Identifying significant variables for the high-dimensional logistic regression model is a fundamental problem in modern statistics and machine learning. This paper introduces a stability knockoffs (StabKoff) selection procedure by merging stability selection and knockoffs to conduct controlled variable selection for logistic regression. Under some regularity conditions, we show that the proposed method achieves FDR control under the finite-sample setting, and the power also asymptotically approaches one as the sample size tends to infinity. In addition, we further develop an intersection strategy that allows better separation of knockoff statistics between significant and unimportant variables, which in some cases leads to an increase in power. The simulation studies demonstrate that the proposed method possesses satisfactory finite-sample performance compared with existing methods in terms of both FDR and power. We also apply the proposed method to a real data set on opioid use disorder treatment.

识别高维逻辑回归模型的重要变量是现代统计学和机器学习中的一个基本问题。本文介绍了一种稳定性仿制品（StabKoff）选择过程，通过合并稳定性选择和仿制品来进行逻辑回归的受控变量选择。在某些规律性条件下，我们表明该方法在有限样本设置下实现了 FDR 控制，并且随着样本量趋于无穷大，功效也逐渐接近 1。此外，我们进一步开发了一种交叉策略，可以更好地分离重要变量和不重要变量之间的仿冒统计数据，这在某些情况下会导致功效增加。仿真研究表明，与现有方法相比，该方法在 FDR 和功耗方面均具有令人满意的有限样本性能。我们还将所提出的方法应用于阿片类药物使用障碍治疗的真实数据集。