When scientists use machine learning to pick which covariates matter, they need to avoid false positives. Traditional variable selection often depends on specific model assumptions, and Model-X knockoffs require knowing the covariate distribution. A new paper on arXiv offers a complementary approach.

The method, called Model-Agnostic Selection with agent-X importance (MAS), uses sample-splitting and a novel importance measure. The measure compares oracle predictiveness on the original distribution and on a distribution with an 'agent' variable set to a constant, estimated via cross-fitting. This removes the need for covariate distribution assumptions.

MAS also uses a resplit-per-covariate strategy to decouple dependencies between estimators and a data aggregation technique to boost power. The authors prove asymptotic FDR control under mild conditions and show in simulations and real-data analyses that MAS matches existing methods in FDR control, computational efficiency, and power.