Analysis of accident potential at unsignalised urban junctions in Ghana

Abstract

Until now the unsafety or accident potential of road locations has been assessed usually solely from their accident history. But this approach has been criticised as fundamentally misleading and inaccurate by many authors. Safety evaluation, particularly in Ghana, is also still restricted to isolated blackspot analysis. Because of its site-specific nature, this represents a less efficient use of scarce resources whilst there is generally a dearth of knowledge regarding the accident potential associated with various road locations and features. Based on a case-study of unsignalised urban junctions in Ghana, this dissertation presents the Empirical Bayesian procedure for the estimation of site-specific accident potential as a superior alternative that automatically addresses the shortcomings inherent in the sole use of recorded accident counts. The refined estimate is produced from a uniquely weighted combination of the recorded accident counts of the particular site and the expected accidents for sites with similar characteristics as the one under study. A unique feature of this study is the demonstration of the framework for integrating comprehensive accident model predictions with the Empirical Bayesian procedure, to improve further upon the quality of the estimates and extend the applicability of the procedure to relatively smaller reference populations. Despite the acknowledged advantages of this approach, little work has been done in this direction, largely due to the absence of appropriate prediction models, particularly for traffic conditions in developing countries. Thus, a key part of this study has been the development of accident prediction models for unsignalised T- and X -junctions. The models were of two types, namely, the coarse or flow-based, which included only traffic exposure functions as the explanatory variable, and the full or comprehensive models, containing both exposure functions and other significant road and traffic variables. Separate models have been developed to predict the three-year frequency of total accidents, injury accidents and five other types of accidents defined by the primary collision types involved. The unique ability of the Bayesian analysis to treat accident potential at a site as a random variable was utilised to outline new probabilistic criteria for accident blackspot identification. Existing criteria for ranking accident blackspots for treatment have also been revised and an improved criterion called the Amended Potential Accident Reduction proposed.