Consistent model selection criteria for quadratically supported risks (Q510678)

In a regression problem, the regression coefficients of covariates are usually estimated by minimizing a risk function defined under a specific loss function. This paper defines a class of loss functions which includes those used in linear regression with Gaussian errors, generalized linear regression, robust regression and quantile regression. When the dimension of the covariates is much larger than the sample size, the optimal model can be selected based on the generalized information criterion (GIC), which is defined based on the loss function with a penalty for the number of selected covariates. A set of sufficient conditions is provided for the consistency of this GIC procedure. Similar results are obtained when the calculation of GIC values for all possible submodels is impossible and GIC is applied only to submodels that correspond to a solution path of a penalized estimator. A data-adaptive model selection procedure which is selection consistent is also proposed to choose a GIC that works well with moderate sample sizes.