Optimal treatment regime is the individualized treatment decision rule which yields the optimal treatment outcomes in expectation. A simple case of treatment decision rule is the linear decision rule, which is characterized by its coefficients and its threshold. As patients heterogeneity data accumulates, it is of interest to estimate the optimal treatment regime with a linear decision rule in high-dimensional settings. Single timepoint optimal treatment regime can be estimated using Concordance-assisted learning (CAL), which is based on pairwise comparison. CAL is flexible and achieves good results in low dimensions. However, with an indicator function inside it, CAL is difficult to optimize in high dimensions. Recently, researchers proposed a smoothing approach using a family of cumulative distribution functions to replace indicator functions. In this paper, we introduce smoothed concordance-assisted learning (SMCAL), which applies the smoothing method to CAL using a family of sigmoid functions. We then prove the convergence rates of the estimated coefficients by analyzing the approximation and stochastic errors for the cases when the covariates are continuous. We also consider discrete covariates cases, and establish similar results. Simulation studies are conducted, demonstrating the advantage of our method.

翻译：暂无翻译