Regularized quantile regression for ultrahigh-dimensional data with nonignorable missing responses (Q2189749)

The authors consider the quantile regression \(Y_i=x_i^T\beta (\tau)+e_i(\tau),\;i=1,\ldots,n,\) where \(x_i=(1,X_{i1},X_{i2},\ldots,X_{ip_n})^T\), is the covariate vector, \(Y_i\) are the responses, \(\beta(\tau)=(\beta_0(\tau),\beta_1(\tau),\ldots, \beta_{p_n}(\tau))^T\) is an unknown coefficient vector and \(e_i(\tau)\) is an independent random error variable with \(\tau-\)th quantile \(0\). Here data are supposed to be ultrahigh-dimensional with missing responses and a regularized estimator of \(\beta\) is looked for. The propensity score is specified by the semiparametric exponential tilting model. The Pearson Chi-square type test statistic is used for identification of the important features in the sparse propensity score model and the adjusted empirical likelihood method is used for estimation of the parameters in the reduced model. With the estimated propensity score model, the authors suggest an inverse probability weighted and penalized objective function for regularized estimation with non convex SCAD penalty and MCP functions. The oracle properties of the proposed regularized estimators are estblished, assuming the propensity score model is of low dimension. Simulation study and real data analysis are given to examine the finite sample performance of the proposed approaches.