| 喻开志,司国庆.双重差分模型识别设计的理论与应用——基于数据结构差异视角[J].数量经济技术经济研究,2026,(7):235-260 | | 双重差分模型识别设计的理论与应用——基于数据结构差异视角 | | Theory and Application of Identification Design in Difference-in-Differences Models: Perspective of Differences in Data Structure | | | | DOI： | | 中文关键词: 数据结构 双重差分法 双重机器学习 双重固定效应 | | 英文关键词: Data Structures Difference-in-Differences Double Machine Learning Two-Way Fixed Effects | | 基金项目: | | | 中文摘要: | | 双重差分法是社会科学领域应用最为广泛的研究手段之一，但其目标参数和识别设计却在不同数据结构类型下存在显著差异。实践中，忽视数据样本构成变化而机械套用传统双重差分法，会加剧因模型误设、数据维度过高等问题引发的估计偏误，导致结果偏离真实可识别目标参数。本文从数据结构差异视角，针对传统双重差分设计的内在局限，依托半参数理论，梳理总结多种双重稳健双重差分法。通过对比识别假设、目标参数与估计策略，厘清各类方法的识别设计与估计差异，并系统阐明机器学习在其中的应用机理，据此提出操作性强的使用建议。本文同时结合数值模拟与实际案例，进一步剖析了不同方法的实际估计表现。本文旨在从数据结构差异视角，为研究者理解与运用前沿双重差分设计提供理论指引与实践指南，以助力其把握各类方法的适用条件与使用边界，进而提升在经验研究中对因果效应的准确识别能力。 | | 英文摘要: | | Difference-in-Differences (DID) is one of the most widely used research methods in the social sciences. However, its target parameters and identifying assumptions differ markedly across data structures. In practice, a mechanical applicationof traditional DID methods without accounting for changes in sample composition often exacerbates the estimation biases arising from issues such as model misspecification and high-dimensional data, thereby leading to results that deviate from the true identifiable target parameters. From the perspective of data structure differences, this paper summarizes the inherent limitations of traditional DID designs and, drawing on the semiparametric theory, it systematically reviews several Doubly Robust (DR) DID methods. By comparing their identifying assumptions, target parameters, and estimation strategies, we clarify how these methods differ in terms of identification design and usage across different data structures. Moreover, we systematically identify the mechanisms of application of machine learning in these methods, providing actionable recommendations for empirical researchers. Furthermore, using numerical simulations and real-world empirical applications, we assess the actual estimation performance of different methods. The goal of this paper is to offer both theoretical guidance and practical advice to those empirical researchers who seek to understand and apply cutting-edge DID methods from the perspective of data structure differences, helping them grasp the applicability conditions and boundaries of each method and, ultimately, improving the ability to accurately identify causal effects in empirical research.
This study has three key findings. First, the structure of the data determines the ultimately identifiable target parameters. Panel data can be used to identify the Average Treatment Effect on the Treated (ATT); in repeated cross-sectional data, depending on whether the No Compositional Changes (NCC) assumption holds, one can identify either the ATT or other relevant parameters. Second, the traditional Two-Way Fixed Effects(TWFE) model exhibits significant limitations in settings characterized by heterogeneous treatment effects, complex variable relationships, and high-dimensional covariates, making it difficult to obtain consistent estimates of the target parameters. Third, semiparametric approaches such as DR-DID can effectively alleviate the abovementioned limitations by incorporating machine learning and cross-fitting techniques, and identify specific target parameters according to the characteristics of the data distribution.
The contributions of this study to current research are threefold. First, it broadens the perspective of existing DID review articles. While the majority of existing reviews focused on a single data type, this study adopts the perspective of data structure differences to clarify the identification designs and estimation differences of DID methods in repeated cross-sectional versus panel data. This helps empirical researchers have a better grasp of the identifying assumptions and the applicability boundaries of different methods. Second, it expands the theoretical framework of existing DID methodology reviews and complements the current literature. Current domestic methodological reviews on DID largely focused on parametric models. In contrast, this study systematically explains DR-DID designs from a semiparametric perspective and clarifies the application mechanisms of machine learning in these methods, thereby offering empirical researchers a richer and more diverse set of methodological choices. Third, it strengthens the link between theoretical exposition and practical application. Compared to previous research, this study performs a comprehensive comparative analysis of DID designs under different data structures. On this basis, it develops an integrated application guide covering method selection, parameter estimation, and statistical testing, providing more operational practical guidance for empirical researchers to flexibly use cutting-edge DID methods.
Based on these findings, this study proposes the following recommendations for method selection. First, in case of repeated cross-sectional data, researchers should test whether the sample satisfies the NCC assumption. If the NCC assumption does not hold, to avoid estimation bias due to model misspecification and heterogeneous treatment effects, we recommend using the methods of Zimmert (2020) or Sant’Anna and Xu (2026), especially the former that is more efficient. If the NCC assumption holds, then in low-dimensional settings, the various DID methods discussed in this study can be used comprehensively. However, as DR-DID designs are more robust against model misspecification and heterogeneous treatment effects,in high-dimensional settings, we recommend using the more efficient Double Machine Learning (DML) DID method proposed by Zimmert (2020). Second, in case of panel data, if the covariate dimension is low, we recommend the comprehensive useof a variety of estimation methods. In contrast, in settings with heterogeneous treatment effects and complex variable relationships, DR-DID estimates are relatively more robust. If the covariate dimension is high, we recommend using the more efficient DML-DID method.
In synthesis, this study advocates that researchers should ground themselves in the research question and carefully choose estimation methods after comprehensively weighing economic theory and data characteristics. In simple, low-dimensional settings, traditional DID methods remain applicable. However, when the data distribution is more complex, treatment effect heterogeneity is pronounced, and covariates are high-dimensional and with intricate relationships, the identification designs represented by DR-DID exhibit stronger adaptability and robustness. | | 查看全文 相关附件： 下载数据代码附录 |
| | |
