CMStatistics 2018: Start Registration
View Submission - CMStatistics
Title: Quantifying the causal effect of speed cameras using two-stage Bayesian bootstrap Authors:  Prajamitra Bhuyan - Imperial College London (United Kingdom) [presenting]
Abstract: A causal doubly-robust (DR) approach combines propensity score (PS) and outcome regression (OR) models to give an average treatment effect estimator that is consistent under correct specification of either of the two component models. In this set-up, standard Bayesian methods are difficult to apply because restricted moment models do not imply fully specified likelihood functions. To avoid this difficulty, the existing methods are restricted to utilize full Bayesian features and involve frequentist estimates of the propensity score. As a result, these methods inherited some of the deficiencies involved in the frequentist DR approach under misspecification of component models and the estimate becomes biased. A two-stage Bayesian bootstrap approach is proposed which allows incorporation of prior information and uncertainty quantification associated with both OR and PS model. Simulations show that the approach performs well under various sources of misspecification of the outcome regression or propensity score models. The proposed method is used to quantify the effect of speed cameras on road traffic collisions in British cities.