NBER Working Papers · 2026 · Brian Curran、Bruce D. Meyer、Derek Wu
中文摘要
在估计二元结果的非线性模型（如probit和logit模型）时，研究者通常依赖平均偏效应（APEs）来概括某一回归变量的效应。由于这些模型中某一变量的边际效应取决于所有其他变量的取值，APE的数值取决于用于计算的样本部分。当基于并非用于界定研究对象的子总体所抽取的样本部分进行平均时，APE可能具有误导性。本文强调了若干常见情形——例如包含次要组的均值差异以及双重差分设计——在这些情形中，针对全样本计算的APE会偏离适当样本部分的边际效应。我们针对特定情形提出一种简单且零成本的解决方案，并通过模拟表明，在适当子样本上重新计算APE可得到无偏结果。通过重新审视多篇已发表论文的结果，我们发现报告的估计值与适当计算的APE之间存在统计显著的差异。
Abstract
When estimating nonlinear models for binary outcomes, such as probit and logit models, researchers often rely on average partial effects (APEs) to summarize the effect of a regressor. Because the marginal effect of a variable in these models depends on the values of all other variables, the value of an APE hinges on the portion of the sample used for the calculations. When averaged over parts of the sample drawn from a subpopulation not used to define the object of interest, the APE may be misleading. This paper highlights common situations, such as differences-in-means with a secondary group and difference-in-differences designs, where APEs calculated for the full sample deviate from marginal effects for the appropriate part of the sample. We propose a simple and costless solution in specific cases and demonstrate through simulations that recalculating APEs over the appropriate subsample yields unbiased results. Reexamining published results from multiple papers, we find statistically significant discrepancies between the reported estimates and the appropriately calculated APEs.