UCLA

Observational studies investigating chronic diseases are valuable in informing public health guidelines. Yet, it is pertinent that studies are evaluated for threats of biases including uncontrolled confounding, which may produce misguided conclusions if left unaddressed. The fitting of marginal structural models using g-computation is a method capable in evaluating time-varying study designs but methods in addressing the bias due to time-varying uncontrolled confounding are understudied. This dissertation graphically describes the types of time-varying uncontrolled confounding structures. Combined with g-computation, the dissertation highlights the development of two novel bias analysis methods to address time-varying uncontrolled confounders in a plasmode simulation framework. The first method utilizes a bias offset to subtract confounding from the outcome and formally is applied to an investigation of elevated depressive symptoms with cancer outcomes. The last method employs an inverse probability of uncontrolled confounder weight as the bias weighting method and is applied to a study of elevated depressive symptoms with cardiovascular disease. Several settings can describe time-varying uncontrolled confounding including relations into subsequently measured confounders, exposures, and the outcome. Simulated results demonstrated the most biased estimates when time-varying uncontrolled confounding were consistently strong over time. In simulations, both the bias offset and bias weighting methods could recover a true estimate from mis-specified, biased models across a series of different bias parameters and effect estimates. In applied illustrations, null and modest relationships were observed between elevated depressive symptoms and incidence of cancer outcomes and cardiovascular disease, respectively. For all illustrations, applied bias analysis suggested robust results to modest levels of confounding. Time-varying uncontrolled confounding can immensely impact observed estimates. This dissertation provides a principled approach to alternative explanations (due to mixing of effects) to enable more credible causal inference in the health sciences. Applied examples demonstrate two novel bias analysis methods and a guided framework of how bias analysis can be combined with causal inference methods. Longitudinal observational studies are integral in advising public health and thus the impact of sources of bias warrants more recognition and careful evaluation using bias analysis methods.