Spatio-Temporal Modelling, Analysis and Forecasting of Road Traffic Accidents in Oman

Road traffic accident (RTA) is one of the most significant reasons for deaths and disabilities globally. Although there have been many attempts over the years, a number of insights yet need to be investigated, which would provide a better understanding of road traffic accidents (RTAs) in countries with a higher number of accidents. Many advanced mathematical models have been developed and applied to uncover a wide variety of issues of RTAs for many countries, for instance, accident frequencies, observed and unobserved factors, unobserved heterogeneity, endogeneity, etc. Oman is one of the wealthy middle eastern countries that experiences a very high volume of RTAs every year. However, little is known about the nature, variation, and causal factors for road traffic accidents in the Sultanate of Oman. Therefore, this study attempts to characterise the temporal, spatial, and combined space-time patterns of RTAs in Oman by mathematical modelling techniques. The RTAs data have been collected for the Sultanate of Oman for this study. Firstly a number of temporal models have been applied to discover the trend, other temporal components and forecast of the road traffic accidents and injuries (RTI) in Oman. The study found that the SARIMA model is the best temporal model with the highest goodness of fit for RTA and RTI. The higher number of RTAs occurred in June-August period, and the peak is in July during the summer holiday. Similarly, the months of May, June and July are expected to have the highest number of RTIs in Oman forecasted by the SARIMA model. Secondly, a variety of spatial models including spatial lag model (SLM), spatial error model (SEM) and geographically weighted Poisson regression (GWPR) model have been applied to capture spatial effects and influencing factors for the RTAs in Oman. One of the significant findings to emerge from this study is that the spatial lag model (SLM) outperformed the spatial error model (SEM) due to the best values in diagnostic indicators. However, the spatial variations in these models (SLM and SEM) are taken into account only through the spatial error structure. Another type of spatial modelling approach that provides a set of local models obtained by calibrating multiple geographical entities simultaneously is the geographically weighted Poisson regression (GWPR).

The main challenge with the GWPR models is to set appropriate kernel function to give weights for each neighbouring point during the model calibration. Likelihood function, parameter estimation and model selection criteria have been shown in details and model formulation has been applied to the RTAs data in Oman. A GWPR model has been developed for five different kernel weighting functions: box-car, bi-square, tri-cube, exponential and Gaussian weighted function. The study found that GWPR models can substantially capture the heterogeneity of the spatial factors over the regions or spatial units. The crucial finding to emerge from this study is that GWPR model with exponential kernel function and adaptive bandwidth is the most suitable for modelling, fitting and analysing RTA data in Oman. Finally, a combined approach of spatio-temporal modelling have been investigated, which would capture the temporal and spatial effects simultaneously. Spatio-temporal modelling is taken into account of both spatial and temporal correlations dependencies. The main challenge is building up the spatio-temporal model consisting of three components: choosing the perfect techniques to parts for space effects, time effects and the interaction between space and time effects. Spatio-temporal models have been built based on a Bayesian hierarchical framework to generate stable estimation that provides good characteristics to benefits even with low counts of RTAs. Integrated nested Laplace approximation (INLA) is fully Bayes tools used for fitting models, estimating fixed, random or posterior parameters. One major challenge is to model the interaction between spatial effects and temporal effects for count data models. This study develops Bayesian hierarchical spatio-temporal interaction models taking Oman RTA data. This study founds that unseparated spatio-temporal models are playing essential roles to capture unobserved heterogeneity for various factors. The most exciting finding is that the best performance pf a spatio-temporal model is the spatio-temporal interaction typeII model. Spatio-temporal interaction typeII acts by interacting between unstructured spatial effects and structured temporal effects. It suggests considering the second-order random walk (RW2) and Leroux CAR (LCAR) model into the spatio-temporal interaction model with a longer time scale and multivariate levels.