A0766
Title: Improving prediction with linear regression by shrinking the contributions of the predictors
Authors: Masao Ueki - Nagasaki University (Japan) [presenting]
Abstract: An approach is developed to improve prediction with linear regression by shrinking the contributions of the predictors to the predicted value, where the contribution of a predictor is defined as the predictor variable multiplied by the corresponding regression coefficient. To shrink the contribution of a predictor, the soft-thresholding function and its extension induced by the elastic net are applied. As with the familiar variable selection approach, it is possible to eliminate predictor variables if the contribution of a predictor for all subjects is thresholded at zero, but the present approach can result in a more detailed sparse output of the contributions of the predictors for individual subjects' predicted values. Simulation studies confirm that the proposed thresholding improves the linear regression and shows good performance relative to other statistical and machine learning methods. Real data applications, including applications to polygenic risk scores, show an improved predictive performance, as in simulation studies.
