CoChat

Difference-in-Differences Estimators of Intertemporal Treatment Effects

Open access
de Chaisemartin, Clément, D’Haultfœuille, Xavier · 2026 · Review of Economics and Statistics

The DOI matches its Crossref record, with a small difference in title or year; 82 of the 82 references it cites were found in a registry or an index; 28 of the 40 sources its claims cite were checked for support.

Abstract

We study treatment-effect estimation using panel data. The treatment may be nonbinary, nonabsorbing, and the outcome may be affected by treatment lags. We make a parallel-trends assumption and propose event-study estimators of the effect of being exposed to a weakly higher treatment dose for ℓ periods. We also propose normalized estimators that estimate a weighted average of the effects of the current treatment and its lags. We also analyze commonly used two-way fixed-effects regressions. Unlike our estimators, they can be biased in the presence of heterogeneous treatment effects. A local-projection version of those regressions is biased even with homogeneous effects.

Overview

CoChat Summary

TL;DR

The paper develops difference-in-differences estimators for panel data where treatment can be nonbinary, nonabsorbing, and have lagged effects, without imposing homogeneous treatment effects. It proposes reduced-form and normalized event-study estimators plus a cost-benefit parameter, shows common TWFE dynamic regressions can be biased or contaminated, and revisits banking deregulation effects.

Key findings

  • The proposed \(DID_\ell\) and normalized \(DID^n_\ell\) estimators are conditionally unbiased under no-anticipation and parallel trends conditional on the period-one treatment; \(DID^n_\ell\) estimates a weighted average of the current treatment and its first \(\ell-1\) lags, and a weighted sum of \(DID_\ell\)’s yields a cost-benefit parameter interpretable as an average total effect per unit of treatment.
  • In a census of 100 highly cited AER papers from 2015–2019, 26 estimate TWFE regressions, but only four have a binary-and-staggered treatment; among the 20 non-binary/non-staggered papers, 11 estimate only static TWFE and nine estimate dynamic effects using intensity-period FE, local-projection, or distributed-lag regressions.
  • The paper shows those three common TWFE dynamic methods may not estimate a convex combination of the paper’s \(\delta_{g,\ell}\) effects, may violate no-sign reversal, and local-projection panel regressions are contaminated by other exposure lengths and can be biased or wrong-signed even with constant effects.
  • In the banking-deregulation application, local-projection coefficients are positive and significant through \(\ell=5\), insignificant at \(\ell=6,7,9\), and significantly negative at \(\ell=8\); the paper’s \(DID_1=0.043\) (s.e. \(0.035\)), \(DID_3=0.081\) (s.e. \(0.049\)), \(DID_5=0.148\) (s.e. \(0.064\)), placebos jointly insignificant (\(p=0.400\)), and the test that \(\ell \mapsto DID^n_\ell\) is constant is not rejected (\(p=0.328\)). For house prices, \(DID_1=0.003\) (s.e. \(0.004\)), \(DID_4=0.016\) (s.e. \(0.009\)), \(DID_5=0.026\) (s.e. \(0.010\)), and placebos are jointly insignificant (\(p=0.703\)).

Methods

  • Panel of \(G\) groups observed over \(T\) periods, with dynamic potential outcomes allowing current outcomes to depend on current and lagged treatments; identifying assumptions include no anticipation, parallel trends for the status-quo outcome conditional on the period-one treatment, and a no-crossing condition, plus a design restriction requiring groups with the same period-one treatment and variation in first-treatment-change timing.
  • Estimators compare switchers to not-yet-switchers with the same period-one treatment; the paper aggregates group-level \(DID_{g,\ell}\)’s into event-study effects, proposes normalized versions, placebo estimators, and asymptotic inference as \(G\to\infty\) under independent groups and regularity conditions, with confidence intervals conservative unless groups are i.i.d.
  • Application revisits Favara and Imbs (2015) using 1994–2005 county-level banking deregulation data; treatment is the number of restrictions lifted (0–4); eight states never deregulate, 33 deregulate once, eight twice, and one three times; 38 of 42 deregulating states first do so in 1995–1998; \(DID_1\) and \(DID_8\) apply to 905 and 773 counties, respectively, and only three placebos apply to more than 50% of the 905 counties whose treatment changes at least once.

