UCLA

Nonrandomized epidemiologic studies face many obstacles in attempting to quantify the effect of an exposure on an outcome. Studies of the associations between Parkinson’s disease and cancer may be particularly vulnerable to uncontrolled confounding, information bias, and selection bias. This dissertation provides graphical models and descriptions of how PD-cancer studies may be affected by biases. An overview is also provided of how investigators have attempted to control for these biases. A novel multiple bias modeling method called simultaneous multi-bias adjustment is developed to address this problem within a data fusion framework. Simulation studies are used to support the validity of this method and compare it to the more traditional approach of sequentially adjusting for multiple biases. Simultaneous multi-bias adjustment is then applied to study of the effect of PD on cancer in a retrospective cohort study using Danish population registry data combined with behavioral information collected from questionnaires and surveys. The observed effect of PD on overall cancer in this data set is approximately null. The effect estimate remains null after simultaneous multi-bias adjustment is applied, accounting for PD misclassification and selection bias related to participation and censoring. The simulation studies and Danish cohort study reveal computational and methodological challenges in performing simultaneous multi-bias adjustment. An interactive website and R package are developed to make this method more accessible to others. By developing and demonstrating how to perform simultaneous multi-bias adjustment, showing its validity in simulated data, applying it to real-world data, and making tools for its usage, this dissertation aims to make multiple bias adjustment in causal modeling more accessible to other investigators.