Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Rigorous theory for covariate-adaptive MAS thresholding

Establish rigorous theoretical justification for applying the p-value-free threshold rule to the covariate-adaptive agent-X importance statistics and the resulting covariate-adaptive Model-Agnostic Selection with agent-X importance (MAS-CA) procedure.

Background

The paper defines a covariate-adaptive agent-X importance by accumulating local importance measures over the steps of learning algorithms with solution paths, such as stepwise regression and tree-based methods. The resulting statistics are evaluated using the same p-value-free threshold rule as the main MAS procedure to identify discoveries.

Numerical experiments reportedly show that the covariate-adaptive construction can achieve higher power than the standard MAS algorithm. However, the paper does not establish that the threshold rule continues to control the false discovery rate for these covariate-adaptive statistics, leaving the validity of the procedure unresolved.

References

Rigorous justifications on this threshold rule are left for future work.

— Controlling the False Discovery Rate via Model-Agnostic Importance Measures  (2610.21360 - Wen et al., 15 Oct 2026) in Section 4, subsection “Covariate-adaptive importance”