Speaker: Guoliang Ma, PhD Candidate, Department of Statistics, Iowa State University

Title: Estimation of Treatment Effects Without Ignorability Using Observational Study

Abstract: Most causal inferences (CI) assume ignorable treatment assignments, i.e. the treatment selection probability depends on observable covariates only. However, real-world applications often reveal a dependency of the selection probability on potential outcomes, leading to incorrect treatment effect estimation if selection bias is disregarded. In non-ignorable scenarios, counterfactual outcomes are missing not at random (MNAR). Traditional propensity score-based (PS) estimators fail to handle such situations due to the unobservability of potential outcomes within the potential outcomes framework. To address this issue, we propose a new iterative evaluating-solving (ES) algorithm to impute potential outcomes via semiparametric quantile regression under the MNAR assumption. We thoroughly discuss model identification under MNAR and validate the proposed estimator through convergence theories, large sample properties, and simulation studies. Variance estimation is also provided. Real data application of the proposed method on the first eight cross-sections of the NLSY97 survey confirms the presence of non-ignorable selection. Estimated average treatment effects show increasing returns on college experiences from 2004 to 2011.