Contra: Contrarian statistics for controlled variable selection.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Holdout randomization tests (HRTs) identify predictive covariates but can inflate the false discovery rate (FDR) when covariate distributions are estimated from data.
• Accurate FDR control is crucial for reliable covariate selection in statistical modeling.

Purpose of the Study:

• To introduce the contrarian randomization test (CONTRA) for robust FDR control in scenarios with estimated or misspecified covariate distributions.
• To offer a computationally efficient and powerful alternative to existing methods for high-dimensional data.

Main Methods:

• CONTRA utilizes an equal mixture of two probabilistic models: one fitted with real data and another with a modified dataset where the tested covariate is replaced by estimated samples.
• This approach explicitly addresses challenges arising from unknown or misspecified covariate distributions.

Main Results:

• CONTRA effectively reduces FDR inflation compared to state-of-the-art methods when covariate distributions are misspecified.
• The method demonstrates asymptotic power of 1 and computational efficiency for high-dimensional and large sample size datasets.
• Effectiveness validated on synthetic benchmarks and a genetic dataset.

Conclusions:

• CONTRA provides a robust and efficient solution for covariate selection with improved FDR control, particularly when dealing with estimated or misspecified covariate distributions.
• The method shows promise for applications in bioinformatics and other fields requiring reliable statistical inference.