Abstract. Causal discoveries are difficult to obtain in the presence of high-dimensional observational datasets. We begin by showcasing how to utilize additional information on covariates to obtain confidence sets of average treatment effect and treatment effect size of a genotype mutation on HIV drug resistance. We then highlight a new approach for dynamic treatment discovery in the presence of “poor” data – we do not require propensity score estimation and yet provide treatment effect confidence sets that account for the complete individual history. Lastly, we showcase a new method for directed causal “parent-child” discoveries in the presence of both interventional as well as observational data. We work with Yeast genetic network under wild-type and gene-knockout interventional data. If time permits we will showcase some new results for censored, EHR data.
Joint work with my Ph.D. students: Jue Hou, Denise Rava, Davide Viviano, Yuqian Zhang, and my long-term collaborators: prof. Yinchu Zhu and Ronghui Xu.