Edinburgh Napier University

  In contemporary accident research, unobserved heterogeneity is recognized as one of the most challenging statistical modeling misspecification issues.Unobserved heterogeneity (i.e.,unobserved characteristics varying systematically across observations or groups of observations)typically results in biased and inefficient parameter estimates, incorrect predictions, andininconsistent statistical inferences. Theimplications of unobserved heterogeneity in model estimationhave been addressed –to some extent –over the last decadesthrough the development of several statistical and econometric methods. However, limitations of the newly emerged methodological advances, in combination with inherent restrictions of the highly-dimensional accident datasets, impose significant barriers on the way various aspects and patterns of unobserved heterogeneity are addressed. This dissertationaims to addressunder-explored unobserved heterogeneity structures, by extending the formulations of popularstatistical approaches that –by definition –account for unobserved variations across the sample population: the random parameters and latent class modeling approaches. For this purpose, the simultaneous effect of non-time-varying (stationary) and time-varying factors is investigated within two distinct accident occurrence and injury-severity settings.

In the context of the accident occurrence analysis, a binary random parameters (mixed) logit framework is employed in an effort to identify pre-crash stationary and time-varyingfactors of accident occurrence, while,at the same time, accounting for unobserved heterogeneityand panel effects. To capture unobserved heterogeneity interactions between time-varying and stationary factors, the employed framework allows for the estimation of correlated random parameters.

In the context of the accident injury-severity analysis, to identify thedeterminants of injury-severity outcomes, the correlated random parameters framework isalsoemployedwithin an ordered probit modeling setting. In addition, to account for threshold heterogeneity, arising from the fixed nature of the thresholds of the traditional ordered probability models, and simultaneously for possible systematic variations of unobserved factors across the accident observations, a random thresholds hierarchical ordered probit approach with random parameters is developed and applied.Furthermore, to investigate how thedata grouping of contemporaryaccident datasets affects the patterns of unobserved heterogeneity, highway segment-and accident-based latent class ordered probit models of accident injury-severities are estimated. The employed latent class approaches also accommodate heterogeneity stemming from theprobabilistic assignment of the highway segments or accident observations in the latent classes, by incorporating class probability functions in model estimation.

The estimation of the accident occurrence and injury-severity models is based on accident data collectedbetween 2011 and 2013 from urban and rural highway segments in the state of Washington.The findings of the dissertation as well as the estimated model specifications can be effectively utilized asinput andalgorithmic modules, respectively,for newly emerged research tools aiming to assess the accident risk of highway segments over time.