Royal Economic Society Annual Conference 2022 Special Issue on The New Difference-in-Differences
本文是《计量经济学杂志》为皇家经济学会2022年会组织的特刊，由Xavier D'Haultfoeuille和Jeffrey Wooldridge介绍双重差分法的最新进展，重点讨论传统双向固定效应模型在复杂处理设计中的局限性及替代估计方法。
Each year, The Econometrics Journal organises a Special Session on a subject of current interest and importance at the Annual Conference of the Royal Economic Society.With these sessions, the journal intends to promote econometric theory and methods of substantive direct or potential value in applications and their actual empirical application.At the Society's online 2022 Conference, Xavier D'Haultfoeuille (CREST-ENSAE) and Jeffrey Wooldridge (Michigan State University) presented in a Special Session on The New Difference-in-Differences.Dif ference-in-dif ferences (DiD) is a popular approach to estimating treatment effects from observational data.In its simplest form, it compares the change in outcomes o v er two periods between a treatment group that switches into treatment in the second period and a control group that remains untreated in both periods.Under the 'parallel-trends' assumption that the treatment group would have seen the same trend in outcomes as the control group if it would not have been treated, such a DiD estimates an average treatment effect on second-period outcomes in the treatment group.In empirical practice, this is often implemented by estimating a tw o-w ay fixed-effects (TWFE) model that specifies mean outcomes to be linear in group and time effects and the effect of treatment.Simply differencing outcomes o v er time kills the group effects, but leaves both the treatment effect (for the treatment group) and any time trend.A second difference, across groups, subsequently isolates the treatment effect.Indeed, the usual (least-squares) TWFE estimator of the slope parameter on the treatment indicator is the DiD estimator.The TWFE model straightforwardly extends to settings with more than two periods and treatments, in which many more treatment paths are possible.For example, if there is a control (never-treated) group, a group that enrols early in treatment, and one that enrols late ('staggered' enrolment), this can be captured by three different time paths of the treatment variable in the TWFE model.This versatility of the TWFE model, together with the clear DiD intuition of the simplest case, may explain its popularity in empirical economics.Recent work in econometrics shows that care should be taken in specifying the TWFE model and interpreting its estimator in these more general cases.The TWFE model is often specified with a constant (across groups and time) slope parameter on the treatment indicator, which is then interpreted as an average treatment effect.This is trivially correct if treatment effects are indeed constant across groups and time, but may be false if they are not.It turns out that, for various common treatment designs, the TWFE treatment effect estimand is a weighted average of groupand time-specific average treatment effects, with weights that may be negative.Consequently, it may not equal a well-defined average treatment effect.In one e xtreme e xample, it may be positive even if all group-and time-specific average treatment effects are negative.In response to this, the recent literature has proposed a variety of alternative estimators that more carefully compose group-and time-specific treatment effects.In his presentation, Xavier D'Haultfoeuille re vie wed this literature, based on the paper with Cl ment de Chaisemartin included in this Special Issue.