Limitations

  • The paper states that the initial-conditions problem is not accounted for in the baseline notation; Assumption 1 rules out continuous period-one treatment, extremely non-persistent treatment, and no or universal treatment change; Assumption 4 may be implausible for some baseline-treatment values; \(\delta_\ell\)’s interpretation is complicated outside Design 1; \(\delta^n_\ell\) averages different lags and individual lag slopes are not estimable under Assumptions 3–4; cost-benefit analysis requires converting outcomes to monetary equivalents and assumes linear, non-stochastic treatment costs; inference is conservative in the i.n.i.d. setup unless groups are i.i.d.; and the local-projection issues are specific to the panel Regression 2, absent in time-series local projections and in Dube et al. (2023).
  • The paper states placebos may fail to detect parallel-trends violations large enough to substantially bias \(DID_\ell\); in the application, only three placebos apply to more than 50% of the 905 counties, parallel trends can only be tested over three years whereas eight years are needed for \(DID_8\), loan-volume placebos are relatively large and imprecisely estimated, and \(DID_9\)–\(DID_{11}\) are not reported because they apply to small subsamples (357, 238, and 1 county).
  • Evident from the text: several application estimates are imprecise (e.g., \(DID_1=0.043\) with s.e. \(0.035\); \(DID^n_1\) is much larger but imprecisely estimated); the paper proves weights sum to less than one in binary-and-staggered designs but only finds this in the non-binary/non-staggered application without a formal proof; and the design-conditional parameters may only partially uncover the structural function mapping current, lagged, and future treatments to outcomes.

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

Integrity signals

Registry

Crossref · 10.1162/rest_a_01414

Cites

82 of 82 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Effect of banking deregulations on loan volume and houses prices.
Figure 1 Effect of banking deregulations on loan volume and houses prices.From the PDF

Verification

Claims the paper makes

19 Valid · 10 Not valid · 7 Not checked · 214 Own results and interpretations

140 more claims · 6 with their cited sources checked

Create a free account to see all 250 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 52
  • Probably not a substantial claim. 60
  • Borderline: one of the tests for a substantial claim is unclear. 1
  • The quote only partly states this claim. 2

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 82 references

Cited by

Items 201–300 of 746

Metadata

Authors
de Chaisemartin, C., & D’Haultfœuille, X.
Affiliations
  • Sciences Po Paris, Economics Department
  • Centre de Recherche en Économie et Statistique (FR)
Venue
Review of Economics and Statistics
Publisher
MIT Press
Published
2026-07
Volume
108
Issue
4
Pages
863-880
ISSN
  • 0034-6535 (ISSN-L)
  • 1530-9142
OpenAlex
W3041109582
Cited by
740
References
81
Citations by year
2026: 147
FWCI
63.78
Type
journal article
Language
en
Open access
Locations
  • The Review of Economics and Statistics (publishedVersion)
  • SSRN Electronic Journal (publishedVersion)
  • National Bureau of Economic Research (publishedVersion)
  • SSRN Electronic Journal (acceptedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • HAL (submittedVersion, cc-by-nc-nd)
  • HAL (submittedVersion, cc-by-nc)
  • HAL (submittedVersion)
  • arXiv (Cornell University)
Funders
  • European Commission
  • Agence Nationale de la Recherche (ANR-17-CE26-0015, ANR-17)
  • HORIZON EUROPE European Research Council (101043899)
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
difference-in-differences, parallel trends assumption, treatment effect heterogeneity, two-way fixed effects, local projections, panel data

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

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.