| 黄炜,郑姜崇杰,王一平.重复截面数据下的计量分析工具 ——拟事件研究法的理论逻辑与中国实践[J].数量经济技术经济研究,2026,(9):218-241 | | 重复截面数据下的计量分析工具 ——拟事件研究法的理论逻辑与中国实践 | | The Pseudo-event Study Approach with Repeated Cross-sectional Data: Identification, Implementation, andApplications in China | | | | DOI： | | 中文关键词: 拟事件研究法 重复截面 个体化冲击 | | 英文关键词: Pseudo-event Study Repeated Cross-sectional Data Idiosyncratic Shocks | | 基金项目: | | | 中文摘要: | | 在缺乏长期追踪面板数据的背景下，如何利用重复截面数据识别个体化冲击的动态效应，是中国微观实证研究面临的一个重要方法问题。本文立足中国重复截面数据资源较为丰富的现实条件，系统阐述拟事件研究法的识别逻辑，并梳理一套通过可观测特征匹配重构个体事件前后路径的分析框架。随后在潜在结果框架下明确其估计对象、反事实构造方式与关键识别假设，通过数值模拟考察匹配预测误差、群组构建偏差和异质性处理效应错配等偏误来源，提出事件时点预测、匹配设定敏感性、预定协变量平衡和面板结果对照等诊断方法。本文还系统讨论了该方法在生育、迁移、健康、就业和家庭决策等中国研究场景中的适用条件与边界。 | | 英文摘要: | | Event-study designs have emerged as a central tool to estimate the way in which individuals and households respond to shocks over time. However, they usually require panel data that follow the same units both before and after treatment. This requirement is restrictive in several empirical settings, especially in cases where long and representative panels are unavailable, too short, or too small for subgroup analysis. This paper offers a methodological discussion of the pseudo-event study approach, which combines the logic of event studies with pseudo-panel construction from repeated cross-sectional data. The key idea of this paper is to recover an event-time path even when different individuals are observed in different survey waves. For a treated individual observed after a certain event, such as first childbirth, this method examines earlier cross-sections for untreated individuals with comparable predetermined characteristics, and uses their outcomes to approximate the missing pre-event states of the treated individual. By linking these synthetic pre-event observations with actual post-event observations, researchers can estimate the dynamic treatment effects in an event-study framework. This paper clarifies how this approach relates to conventional event studies and Deaton-style synthetic cohort methods. While standard event studies use panel variation to compare treated units with suitable counterfactual trends, synthetic cohort methods aggregate repeated cross-sections into groups defined by stable characteristics. Pseudo-event studies extend these ideas by using cross-sectional matching to impute individual event histories. This is especially useful when the timing of a life-cycle event can be predicted by observed variables such as age, education, hukou status, cohort, province, or family structure, but the data do not contain genuine longitudinal histories. The paper then sets out the key identification conditions. In addition to parallel trends and no anticipation, pseudo-event studies require a conditional independence condition: after conditioning on predetermined characteristics, matched untreated observations must provide valid counterfactual outcomes for treated individuals before the event. This requirement highlights the distinction between statistical predictability and behavioral anticipation. Strong statistical predictability is needed to construct credible synthetic histories; however, this does not necessarily imply that individuals strategically adjust their behavior before treatment. The design is credible only when the matching variables have sufficient predictive power, common support is adequate, group composition is stable, and unobserved determinants of treatment timing do not systematically affect potential outcomes. To operationalize these conditions, this paper conducts numerical simulations and provides implementation procedures. The results of these simulations show that conventional cross-sectional cohort approaches can understate persistent treatment effects, because they observe only post-event outcomes and lack a true pre-event benchmark. When the matching variables capture the determinants of event timing, pseudo-event studies closely reproduce the results from a standard panel-data event study. When important variables, such as hukou status, education, or province, are omitted, the reconstructed pre-trends become biased and the estimated dynamic effects deviate from the true path. These exercises demonstrate both the value of the proposed method and the need for prediction checks, balance tests, rolling-window comparisons, and sensitivity analyses. Furthermore, this paper discusses the applications in the Chinese data context. Using fertility and child penalties as the leading example, it explains why repeated cross-sectional sources, such as population censuses, one-percent population surveys, time-use surveys, CGSS, and other large-scale micro datasets, can support dynamic causal analysis. Evidence based on Chinese census-style data shows how the proposed method can recover changes in female labor supply around childbirth and identify spatial and temporal heterogeneity in child penalties. Additional validation using CFPS panel data compares pseudo-event estimates with standard event-study estimates and examines predetermined covariates, providing evidence on the empirical reasonableness of the reconstructed pseudo-panel. Taken together, this paper transforms recent empirical practice into a clearer methodological framework. It defines the estimand, explains the data construction process, separates identification assumptions from implementation choices, and provides diagnostics for assessing when the method is likely to work. Rather than treating pseudo-event studies as a substitute for genuine panel data, this paper presents them as a complementary tool for data-constrained environments. The proposed framework expands the applied microeconometric toolkit to investigate fertility, labor supply, migration, health shocks, household structure, and policy-relevant life-cycle events in China and in other settings with limited longitudinal data. | | 查看全文 相关附件： 下载数据代码附录 |
| | |
