The DOI matches its Crossref record; 48 of the 48 references it cites were found in a registry or an index; 17 of the 81 sources its claims cite were checked for support.
Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects
Open accessAbstract
Linear regressions with period and group fixed effects are widely used to estimate treatment effects. We show that they estimate weighted sums of the average treatment effects (ATE ) in each group and period, with weights that may be negative. Due to the negative weights, the linear regression coefficient may for instance be negative while all the ATEs are positive. We propose another estimator that solves this issue. In the two applications we revisit, it is significantly different from the linear regression estimator. (JEL C21, C23, D72, J31, J51, L82)
Overview
CoChat Summary
TL;DR
Two-way fixed effects (FE) and first-difference (FD) regressions estimate weighted sums of group-period average treatment effects, and the weights can be negative, so the coefficient may have the opposite sign from all treatment effects. The authors propose DID_M, an estimator of the average effect in switching cells, and in two applications it differs significantly from the FE/FD estimators.
Key findings
- Under common trends, β_fe and β_fd are weighted sums of ATEs in treated (g,t) cells; weights sum to one but may be negative. The paper’s example: 1.5×1 − 0.5×4 is negative while both components are positive. In Gentzkow et al. data, 40% of weights attached to β_fe are negative; 46% attached to β_fd are strictly negative. In Enikolopov et al., 52.6% of 1999 weights are negative, negative weights sum to −2.26, and σ_fe = 0.91 and σ_fe = 1.23. In the union application, 820 weights are positive and 196 negative, negative weights sum to −0.01, and σ_fe = 0.097.
- Applications: Gentzkow et al.: β_fd = 0.0026 (s.e. 0.0009), β_fe = −0.0011 (s.e. 0.0011), DID_M = 0.0043 (s.e. 0.0015), 66% larger than β_fd, t-stat = 1.69. Union premium: β_fe = 0.107 (s.e. 0.030), β_fd = 0.060 (s.e. 0.032), DID_M = 0.041 (s.e. 0.034), significantly different from β_fe (t = 2.60) and β_fd (t = 2.36); placebo DID_pl_M = 0.094 (s.e. 0.038), t = 2.49; joiners’ effect 0.059, leavers’ effect 0.021.
- Survey: of 337 AER papers published 2010–2012, 33 (9.8%) use FE/FD or closely related regressions; excluding theory papers and lab experiments, 19.1%. Of the 33, 26 are sharp designs and 7 fuzzy; stable groups: 12 yes, 14 presumably yes, 5 presumably no, 2 no.
- DID_M estimates the ATE of switching cells without treatment-effect homogeneity; it is asymptotically normal as G→∞, and the proposed confidence interval is asymptotically conservative.
Methods
- Theoretical decomposition of FE and FD regressions under Assumptions 1–5: balanced panel of groups, sharp design, independent groups, mean independence, and common trends. Extensions to fuzzy designs, non-binary treatments, covariates, and 2SLS are in the Web Appendix.
- Survey of all AER papers 2010–2012, plus two empirical applications: Gentzkow et al. (2011) on newspaper entry/exit and voter turnout, and Vella and Verbeek (1998) on union wage premia. The authors use the twowayfeweights, fuzzydid, and did_multiplegt Stata packages; in the union data they discard changes in union status happening twice in three consecutive years, discarding half of union status changes.
- Asymptotic analysis as the number of groups G tends to infinity, establishing consistency/asymptotic normality of DID_M and conservative confidence intervals.
Limitations
- Stated: DID_M may have larger variance than β_fe, creating a bias-variance trade-off. Assumption 11 is necessary for unbiasedness but not for consistency. The asymptotic variance can only be conservatively estimated, so the confidence interval is asymptotically conservative. DID_M estimates only switchers at the time they switch, not all ATEs.
- Stated: DID_M requires stable groups; in the survey, for about half the papers reading the paper was not enough to assess this, and 5 papers presumably do not have stable groups while 2 do not. In the union application, the placebo suggests a positive pre-trend and that DID_M may overestimate the union premium; the leavers’ effect is close to 0.
- Evident from the text: the main results rely on common trends and mean-independence assumptions that rule out Ashenfelter’s dip; the main text focuses on binary treatment, with non-binary and fuzzy cases relegated to the Web Appendix. Negative weights are a concern when treatment effects are heterogeneous across groups or over time.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
11 Valid · 10 Not valid · 59 Not checked · 183 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“When the data has two groups and two periods, the WaldDID estimand considered therein is equal to β fe and β fd .”
