中文摘要
缩尾与截尾用于尽可能减小离群值对处理效应估计值的影响。在随机对照试验（RCT）中，通常的做法是将处理组与对照组合并，对整个样本的尾部进行缩尾或截尾。这可能导致处理组与对照组中被缩尾或截尾的观测值比例不相等。另一种方法是分层缩尾／截尾，该方法分别对各处理组进行缩尾，确保每个实验组别中被缩尾或截尾的观测值比例相等。一个形式化框架以及针对一项RCT的蒙特卡洛模拟表明，与传统方法相比，分层缩尾／截尾降低了处理效应估计偏误和第二类错误的风险，但代价是第一类错误发生的可能性更高。将这些方法应用于Angelucci等（2023）和Jack等（2023）的研究表明，所选择的缩尾或截尾技术可能影响处理效应的量级和统计显著性。本文还讨论了面向开展RCT且希望对离群值进行缩尾或截尾处理的研究者的实用指南。
Abstract
Winsorizing and trimming are used to minimize the effects of outliers on estimated treatment effects. In Randomized Controlled Trials (RCTs), the typical approach winsorizes/trims the tails of the whole sample, pooling together treatment and control groups. This can have as a consequence that observations from treatment and control groups are disproportionately winsorized/trimmed. An alternative approach, Stratified Winsorizing/Trimming, winsorizes treatment groups separately, ensuring that an equal proportion of observations are winsorized/trimmed per experimental arm. A formal framework and Monte Carlo simulations of an RCT illustrate that Stratified Winsorizing/Trimming reduces the treatment effect bias and risk of Type II errors compared to the traditional approach, although at the cost of a greater likelihood of Type I errors. Applications to Angelucci et al. (2023) and Jack et al. (2023) illustrate that the chosen winsorizing/trimming technique can affect the magnitude and statistical significance of treatment effects. Practical guidelines for researchers conducting RCTs that want to winsorize/trim outliers are discussed.