Doubly Robust Estimation of Causal Effects in Observational Data with Time-to-event OutcomesLi, Runjia (2024) Doubly Robust Estimation of Causal Effects in Observational Data with Time-to-event Outcomes. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractThere is a growing need for novel methods to estimate causal effects in observational data with time-to-event outcomes to provide reliable guidance for treatment strategies in life-threatening conditions. Bias in causal effect estimation can arise from treatment effect heterogeneity, model misspecification, and unmeasured confounders. This dissertation proposes novel approaches to address these challenges. In the first part of the dissertation, I propose a framework for estimating conditional average treatment effects (CATE) in time-to-event data with competing risks. It accounts for treatment effect heterogeneity and protect against model misspecification. Using targeted maximum likelihood estimation (TMLE), I develop a substitution estimator based on cumulative incidence functions (CIF), derived from the efficient influence function (EIF). This estimator is doubly robust, and achieves asymptotic efficiency under mild conditions. Simulations demonstrate its favorable performance across various settings, confirm the double robustness, asymptotic normality and the flexibility of the framework incorporating different regression and machine learning models. Additionally, I construct variable importance measures to identify variables contributing to treatment effect heterogeneity and estimation, providing guidance for clinicians on the critical biomarkers or information to collect. The method is applied to electronic health record data to evaluate the treatment effect of steroids on ICU mortality among sepsis patients. In the second part, I develop a novel instrumental variable (IV) method for estimating average treatment effects in data with unmeasured confounding. Derived from the EIF, this model-free estimator achieves double robustness and asymptotic efficiency under certain mild conditions. Defined by CIF, the method is adaptable to time-to-event data with competing risks. Our method also enables the incorporation of various models for outcome, treatment, and censoring. Extensive simulations demonstrate the double robustness, asymptotic normality, and the capability to analyze complex data. This proposed IV method is applied to investigate the effect of hydrocortisone on mortality among ICU patients with vasopressor-dependent septic shock. Public health significance: The proposed methods address key challenges in estimating causal treatment effects in time-to-event data, including treatment effect heterogeneity, model misspecification, and unmeasured confounders. This dissertation provides powerful tools for optimizing treatment strategies, improving estimation reliability, and advancing healthcare research. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[img]](https://d-scholarship.pitt.edu/style/images/fileicons/application_pdf.png)
![[feed]](/images/feed-icon-32x32.png){kind=icon}