2025 Theses Doctoral

Methods to Reduce Selection and Confounding Bias in Observational and Clinical Studies

Pitts, Amy Johnson

This dissertation explores statistical methods to enhance the robustness and reliability of inference in observational and clinical studies, with a focus on mitigating selection and confounding biases.

First, we address challenges in survey inference due to sampling bias introduced by disruptions from the COVID-19 pandemic. Using data from an ongoing survey study of people living with HIV in New York City, we assess the impact of recruitment bias on estimates of HIV viral suppression and mental and physical wellness. To correct for deviations in sample representation we propose a model-based approach for survey inference when only the marginal distributions of the population characteristics referred to in this text as an adaptation of multilevel regression with poststratification (MRP) approach. Our findings demonstrate that the MRP adaptation approach improves inference by adjusting for selection bias introduced during the COVID-19 pandemic.

Next, we investigate a causal inference challenge involving extreme positivity violations in an observational study of prenatal anesthesia exposure. The goal is to estimate the causal effect of mothers receiving anesthesia during pregnancy on the diagnosis of disruptive or internalizing behavioral disorders (DIBD) in children, separate from the confounding effects of surgery. Since anesthesia and surgery are deterministically linked, standard methods fail to disentangle their effects. To address this, we employ the separable effects model of Robins and Richardson (2010), which isolates the direct effect of anesthesia by blocking pathways through variables that fully mediate the effect of surgery on DIBD.

Finally, we examine generalizability and transportability methods in a multi-study, multi-outcome setting, focusing on medication for opioid use disorder (MOUD) and its impact on family/social status. In this application, only one clinical trial directly measured the outcome of interest, limiting statistical power. We develop semi-parametric estimation techniques that leverage additional post-exposure variable proxies across harmonized clinical trials from the National Institute on Drug Abuse (NIDA) Clinical Trials Network (CTN). These methods leverage additional post-exposure variables to help gain insight into previously underpowered outcomes.

Together, these projects contribute to the development of statistical methods that improve the validity and interpretability of findings in complex real-world data settings, advancing both survey methodology and causal inference in clinical and observational research.