Two-stage subsampling variable selection for sparse high-dimensional generalized linear models

Although high-dimensional data analysis has received a lot of attention after the advent of omics data, model selection in this setting continues to be challenging and there is still substantial room for improvement. Through a novel combination of existing methods, we propose here a two-stage subsampling approach for variable selection in high-dimensional generalized linear regression models. In the first stage, we screen the variables using smoothly clipped absolute deviance penalty regularization followed by partial least squares regression on repeated subsamples of the data; we include in the second stage only those predictors that were most frequently selected over the subsamples either by smoothly clipped absolute deviance or for having the top loadings in either of the first two partial least squares regression components. In the second stage, we again repeatedly subsample the data and, for each subsample, we find the best Akaike information criterion model based on an exhaustive search of all possible models on the reduced set of predictors. We then include in the final model those predictors with high selection probability across the subsamples. We prove that the proposed first-stage estimator is $n^{1/2}$-consistent and that the true predictors are included in the first stage with probability converging to 1. In an extensive simulation study, we show that this two-stage approach outperforms the competitors yielding among the highest probability of selecting the true model while having one of the lowest number of false positives in the settings of logistic, Poisson, and linear regression. We illustrate the proposed method on two gene expression cancer datasets.