False Discovery Rate Control via Data Splitting

Selecting relevant features associated with a given response variable is an important problem in many scientific fields. Quantifying quality and uncertainty of selection result via false discovery rate (FDR) control has been recent interest. This article introduces data-splitting method (referred to as “DS”) asymptotically the FDR while maintaining high power. For each feature, DS constructs test statistic by estimating two independent regression coefficients data splitting. achieved taking advantage statistic’s property that, for any null its sampling distribution symmetric about zero; whereas positive mean. Furthermore, Multiple Data Splitting (MDS) proposed stabilize boost Surprisingly, under control, MDS not only helps overcome power loss caused splitting, but also results lower variance proportion (FDP) compared all other methods consideration. Extensive simulation studies real-data application show that are robust unknown features, easy implement computationally efficient, often most powerful ones among competitors especially when signals weak correlations or partial high. Supplementary materials this available online.