Semi-parametric Panel Count Model for Drug Safety Evaluation

Abstract

In this dissertation, we study issues related to drug safety evaluation. In the first part, we focus on the adverse event (AE) signal detection in post-market surveillance systems. We derive a semi-parametric panel count model to search for safety signals by accounting for background noise, issues associated with such as zero-inflated data count, and covariates information. We develop an estimating procedure with Expectation-Maximization (EM) algorithm to estimate the model. In each M-step, the maximization of the non-parametric component is reformulated as an optimization problem in isotonic regression. The strong consistency and asymptotic distributions of the model estimators are formally derived. We conduct simulation studies to evaluate the finite sample performance of the method proposed and to demonstrate the advantages of the proposed method in signal detection with high power for signal detection, high specificity, and sensitivity. The proposed method is applied to WHO VigiBase System and FAERS with several relevant covariates yielding new signals not found with standard approaches and reduced false positive rates. In the second part, we develop the doubly robust estimator in the panel count model to improve the method of inferring causal effects of medicines/vaccines on adverse events (AE) from data with Poisson outcome. Simulation studies demonstrate its robustness with respect to misspecifications of the propensity score or outcome model.