Integrating Information From Prior Research Into A Before-After Road Safety Evaluation Through Bayesian Approach And Data Sampling
thesisposted on 08.06.2021, 11:28 by Chen Yongsheng
Before-after road safety evaluation (B/A) to measure safety treatment effect is a key mission in road safety management, and has fueled considerable research. However, previous research in this area has been overwhelmingly dedicated to safety model estimation with less emphasis on other methodological issues. As a result, there continues to be uncertainty in the validity of treatment effect estimates. This study seeks, with innovative paradigms, a systematic solution by solidifying methodologies for every essential step of a thorough B/A process to secure its ultimate validity. Methodologies of data sampling and processing, and before and after model development, both vital procedures that have been historically neglected, are investigated. A pre-test data sampling approach to select reference groups is established in the context of B/A application. A post-assignment propensity score matching method is developed in order to further eliminate statistical bias while the treatment effect indicator – collision reduction ratio (CRR) – is being estimated. Rather than focus on single safety model development as is common in traffic safety research, this study seeks all viable knowledge by employing various safety measures including collision and safety surrogates, by embedding several adaptable random distributions, by fitting models through both "Frequentist" and "Bayesian" approaches, and by exploring a variety of model forms and components. Accordingly, the output of this study is not a "best" single model, but rather an amalgamation of diversified models. The diversity is shown to be attractive in terms of information conveyed, especially for the B/A process. Finally, this study succeeds in finding a methodology to integrate all of the diverse knowledge sources. The Bayesian Model Averaging (BMA) method is investigated and developed to integrate a variety of statistical significant models without exclusion, in forging a unified model. All methodologies explored and developed in this study are essential to secure the validity of the B/A process. As important, they are substantially connected to each other. Should one method be deficient, the remaining steps cannot guarantee validity of B/A process. As a whole, these methodologies, if properly developed and applied, constitute a logical chain to estimate treatment effect with minimal errors and high validity.