Causal Inference for Error-Prone Exposures

The potential pitfalls that confounding and measurement bias can impose on a study are enormous and have been known and discussed in the scientific literature for over a century. Development of statistical methods to adjust for these phenomena has grown rapidly over the last few decades, producing broad methodologies for causal inference and measurement error correction which span many disciplines. However, there have been comparatively few papers at the intersection of these two fields, despite confounding and measurement error often coinciding in the same data. Of the few available methods, most rely on investigators having supplemental data such as replication, validation, or instrumental data. Access to such additional data is uncommon in many areas of practice. Thus, the goal of this dissertation is to provide causal inference methods for error-prone exposures that have desirable properties even in the absence of supplemental data. In Chapter 2, a set of such methods are proposed, which extend the conditional score approach for measurement error correction such that it also adjusts for observed confounding in point-exposure studies. Three estimators are proposed based on g-computation, inverse probability weighting, and doubly-robust estimation. The proposed estimators are consistent and asymptotically normal and the methods are shown to perform well in finite samples using a simulation study. Their application is illustrated using the HVTN 505 vaccine trial. In Chapters 3 and 4, the methods move beyond the point-exposure setting to the longitudinal setting. In particular, the problem of joint modeling of longitudinal and survival data is considered, which can provide consistent estimation when longitudinal exposures are measured with error and/or measured sparsely over time. In Chapter 3, a framework for causal joint modeling is defined and estimating equations weighted by inverse probability of treatment weights are proposed to estimate parameters of causal joint models. In Chapter 4, the study of causal joint modeling is continued, but focus is moved to counterfactual survival curve estimands and to estimators based on g-computation. The methods are applied to the AMP study to evaluate the effect of a broadly-neutralizing antibody on HIV acquisition in the context of participant utilization of pre-exposure prophylaxis.Doctor of Philosoph