Multilevel models in traffic safety research: An investigation and illustration of a flexible solution to improve upon classical statistical analysis techniques

Abstract

This Ph.D. thesis deals with the issue of correct versus incorrect usage of statistics. More precisely, the objective is to demonstrate the applicability, usefulness and added value of multilevel models in the field of traffic safety. Multilevel models, also known as mixed models or random effects models are a family of techniques that can be considered a flexible solution to overcome some limitations of classical analysis techniques. To reach this objective, several sub-goals have been formulated. First, multilevel models are described using an intuitive and a mathematical approach. Second, the use of multilevel models is justified by explaining and illustrating the consequences of not using such models when it is necessary, or at least advisable to do so. It is explained that multilevel models are particularly useful for dealing with hierarchical or nested data, a type of data that is common across scientific disciplines including traffic sciences. Due to the nature of hierarchical data careful consideration is required with respect to the dependence of nested observations and context. Given the dependence of nested observations, ignoring the hierarchical nature of the data will lead to an underestimation of standard errors and an increased level of committing type I errors. Ignoring contextual aspects of hierarchical data typically results in an impoverished conceptualization of the research problem. The third sub-goal of this thesis is to apply multilevel models to a variety of traffic safety research topics and to illustrate their added value. Three case-studies are used to achieve this goal; they involve a case study on drinking driving, one on sleepiness among night-time drivers and one on the effectiveness of graduated driver licensing programs. In the first two case-studies a two-level logistic regression model is used to analyze the data; in the third case study a multilevel meta-regression analysis is carried out using both a 'frequentist' and a 'Bayesian' approach. It is demonstrated that multilevel models are particularly elegant and productive in generating new knowledge about each of the traffic safety issues of interest and that not using such models can lead to faulty conclusions. Finally, the last sub-goal is to define a research agenda. Based on the findings from the case-studies conclusions are formulated regarding the applicability, usefulness and added value of multilevel models in traffic safety research. Special attention is given to practical implications of the findings and their social relevance, both with respect to the applied methods and the actual findings from the analyses. As such, a research agenda for further studying these road safety issues is created.