What the sources say
- Partly supportedcites de Chaisemartin and D'Haultfœuille (2018) sourceExperimental — accuracy not yet measured
Section 2.2 defines the two-group, two-period Wald-DID as the outcome DID divided by the treatment DID and says it equals the estimand from a specified IV regression. It does not establish that this estimand equals both β_fe and β_fd; the copy does not define those coefficients or state that equality. Page numbers were not available in our copy. (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Our main result generalizes Theorem 1 in de Chaisemartin and D'Haultfœuille (2018).”
What the sources say
- Supportedcites de Chaisemartin and D'Haultfœuille (2018) source
The source’s Theorem 2.1 section contains a numbered first result about the Wald differences-in-differences estimand, identifying the prior result the citing authors invoke. Whether their own result generalizes it is the citing authors’ claim, not a finding attributed to this source. Our copy labels the theorem 2.1 rather than 1. (LLM, conf=0.82)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“As union status is often measured with error (see, e.g. Freeman, 1984; Card, 1996), we discard changes in union status happening twice in three consecutive years.”
What the sources say
- Supportedcites Freeman, 1984 source
The paper’s “Evidence on Measurement Error” section reports non-negligible misclassification of union status in two surveys. This supports the claim that union status is often measured with error; the citing authors’ decision to discard repeated changes is their own methodological choice. (LLM, conf=0.93)
- Not checked — no full text of the sourcecites Card, 1996 source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“Enikolopov et al. (2011) find that ̂ β fe = 6 . 65 (s.e.= 1 . 40 ).”
What the sources say
- Could not verifycites Enikolopov et al. (2011) sourceExperimental — accuracy not yet measured
Section 5 (Aggregate-level results; page unavailable) describes a model with regional fixed effects, but the retrieved passages do not report a fixed-effects estimate of 6.65 with a standard error of 1.40. Searches for both numbers and the fixed-effects coefficient did not locate the claimed result in our copy. (LLM, conf=0.72)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“This is a sharp design.”
What the sources say
- Could not verifycites Aaronson et al. (2012) sourceExperimental — accuracy not yet measured
Appendix B (under State-level Data; page unavailable) describes regressions of spending and income on the minimum wage, but does not characterize the regression as a sharp design. Searches for “sharp design,” “sharp,” and discontinuity or cutoff terminology found no passage in our copy that establishes that classification. (LLM, conf=0.68)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
131 more claims
Create a free account to see all 263 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 80
- It may state two separate claims as one. 73
- The quote only partly states this claim. 1
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?
The cited source doesn’t support the claim as written.
In the citing paper
“de Chaisemartin and D'Haultføeuille (2020) focuses on recovering simultaneous treatment effects rather than the estimation of dynamic effects, and it does not allow for covariates.”
From How Much Should We Trust Staggered Difference-in-Differences Estimates? (2022) · cites this paper as de Chaisemartin and D'Haultføeuille (2020)
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredThe cited source doesn’t support the claim as written.
In the citing paper
“de Chaisemartin and D'Haultfœuille (2018) and de Chaisemartin and D'Haultfœuille (2020) had already noted that in a model assuming away effects of lagged treatments on the outcome, parallel trends across groups with different baseline treatments essentially rules out time-varying effects.”
From Difference-in-Differences Estimators of Intertemporal Treatment Effects (2026) · cites this paper as de Chaisemartin and D'Haultfœuille (2020)
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“A rapidly growing literature has shown that two-way fixed effects regressions deliver consistent estimates only with strong assumptions about the homogeneity of treatment effects, and may be biased when treatment effects vary over time or by adoption cohort (Cengiz et al., 2019; de Chaisemartin and D'Haultfœuille, 2020; Sun and Abraham, 2021;Goodman-Bacon, 2021; Callaway and Sant'Anna, 2021;Borusy”
From Generative AI at Work (2023) · cites this paper as de Chaisemartin and D'Haultfœuille, 2020
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“As in de Chaisemartin and D'Haultfœuille (2020), exact inference becomes infeasible when treatment effects are heterogeneous, but conservative inference is possible.”
From Revisiting Event-Study Designs: Robust and Efficient Estimation (2024) · cites this paper as de Chaisemartin and D'Haultfœuille (2020)
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Recent work by Goodman-Bacon (15) and de Chaisemartin & D'Haultfoeuille (9) provides a clearer account of how the βFE parameter represents a variance of treatment–weighted combination of underlying heterogeneous effects.”
From Designing Difference-in-Difference Studies with Staggered Treatment Adoption: Key Concepts and Practical Guidelines (2024) · cites this paper as [9]
Experimental — accuracy not yet measured
22 more claims about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
Identification and estimation of local average treatment effects
Imbens, G. W., & Angrist, J. D. (1994) · Econometrica · doi:10.2307/2951620 · Cited by 4,101
An algorithm for quadratic programming
Frank, M., & Wolfe, P. (1956) · Naval Research Logistics Quarterly · doi:10.1002/nav.3800030109 · Cited by 3,030 · cited 1×
“Hence, by Frank and Wolfe ( 1956 ) and the fact that the linear term in the quadratic problem is 0, the solution exists if and only if the set of constraints is not empty.”
