CoChat

Two-Way Fixed Effects, the Two-Way Mundlak Regression, and Difference-in-Differences Estimators

Open access
Wooldridge, Jeffrey M. · 2021 · SSRN / The Econometrics Journal

The DOI matches its Crossref record; 39 of the 42 references it cites were found in a registry or an index, and 3 are not checked yet; 10 of the 30 sources its claims cite were checked for support.

Abstract

I establish the equivalence between the two-way fixed effects (TWFE) estimator and an estimator obtained from a pooled ordinary least squares regression that includes unit-specific time averages and time-period specific cross-sectional averages, which I call the two-way Mundlak (TWM) regression. This equivalence furthers our understanding of the anatomy of TWFE, and has several applications. The equivalence between TWFE and TWM implies that various estimators used for intervention analysis – with a common entry time into treatment or staggered entry, with or without covariates – can be computed using TWFE or pooled OLS regressions that control for time-constant treatment intensities, covariates, and interactions between them. The approach allows considerable heterogeneity in treatment effects across treatment intensity, calendar time, and covariates. The equivalence implies that standard strategies for heterogeneous trends are available to relax the common trends assumption. Further, the two-way Mundlak regression is easily adapted to nonlinear models such as exponential models and logit and probit models.

Verified 2026-10-10 · 3 references not checked yet

Overview

CoChat Summary

TL;DR

The paper shows that the two-way fixed effects (TWFE) estimator is numerically equivalent to a pooled OLS regression that includes unit-specific time averages and time-period-specific cross-sectional averages—the two-way Mundlak (TWM) regression. It uses this equivalence to show that flexible pooled OLS/extended TWFE, cohort imputation, random effects, and event-study estimators produce identical average treatment effects on the treated under standard staggered difference-in-differences assumptions.

Key findings

  • TWFE and TWM are algebraically equivalent: adding both unit-specific time averages and period-specific cross-sectional averages to pooled OLS reproduces the TWFE coefficients on time-varying regressors.
  • Under no anticipation, conditional parallel trends, linearity, and random sampling, cohort imputation, pooled OLS on cohort dummies, random effects, TWFE, and Borusyak et al. imputation yield the same ATTs; the flexible leads-and-lags estimator is equivalent to Sun and Abraham (2021) and the regression-based Callaway and Sant’Anna (2021) estimator.
  • In the Walmart application using 1,288 U.S. counties over 1977–1999, the lags-only and leads-and-lags weighted average effects are 9.4% and 7.3%, respectively; allowing cohort-specific linear trends reduces the average effect to about 2.6% (t ≈ 2.71), with effects in the first six years between 2% and 3% and a peak just under 5% nine years out.
  • The paper provides aggregation formulas for overall and exposure-time ATTs and shows how to allow heterogeneous cohort trends within the same regression framework.

Methods

  • Theoretical econometric paper deriving algebraic equivalences between TWFE, TWM, pooled OLS, random effects, imputation, and event-study estimators, with proofs in an appendix.
  • Identification is developed for staggered interventions under SUTVA, no bad controls, no anticipation, conditional parallel trends, linearity, and random sampling across units.
  • Empirical application uses Brown and Butts (2025) county data: 1,288 counties, 1977–1999, log retail employment, three time-constant 1980 controls, 14 treated cohorts, county-clustered standard errors, and Stata margins for exposure-time effects; Fig. 2 uses 1,000 bootstrap replications.

Limitations

  • The paper states that the equivalences assume a balanced panel and time-constant covariates; with time-varying covariates, POLS, TWFE, and RE are no longer identical unless the Mundlak equation is expanded, and with unbalanced panels the equivalences no longer hold.
  • The paper states that inference can be problematic with few treated units per treated cohort, and that the Walmart application’s parallel trends assumption is clearly violated, so the lags-only and leads-and-lags estimates cannot be relied on as good causal estimates; cohort-specific linear trends reduce but do not perfectly eliminate pre-trends.
  • The paper notes that linearity of conditional means in covariates is assumed, and that the leads-and-lags estimator can exacerbate bias if parallel trends violations carry into treated periods.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.2139/ssrn.3906345

Cites

39 of 42 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

6 Valid · 7 Not valid · 17 Not checked · 184 Own results and interpretations

109 more claims

Create a free account to see all 214 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 52
  • It may state two separate claims as one. 61
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 42 references

Cited by

Items 301–400 of 759

Metadata

Authors
Wooldridge, J. M.
Affiliations
Michigan State University (US)
Corresponding author
Wooldridge, Jeffrey M.
Venue
SSRN / The Econometrics Journal
Publisher
Elsevier BV
Published
2021
Volume
69
Issue
5
Pages
2545-2587
ISSN
0377-7332 (ISSN-L)
OpenAlex
W4413754779
Cited by
714
References
39
Citations by year
2026: 218 · 2025: 199 · 2024: 155 · 2023: 78 · 2022: 45 · 2021: 11
FWCI
225.81
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Empirical Economics (publishedVersion, cc-by)
  • SSRN Electronic Journal (submittedVersion)
Funders
  • Agricultural and Applied Economics Association
  • University College London
  • College of Engineering, Michigan State University
Topics
  • Physical Sciences › Mathematics › Statistics and Probability › Advanced Causal Inference Techniques
  • Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Spatial and Panel Data Analysis
  • Physical Sciences › Mathematics › Statistics and Probability › Statistical Methods and Bayesian Inference
Keywords
two-way fixed effects, difference-in-differences, estimator equivalence, treatment effect heterogeneity, average treatment effect on the treated, parallel trends assumption, intervention analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.