Valid confidence intervals for post-model-selection predictors (Q2414094)

This paper proposes confidence intervals for post-model-selection predictors in a regression framework. The confidence intervals are constructed in a similar way to the PoSI intervals, and intend to cover the quantity of interest with a use-specified minimal coverage probability, irrespective of the model selection procedure. The quantity of interest is, basically, a projection of the expected value of the dependent variable on the space spanned by the regressors included in the selected model. The validity of the confidence levels is established under different conditions in the paper. The increase in width due to the intervals being obtained by maximizing over all inactive variables is show to be small in simulations and asymptotically. The simulations include model selection employing AIC, BIC, LASSO, SCAD and MCP.