Classical confidence intervals after best subset selection are widely implemented in statistical software and are routinely used to guide practitioners in scientific fields to conclude significance. However, there are increasing concerns in the recent literature about the validity of these confidence intervals in that the intended frequentist coverage is not attained. In the context of the Akaike information criterion (AIC), recent studies observe an under-coverage phenomenon in terms of overfitting, where the estimate of error variance under the selected submodel is smaller than that for the true model. Under-coverage is particularly troubling in selective inference as it points to inflated Type I errors that would invalidate significant findings. In this article, we delineate a complementary, yet provably more deciding factor behind the incorrect coverage of classical confidence intervals under AIC, in terms of altered conditional sampling distributions of pivotal quantities. Resting on selective techniques developed in other settings, our finite-sample characterization of the selection event under AIC uncovers its geometry as a union of finitely many intervals on the real line, based on which we derive new confidence intervals with guaranteed coverage for any sample size. This geometry derived for AIC selection enables exact (and typically less than exact) conditioning, circumventing the need for the excessive conditioning common in other post-selection methods. The proposed methods are easy to implement and can be broadly applied to other commonly used best subset selection criteria. In an application to a classical US consumption dataset, the proposed confidence intervals arrive at different conclusions compared to the conventional ones, even when the selected model is the full model, leading to interpretable findings that better align with empirical observations.

翻译：暂无翻译