Power of the Spacing test for Least-Angle Regression

TL;DR

This paper analyzes the power and unbiasedness of the spacing test for LARS in high-dimensional linear models, extending it to unknown noise variance and comparing its effectiveness with other tests.

Contribution

It provides the first unbiasedness proof for the spacing test under alternatives and extends it to unknown noise variance, with detailed power analysis.

Findings

The spacing test is unbiased and has power at least equal to the significance level.

Its rejection region is optimal for orthogonal predictors.

The test's power exceeds the significance level as it approaches zero.

Abstract

Recent advances in Post-Selection Inference have shown that conditional testing is relevant and tractable in high-dimensions. In the Gaussian linear model, further works have derived unconditional test statistics such as the Kac-Rice Pivot for general penalized problems. In order to test the global null, a prominent offspring of this breakthrough is the spacing test that accounts the relative separation between the first two knots of the celebrated least-angle regression (LARS) algorithm. However, no results have been shown regarding the distribution of these test statistics under the alternative. For the first time, this paper addresses this important issue for the spacing test and shows that it is unconditionally unbiased. Furthermore, we provide the first extension of the spacing test to the frame of unknown noise variance. More precisely, we investigate the power of the spacing test for LARS and prove that it is unbiased: its power is always greater or equal to the significance level $\alpha$. In particular, we describe the power of this test under various scenarii: we prove that its rejection region is optimal when the predictors are orthogonal; as the level $\alpha$ goes to zero, we show that the probability of getting a true positive is much greater than $\alpha$; and we give a detailed description of its power in the case of two predictors. Moreover, we numerically investigate a comparison between the spacing test for LARS and the Pearson's chi-squared test (goodness of fit).