Dissertations & Theses (Open Access)

Date of Award


Degree Name

Doctor of Philosophy (PhD)


Ruosha Li

Second Advisor

Shouhao Zhou

Third Advisor

Christine B. Peterson


My dissertation builds on a systematic review of 125 clinical trials reporting on treatment-related adverse events (AEs) associated with PD-1/PD-L1 inhibitors published from 2010 to 2018. The motivating dataset contained the following study-level components extracted from each publication: trial name, number of treated patients, selected immunotherapy drug, dosing schedule, cancer type, number of AEs within each category, and the pre-specified criteria for AE reporting. The number of AEs were reported based upon all-grade (Grade 1-5) and Grade 3 or higher (Grade 3-5) severity. My overall objective was to increase our understanding of the toxicity profiles of five most common cancer immunotherapy drugs, and to evaluate AE incidence across subgroups in a meta-analysis setting. However, for assessing drug safety in clinical trials, a common challenge is that many published clinical studies do not report rare AEs. In particular, if the number of AEs observed is lower than a pre-specified cutoff value, these events may not always be reported in the publication (i.e., they are censored). My doctoral dissertation research, thus, proposes an innovative statistical methodology for effectively handling censored rare AEs in the context of meta-analysis of immunotherapy trials. First, by deriving exact inference and robust estimates for the missing not at random data, we proposed a Bayesian multilevel regression model in the coarsened data framework to accommodate censored rare event data. We also demonstrated that if the censored information was ignored, the incidence probability of AEs would be overestimated. Second, to select the best Bayesian censored data model among a set of candidate models in the presence of complicated or high-dimensional features, we proposed an alternative strategy to implement Bayesian model selection for censored data analysis in Just Another Gibbs Sampling (JAGS). To generate deviance samples from a Bayesian model using JAGS, if censoring occurs, an existing function incorrectly calculates the value of deviance function because of the “wrong focus”, i.e., the incorrect likelihood computed on the basis of model specification in JAGS. Therefore, we proposed a strategy to establish a simultaneous way to calculate the true value of deviance function in JAGS. The alternative strategy could be generalized to model other types of data and be applied to many other disciplines. Third, we developed a sparse Bayesian selection model with prior specifications on meta-analysis of censored rare AEs to perform selection of pairwise interactions between various study-level factors. Because the toxicity profiles of immunotherapy drugs may not be explained comprehensively by main effects of study-level factors, we identified the high-risk group by considering two-way interactions that impact the outcome of interest. Through simulation studies, we demonstrated that the proposed interaction selection method outperforms others in prediction accuracy and interaction identification in the presence of missing outcome data. Lastly, we also applied the proposed method to our real-world motivating dataset. In sum, my dissertation work makes significant and innovative contributions to the field of applied statistics and cancer research.