Stability Selection is a popular way to make variable selection more reliable, but it resamples observations while leaving the design matrix untouched. The authors of a new arXiv paper point out that this lets an irrelevant feature that is correlated with relevant ones—a confounder—get selected in a substantial fraction of resamples, breaking the exchangeability assumption behind Stability Selection's false-positive bound.

To fix this, Mahdi Nouraie, Houying Zhu, and Samuel Muller propose 2D Stability Selection (2D-SS), which aggregates selection frequencies over half-samples and over a grid of injected design noise; at zero noise it reduces to Stability Selection. In expectation, the injected noise acts as Tikhonov regularisation, making a jittered Lasso fit a noisy elastic net whose support recovery is governed by a noise-indexed irrepresentable condition. Jittering can therefore repair designs where confounders violate the Lasso condition, although it can also hurt.

The paper provides a noise level at which a violated feature's statistic vanishes (for uncorrelated relevant features), exact-recovery guarantees for a single fit, conditions under which aggregation over noise levels recovers the support, and a complementary-pairs bound on the expected number of false positives carried over from Stability Selection. The procedure takes no measurement-error covariance as input, though the guarantees assume i.i.d. Gaussian measurement error. In simulations, 2D-SS excluded confounders that Stability Selection included, and its size-matched version stayed below the nominal closed-form false-positive bound in all 240 simulated cells, whereas Stability Selection exceeded it in 21. It did not generally improve on Stability Selection in F1 and was lower at low noise, so the authors position it as useful when stable but wrong selections are the concern.