Files
Abstract
In this dissertation, we present three novel contributions, including providing a new methodology, examining the proposed method's performances, and extensive the study in literature. The first paper of this dissertation focuses on the statistical techniques to developing biomarkers that provide the integration of reliable indicators of effectiveness for particular mechanism-depending on medication, guiding treatment selection for cases relying on the tumor's biological makeup and the patient's genotype settings. When we directly attempt to evaluate biomarkers' performance without considering the influence of covariates ' on treatment assignment, the result can lead to inaccurate evaluation of biomarkers performance. To minimize the influence of covariates ' on treatment, outcome, or both, that can produce bias, we have employed various causal inference methods in a lung cancer dataset. Chapter 3 aims to present the general framework for the treatment selection process in literature, consisting of the intersection of machine learning, causal inference, and biomarkers. We use parametric, and machine learning techniques to estimate propensity scores and then, applying pair matching techniques that rely on parametric and machine learning methods to adjust in terms offer the existence of extraneous factors. Different associations between treatment or outcome and covariates are studied and assessed in terms of results in outcome models. After that, we use the results of parametric and machine learning methods to evaluate biomarkers that may estimate whether patients benefit from a specific treatment in observational studies. In chapter 4, an essential requirement for inverse probability weighting estimation is the positivity assumption. However, when the positivity assumption is violated in propensity score distributions between treatment groups, some weights can be approximately 0 and 1. These weights led to uncertainty bias and large variance estimations. We study various techniques to rid of overlap issues. We propose different levels of non-overlap scenarios to examine the performance of balance weighting family and generalized propensity score matching across true propensity model and misspecified propensity score model in multiple treatment cases. We present results of different methods of variance estimator( i.e., a robust sandwich-type variance estimator and a bootstrap variance estimator) to estimate the causal effect.