Duflo, E. (2001) · American Economic Review · doi:10.1257/aer.91.4.795 · Cited by 1,940
Bustos, P. (2011) · American Economic Review · doi:10.1257/aer.101.1.304 · Cited by 1,552
Autor, D. H. (2003) · Journal of Labor Economics · doi:10.1086/344122 · Cited by 1,534 · cited 1×
“This test differs from the standard event study pretrends test ( see Autor 2003 , which has been shown to be invalid when treatment effects are heterogeneous ( see Abraham and Sun 2018 .”
Estimating the Effect of Training Programs on Earnings
Ashenfelter, O. (1978) · The Review of Economics and Statistics · doi:10.2307/1924332 · Cited by 1,515 · cited 1×
“This rules out the possibility that a group gets treated because it experiences negative shocks, the so-called Ashenfelter’s dip ( see Ashenfelter 1978 .”
Algan, Y., & Cahuc, P. (2010) · American Economic Review · doi:10.1257/aer.100.5.2060 · Cited by 1,286
House Prices, Home Equity–Based Borrowing, and the US Household Leverage Crisis
Mian, A., & Sufi, A. (2011) · American Economic Review · doi:10.1257/aer.101.5.2132 · Cited by 1,161
The Effects of Rural Electrification on Employment: New Evidence from South Africa
Dinkelman, T. (2011) · American Economic Review · doi:10.1257/aer.101.7.3078 · Cited by 1,121
Americans Do IT Better: US Multinationals and the Productivity Miracle
Bloom, N., Sadun, R., & Reenen, J. V. (2012) · American Economic Review · doi:10.1257/aer.102.1.167 · Cited by 1,028
Research works printed but not linked: 2
- de Chaisemartin Clément, 2011, Center for Research in Economics and Statistics Working Paper, 2011
- Freeman Richard B, 1984, ILR Review, 38, 244
Cited by
Difference-in-Differences with multiple time periods
Callaway, B., & Sant’Anna, P. H. C. (2021) · Journal of Econometrics · doi:10.1016/j.jeconom.2020.12.001 · Cited by 9,126
Difference-in-differences with variation in treatment timing
Goodman-Bacon, A. (2021) · Journal of Econometrics · doi:10.1016/j.jeconom.2021.03.014 · Cited by 8,046
Estimating dynamic treatment effects in event studies with heterogeneous treatment effects
Sun, L., & Abraham, S. (2021) · Journal of Econometrics · doi:10.1016/j.jeconom.2020.09.006 · Cited by 5,518
How Much Should We Trust Staggered Difference-in-Differences Estimates?
Baker, A. C., Larcker, D. F., & Wang, C. C. Y. (2022) · Journal of Financial Economics · doi:10.1016/j.jfineco.2022.01.004 · Cited by 3,149
Revisiting Event-Study Designs: Robust and Efficient Estimation
Borusyak, K., Jaravel, X., & Spiess, J. (2024) · Review of Economic Studies · doi:10.1093/restud/rdae007 · Cited by 2,280
Metadata
- Authors
- de Chaisemartin, C., & D’Haultfœuille, X.
- Affiliations
- University of California at Santa Barbara (email: )
- Centre de Recherche en Économie et Statistique (FR)
- Venue
- American Economic Review
- Publisher
- American Economic Association
- Published
- 2020-09-01
- Volume
- 110
- Issue
- 9
- Pages
- 2964-2996
- ISSN
- 0002-8282 (ISSN-L)
- OpenAlex
- W2979586175
- Semantic Scholar
- f2b059a83a75bd2f8fe4b4f29f53b865eedc024f
- Cited by
- 5,031
- References
- 14
- Citations by year
- 2026: 1247 · 2025: 1194 · 2024: 1022 · 2023: 679 · 2022: 483 · 2021: 254 · 2020: 91
- FWCI
- 265.97
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- American Economic Review (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- RePEc: Research Papers in Economics (submittedVersion)
- Funders
- Agence Nationale de la Recherche (ANR-11-IDEX-0003, ANR-17-CE26-0015, ANR-11-LABX-0047)
- 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, treatment effect heterogeneity, average treatment effect, negative weights, treatment effect estimation
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 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects. https://app.cochat.ai/papers/10.1257/aer.20181169
- The paper
- de Chaisemartin, C., & D’Haultfœuille, X. (2020). Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects. American Economic Review. https://doi.org/10.1257/aer.20181169
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.