Difference-in-Differences Estimators of Intertemporal Treatment Effects

Cl\'ement de Chaisemartin; Xavier D'Haultf{\oe}uille

arxiv: 2007.04267 · v14 · submitted 2020-07-08 · econ.EM

Difference-in-Differences Estimators of Intertemporal Treatment Effects

Cl\'ement de Chaisemartin , Xavier D'Haultf{\oe}uille This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-05-24 13:48 UTCgrok-4.3open to challenge →

classification econ.EM

keywords difference-in-differencesevent studytreatment effectspanel dataintertemporal effectsparallel trendsheterogeneous effectstwo-way fixed effects

0 comments

The pith

Event-study difference-in-differences estimators recover the effects of sustained higher treatment doses for a given number of periods under parallel trends.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops estimators for panel data settings where a treatment can be non-binary, non-absorbing, and can affect outcomes through current and lagged values. It proposes event-study estimators that identify the causal effect of exposure to a weakly higher treatment dose over ℓ periods. It also introduces normalized estimators that recover a weighted average of the effects of the current treatment and its lags. These estimators remain consistent even when treatment effects vary across units and time periods. In contrast, standard two-way fixed-effects regressions can be biased by such heterogeneity, and a local-projection version of those regressions is biased even when effects are homogeneous.

Core claim

Under a parallel-trends assumption, event-study estimators identify the effect of being exposed to a weakly higher treatment dose for ℓ periods, while normalized estimators identify a weighted average of current and lagged treatment effects; two-way fixed-effects regressions are biased by heterogeneous treatment effects, and their local-projection versions remain biased even under homogeneous effects.

What carries the argument

Event-study and normalized difference-in-differences estimators that compare changes in outcomes for units experiencing different treatment paths over time.

If this is right

Applied researchers can obtain unbiased estimates of dynamic treatment effects in settings with time-varying and non-absorbing treatments.
Normalized estimators deliver a single interpretable summary of the combined effect of a treatment and its lags.
Standard two-way fixed-effects regressions should not be used for intertemporal treatment effect estimation when effects may differ across groups.
Local-projection regressions inherit the same bias problems and cannot be relied upon even in homogeneous-effects cases.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The estimators could be applied to policy evaluations involving gradual or intensity-varying rollouts, such as changes in minimum wages or regulatory exposure.
Extensions might combine these estimators with matching or weighting to relax the parallel-trends assumption in specific ways.
The bias results suggest re-examining many published event-study findings that rely on two-way fixed-effects specifications.

Load-bearing premise

Units with different treatment paths would have experienced parallel trends in the outcome in the absence of those treatment differences.

What would settle it

A Monte Carlo simulation or empirical application in which the proposed event-study estimators recover the true intertemporal effects while two-way fixed-effects estimates diverge, under data generated with heterogeneous treatment effects.

read the original abstract

We study treatment-effect estimation using panel data. The treatment may be non-binary, non-absorbing, 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 $\ell$ 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Abstract sketches event-study and normalized DiD estimators for non-binary lagged treatments but only the abstract is available so the claims cannot be checked.

read the letter

The main point is that this paper outlines event-study estimators for the effect of higher treatment exposure over a fixed number of periods and normalized estimators that recover a weighted average of current and lagged effects, all under a parallel-trends assumption. It also flags that two-way fixed-effects regressions can be biased with heterogeneous effects and that local-projection versions remain biased even with homogeneous effects. That setup targets a real issue in applied panel work where treatments like subsidies or regulations change gradually and outcomes respond with delays. If the estimators are derived cleanly from the stated assumption, they could give applied researchers a more reliable alternative to off-the-shelf regressions in those cases. The parallel-trends condition is the usual one, so the contribution would rest on whether the new estimators actually isolate the desired parameters without extra restrictions. The obvious limitation is that only the abstract exists here. There are no estimator formulas, no bias derivations, no simulation checks, and no application, so it is impossible to see whether the bias results hold in the cases the authors have in mind or whether some heterogeneity patterns are left unaddressed. The direction looks consistent with other recent work on DiD limitations, but that is as far as one can go without the details. This kind of paper would mainly interest applied economists who run event studies on time-varying treatments in policy settings. A reader facing exactly those data structures might find the estimators useful once the full derivations and checks are available. Right now the abstract alone is too thin to cite or to bring to a reading group. I would send the full paper out for peer review if the authors supply the math, some Monte Carlos, and at least one empirical illustration, because the problem is common and the proposed fix is worth testing.

Referee Report

1 major / 0 minor

Summary. The paper studies treatment-effect estimation using panel data where the treatment may be non-binary and non-absorbing and outcomes may depend on treatment lags. Under a parallel-trends assumption, it proposes event-study estimators of the effect of exposure to a weakly higher treatment dose for ℓ periods, normalized estimators that recover a weighted average of current and lagged effects, and shows that two-way fixed-effects regressions can be biased under heterogeneous treatment effects while a local-projection version of those regressions is biased even under homogeneous effects.

Significance. If the proposed estimators are consistent under the stated assumptions and the bias characterizations are correct, the results would be significant for applied work in difference-in-differences settings with intertemporal and heterogeneous effects, by supplying alternatives to commonly used regressions that the abstract indicates can be biased.

major comments (1)

Only the abstract is available; the manuscript contains no estimator definitions, derivations, proofs, or numerical examples. This prevents verification of the central claims that the new estimators are unbiased under the parallel-trends assumption while two-way fixed-effects and local-projection estimators are biased (as asserted in the abstract).

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for their report. We address the major comment below and note a standing limitation.

read point-by-point responses

Referee: Only the abstract is available; the manuscript contains no estimator definitions, derivations, proofs, or numerical examples. This prevents verification of the central claims that the new estimators are unbiased under the parallel-trends assumption while two-way fixed-effects and local-projection estimators are biased (as asserted in the abstract).

Authors: We agree that the full manuscript is required to verify the claims about the event-study estimators (for the effect of exposure to a weakly higher treatment dose for ℓ periods under parallel trends), the normalized estimators (recovering weighted averages of current and lagged effects), and the bias results for two-way fixed-effects and local-projection regressions. Since only the abstract is provided here, we cannot supply the requested definitions, derivations, proofs, or examples in this response. revision: no

standing simulated objections not resolved

Only the abstract is available; the manuscript contains no estimator definitions, derivations, proofs, or numerical examples. This prevents verification of the central claims that the new estimators are unbiased under the parallel-trends assumption while two-way fixed-effects and local-projection estimators are biased (as asserted in the abstract).

Circularity Check

0 steps flagged

No circularity detected from available text

full rationale

Only the abstract is provided, which states that event-study estimators are proposed under a parallel-trends assumption without any equations, derivations, parameter fits, or self-citations shown. No load-bearing steps reduce to inputs by construction, and the description matches a standard derivation from an external identifying assumption. This is the expected non-finding when no technical content is available for inspection.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Central claims rest on the parallel trends assumption as the key identifying condition; no free parameters or invented entities are mentioned.

axioms (1)

domain assumption parallel trends assumption
Explicitly invoked in abstract as foundation for the proposed estimators

pith-pipeline@v0.9.0 · 5606 in / 1139 out tokens · 26186 ms · 2026-05-24T13:48:24.016253+00:00 · methodology

Difference-in-Differences Estimators of Intertemporal Treatment Effects

Core claim

What carries the argument

If this is right

Where Pith is reading between the lines

Load-bearing premise

What would settle it

discussion (0)