From Theory to Practice: Improving Causal Conclusions from Healthcare Data

Abstract

Causal inference in biomedical, epidemiological, and health policy research often relies on observational data, such as electronic health records (EHRs), patient registries, or insurance claims, to compensate for the inaccessibility of randomized controlled trials. However, causal inference from observational data depends on strong, often unverifiable assumptions, including exchangeability, parallel trends, and correct model specification. Violations of these assumptions can bias treatment effect estimates, making it essential to assess the sensitivity of causal conclusions--particularly in healthcare applications, where properly interpreting causal relationships and delivering reliable insights is critical for guiding clinical practice and informing system-wide decisions. This thesis contributes to the theoretical and empirical analysis of causal methods under realistic data limitations, with a focus on covariate selection and adjustment, proximal inference for unobserved confounding, and applications of modern estimation techniques to healthcare-relevant settings. In Chapter 2, we investigate the performance and robustness of state-of-the-art machine learning estimators for causal inference when covariate selection for statistical adjustment is performed in a realistically suboptimal manner. Although nonparametric doubly robust methods are asymptotically unbiased, they can perform poorly in finite samples due to slow convergence of nuisance function estimates. Through an extensive simulation study, built upon previous research on statin use and atherosclerotic cardiovascular disease (ASCVD) incidence, we examine how including extraneous covariates---a likely risk when researchers over-adjust to mitigate concerns about unmeasured confounding---may degrade estimator performance. These findings highlight the importance of incorporating domain knowledge to guide covariate selection, even when using flexible data-adaptive methods. In Chapter 3, we explore proximal causal inference, a novel framework designed to address unobserved confounding by leveraging negative control exposures and outcomes to recover the true causal effects. While this approach offers an alternative to the exchangeability assumption, it relies on identification conditions for the proxy variables set and model specifications that remain empirically untestable. We derive closed-form bias expressions under a linear structural equation model to quantify the impact of violating these assumptions and propose a practical bias adjustment strategy using data from an observational ICU study. These results provide a foundation for formal sensitivity analysis and offer insight into the real-world utility of proximal methods. Finally, in Chapter 4 we evaluate the impact of the Meaningful Use Incentive Program on hospital performance, using modern causal methods in a multi-period difference-in-differences (DiD) design. We apply a staggered DiD estimation framework, along with a sensitivity analysis of dynamic treatment effect estimates under potential violations of the parallel trends assumption, across a wide range of quality, safety, and process of care measures. By accounting for treatment timing variation, allowing for heterogeneous effects over a longer follow-up period, and testing for violations of identifying assumptions, our study offers a more rigorous and comprehensive assessment of the causal impact of health information technology (IT) policies introduced by the Meaningful Use program. Our findings help reconcile mixed findings in the literature and inform the design of future hospital incentive programs that aim to promote advanced use of EHRs.