B1723
Title: Variable selection method for the logistic regression model using a model-X knockoffs algorithm
Authors: Takafumi Nakatsu - Chuo University (Japan) [presenting]
Toshinari Kamakura - Chuo University (Japan)
Abstract: In the field of medical research, the logistic regression model is commonly used for binary response variables, such as, treatment success or failure. It is important to select covariates which have substantial relation to the response variables to use the model. Covariates are usually selected with the standard likelihood-based testing methods, the likelihood ratio test, score test, and Wald test, by exploring a combination of covariates which gives true non-zero regression parameters. However, until now, the optimality of selected covariates has not yet been well investigated. Recently, a novel statistical procedure, model-X knockoffs algorithm has provided a way to identify important covariates under controlling False Discovery Rate (FDR). We developed a new method to reach an optimum solution of covariate combination with true non-zero regression parameters, employing the model-X knockoffs algorithm. Features of the method is being investigated in comparative simulation studies. The results of the studies will be presented and discussed.
