Andrew Topp of the Department of Biostatistics defends his dissertation on "Doubly Robust Estimation in Two-stage Dynamic Treatment Regimes in the Presence of Drop-out"
Graduate faculty of the University and all other interested parties are invited to attend.
Various methods exist for causal inference about the effects of different treatments from observational studies or randomized trials. A straightforward approach is to fit a regression model of the outcome as a function of the treatment they received along with observed patient characteristics. Other methods, such as inverse probability weighting, work by instead estimating a patient's probability of receiving treatment and weighting the outcomes by the inverse of this probability. Doubly Robust estimators use both models and provide unbiased estimates as long as either the probability of treatment or outcome is correctly modeled. These techniques can be extended to analyze and compare dynamic treatment regimes, that is, multiple stages of treatment punctuated by decision points concerning what the next treatment should be. This dissertation is concerned with developing more efficient doubly robust estimators for two-stage dynamic treatment regimes, first without and later, with data missing at random. First, we develop a new inverse probability of treatment weighted and doubly robust estimators for analyzing dynamic treatment regimes. Then we compare these methods to the corresponding existing methods for estimating the mean outcome of dynamic treatment regimes in a simulation. We utilize the new doubly robust estimator in the analysis of the STAR*D trial to estimate the mean outcome of patients on different regimes for the treatment of non-psychotic major depressive disorder. Finally, we propose a modification of the new inverse probability of treatment weighted and doubly robust estimators in order to account for missing data.