Variable Selection for Dynamic Treatment Regimens
Author | : Zeyu Bian |
Publisher | : |
Total Pages | : 0 |
Release | : 2022 |
ISBN-10 | : OCLC:1358412429 |
ISBN-13 | : |
Rating | : 4/5 (29 Downloads) |
Book excerpt: "In the precision medicine paradigm, treatment decisions are tailored to each individual, instead of a "one-size-fits-all" approach, which is beneficial in the presence of heterogeneous treatment effects. With the aim of improving individual patients' health outcomes, dynamic treatment regimens (DTRs) recommend effective treatments for individual patients based on their characteristics. However, collected data often contain many irrelevant variables for tailoring treatment. Including all the covariates in an analysis could yield a loss of statistical efficiency and an unnecessarily complicated treatment decision rule, which is difficult for physicians to assess or implement. Thus, variable selection with the objective of optimizing patients' outcome by identifying useful tailoring variables is important. The topic of variable selection in a general context has been well studied, however, it has been less investigated in the area of DTRs. Applying existing variable selection techniques to DTRs estimation methods directly can be challenging: first, the goal of variable selection in DTRs differs from variable selection in a general context. Variable selection for DTRs aims to improve the estimated decision rules instead of predictive performance. Second, in DTRs, we are most interested in selecting variables that may be effect modifiers-a scenario that is rarely considered in the prediction setting. Last, DTRs are often estimated using semi-parametric methods that provide robustness against model misspecification. Many existing methods are complicated and hard to implement (especially for count and binary outcomes), thus it is difficult to extend these to a regularization framework. In such a case, we might want to use a simpler regression-based method.The overarching goal of this thesis is to develop new variable selection techniques in the DTRs setting. This thesis consists of three manuscripts. In the first manuscript, I extend the estimation approach of dynamic weighted ordinary least squares to a penalized framework, where estimation and variable selection for DTRs can be performed simultaneously. I show that this extension has the double robustness and oracle properties under some conditions. The newly proposed method is applied to data from the Sequenced Treatment Alternatives to Relieve Depression study. The second manuscript considers two practical issues that frequently arise in causal inference and variable selection approaches: confounder selection and tuning parameter selection. The approach from the first paper is combined with a confounder selection method, and this is illustrated on data from a pilot sequential multiple assignment randomized trial of a web-based stress management study. In these first two works, I only considered the case in which the outcome is continuous, while in the third manuscript, I extend the doubly robust penalized weighted regression approach to the discrete outcomes setting.In this thesis, I show that with a suitable choice of weights, a weighted penalized regression model still enjoys the desired double robustness property, and yet is straightforward to implement. The advantage of the newly proposed approach compared to alternative regularized DTRs estimation methods lies in the fact that it can be viewed from a minimization perspective. Hence, the implementation is simpler, various penalty functions can be used, and the solution can be found using existing computationally efficient tools"--