Простой и сложный метод разности разностей

Abstract

The paper presents extensions of the popular difference‐in‐differences approach (DD) from 2×2 design on multiple time‐period, multiple groups, fuzzy DD, non‐staggered treatment and approaches to measure distributional treatment effect. The paper describes assumptions for consistent estimation of the treatment effect by two‐way fixed effects model (TWFE) and presents the problem leading to inconsistent estimates justifying the application of alternative estimators. The paper briefly introduces methods developing DD for multiple‐period multiple‐group cases based on TWFE and alternative approaches. The proposed techniques allow treatment evaluation in the frame of DD when canonical TWFE leads to inconsistent estimates. Some approaches allow replacement of the well‐known parallel trend assumption (PTA) for a conditional PTA or time randomisation. The paper refers to implementations of these methods in Stata and R. Simulation modelling demonstrates that the stated properties of the alternative estimators are not always reliable.