Optimizing Dynamic Treatment Regimes with Q-Learning: Complications due to Error-Prone Data and Applications to COVID-19 Data
Abstract
In this thesis, we employ statistical modeling and methods to examine COVID-19 data, and we develop new methods to address new issues that invalidate some standard methods.
In the first study, we employ semiparametric and nonparametric survival models as well as data visualization techniques to examine the epidemiological features of COVID-19. Based on our numerical results, the median incubation time is about 5 days, and the elders are more likely to have longer incubation periods.
In the second study, we use data from 175 countries and investigate possible factors associated with the case fatality rate (CFR) of COVID-19. The Q-learning algorithm is employed to assess optimal preventive policies adopted by individual countries to reduce their COVID-19 CFRs. The data analysis suggests that, in addition to addressing traditional risk factors, policymakers should tailor the strictness of preventive policies to country-specific characteristics and evolving situation to alleviate the risk of death from COVID-19.
The third study investigates the effects of misclassified covariates in developing dynamic treatment regimes with the Q-learning approach. We present two procedures to account for the bias induced by covariate misclassification. The satisfactory performance of these procedures is demonstrated through extensive simulation studies.
The fourth study deals with mixed measurement error and misclassification in covariates within the context of Q-learning with compound outcome. We demonstrate that the presence of such measurement inaccuracies can pose significant challenges to the accurate estimation process in Q-learning. To address this issue, we propose effective correction strategies that successfully alleviate the impact of mismeasurement.