confirm the reliability of the
data used. With a computed p-value greater than 0.05 for all the
independent variables, the null hypothesis of the Dixon test was
accepted and it was concluded that the accident data obtained from
FRSC is devoid of outliers. In addition, with a centered VIF < 10, it
was concluded that there is the absence of multicollinearity between
the dependent and independent variables. With a computed
coefficient of determination (R2) value of 0.9265, artificial neural
network (ANN) was acclaimed better. Road accident prediction
model compare to multiple linear regression model (MLRM) with a
computed R2 value of 0.0617.