Subgroup Identification and Variable Selection from Randomized Clinical Trial Data.

Subgroup Analysis;Variable Selection;Randomized Clinical Trials;Statistics and Numeric Data;Science;Biostatistics

Foster, Jared C.Wang, Lu ;

University of Michigan

关键词: Subgroup Analysis;    Variable Selection;    Randomized Clinical Trials;    Statistics and Numeric Data;    Science;    Biostatistics;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/102382/jaredcf_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

We consider the situation of a randomized clinical trial with a moderate number of baseline covariates. While the overall treatment effect may be modest, it is plausible that there is a subgroup of individuals, defined by their baseline covariate values, that have an enhanced treatment effect. The challenge is to identify a subgroup from the randomized trial data.Such subgroups may be used to aid in future treatment decisions, so it is desirable that they be simple, and depend on only a limited number of covariates.We consider two methods which use randomized clinical trial data to identify such subgroups, the first of which is a penalized monotone single-index model.Though not a stand-alone subgroup identification method, this model can be used to perform the variable selection stage of a subgroup identification procedure.In this method, the expected treatment effect is assumed to be an unknown monotone function of a linear combination of covariates.To estimate this unknown function and the linear combination, an adaptive LASSO-penalized monotone single-index model is employed. The second method is a two-stage subgroup identification procedure.In stage 1, nonparametric regression is used to estimate treatment effects for each subject.In the stage 2, a systematic evaluation of many subgroups of a simple, pre-specified form is performed.Using a criterion which is based on the estimated treatment effects obtained in stage 1, the best of these subgroups is identified.To help reduce the risk of false positive findings, we also propose a number of permutation-based methods for obtaining p-values for treatment-by-covariate interactions, which can be used to test whether or not the identified subgroup is truly enhanced.All methods are evaluated in simulations studies, and illustrated using randomized clinical trial data.