arxiv: 2604.27187 · v1 · submitted 2026-04-29 · 💰 econ.EM

Recognition: unknown

Treatment-effect heterogeneity and interactive fixed effects: Can we control for too much?

Bruno Ferman, Marcelo Fernandes, Murilo Cardoso

Pith reviewed 2026-05-07 10:21 UTC · model grok-4.3

classification 💰 econ.EM

keywords interactive fixed effectstreatment effect heterogeneitybad controlpanel dataaverage treatment effectmulticollinearity

0 comments

The pith

Interactive fixed effects estimators can fail to recover the average treatment effect on treated units when heterogeneity follows a linear factor structure.

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

This paper examines the interactive fixed effects estimator in panel data with heterogeneous treatment effects. It shows that when treatment-effect differences follow a linear factor structure, the estimator absorbs that variation into its fixed effects. The result is a bad-control problem that prevents correct recovery of the average treatment effect among treated units. When factors are time-invariant or loadings are unit-invariant, multicollinearity can cause identification to break down entirely. Methods that exclude treated units after treatment when estimating factors avoid these problems.

Core claim

If the treatment-effect heterogeneity admits a linear factor structure, the IFE estimator could fail to recover the average treatment effect on the treated units. The problem arises because the interactive fixed effects absorb the heterogeneity in the treatment effect, creating a bad-control problem. With time-invariant factors or unit-invariant loadings in the treatment effect heterogeneity, identification may further break down due to multicollinearity. These problems are not present in alternative estimation methods that exclude treated units in post-treatment periods from the factor estimation.

What carries the argument

The interactive fixed effects (IFE) estimator and its absorption of treatment-effect heterogeneity when that heterogeneity has a linear factor structure, which turns the factors into bad controls.

If this is right

The IFE estimator may produce inconsistent estimates of the average treatment effect on the treated under factor-structured heterogeneity.
Multicollinearity arises and blocks identification when the heterogeneity includes time-invariant factors or unit-invariant loadings.
Estimators that exclude treated units from post-treatment periods when constructing factors remain consistent for the ATT.
The structure of treatment heterogeneity determines whether interactive fixed effects are suitable controls in panel settings.

Where Pith is reading between the lines

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

Applied researchers should compare IFE results against methods that isolate factor estimation from treated observations when heterogeneity is suspected to have factor patterns.
The finding extends to other factor-augmented estimators in causal inference where unobserved heterogeneity may overlap with treatment variation.
It raises the question of how to test for factor-structured treatment heterogeneity before choosing an estimator.

Load-bearing premise

Treatment-effect heterogeneity follows a linear factor structure that can be absorbed by the interactive fixed effects.

What would settle it

A Monte Carlo simulation or empirical application in which treatment effects are generated from a known linear factor model, the IFE estimate of the ATT differs from the true value, and an estimator that drops post-treatment treated observations from factor estimation recovers the correct value.

read the original abstract

This paper studies the interactive fixed effects (IFE) estimator in a panel-data setting with heterogeneous treatment effects. We show that, if the treatment-effect heterogeneity admits a linear factor structure, the IFE estimator could fail to recover the average treatment effect on the treated units. The problem arises because the interactive fixed effects absorb the heterogeneity in the treatment effect, creating a \textit{bad-control} problem. With time-invariant factors or unit-invariant loadings in the treatment effect heterogeneity, identification may further break down due to multicollinearity. These problems are not present in alternative estimation methods that exclude treated units in post-treatment periods from the factor estimation.

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

The paper flags a conditional but real bias in IFE estimators when treatment-effect heterogeneity has a linear factor structure, because the factors absorb the heterogeneity and create a bad-control problem.

read the letter

This paper identifies a specific way the interactive fixed effects estimator can miss the average treatment effect on the treated. When the heterogeneity in treatment effects follows a linear factor structure, the estimated factors and loadings pick up that variation along with the untreated outcomes, which biases the ATT. The same setup produces multicollinearity if the factors are time-invariant or the loadings are unit-invariant. The authors derive this directly from the model equations and note that dropping treated units from the post-treatment factor estimation sidesteps the issue. That mechanism is the new piece; prior IFE work did not spell out this absorption channel under factor-structured heterogeneity. The argument is clean and stays within the stated assumptions, so the central claim holds up on its own terms. The main limitation is that everything turns on the heterogeneity actually having that linear factor form, which the paper treats as a plausible but not universal case. Without simulations it is hard to gauge how large the bias tends to be in typical panels, but the theoretical warning is still useful on its own. Readers who apply IFE to policy panels with possible heterogeneous effects will want to see this. It is the sort of targeted limitation note that belongs in the literature and deserves a serious referee, mainly to check the derivations and perhaps add some Monte Carlo checks on magnitude.

Referee Report

1 major / 2 minor

Summary. The paper studies the interactive fixed effects (IFE) estimator in panel data with heterogeneous treatment effects. It shows that if the treatment-effect heterogeneity admits a linear factor structure, the IFE estimator fails to recover the average treatment effect on the treated (ATT) because the interactive fixed effects absorb the heterogeneity into the estimated factors and loadings, creating a bad-control problem. Under time-invariant factors or unit-invariant loadings in the heterogeneity, multicollinearity can cause identification to break down. These issues are avoided by alternative estimators that exclude treated units in post-treatment periods from the factor estimation step.

Significance. If the result holds, the paper makes a useful contribution to the econometrics literature on panel causal inference by identifying a specific, assumption-driven failure mode for the popular IFE estimator. The direct derivation of the bias from the model equations and factor-structure assumption on heterogeneity provides a clear, falsifiable condition under which over-controlling produces bias rather than robustness. This should prompt applied researchers to consider whether their treatment-effect heterogeneity is plausibly factor-structured before defaulting to IFE, and it usefully contrasts with exclusion-based alternatives.

major comments (1)

The central claim that absorption produces a bad-control bias for the ATT is load-bearing and follows from the joint estimation of factors/loadings on the full panel under the linear factor structure for heterogeneity. The manuscript should present the explicit model equations and the resulting bias expression (likely in the theoretical section following the abstract) so that readers can verify the absorption mechanism without relying solely on the verbal description.

minor comments (2)

The abstract states that 'these problems are not present in alternative estimation methods...' A one-sentence description of one such exclusion-based estimator would improve accessibility for readers who do not reach the main text.
Notation for the treatment-effect heterogeneity factors, loadings, and the ATT should be introduced with a clear table or equation block early in the paper to avoid any ambiguity when the multicollinearity cases are discussed.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading, positive assessment, and constructive suggestion. We agree that explicitly presenting the model equations and bias expression will improve transparency and verifiability of the central absorption mechanism. We will incorporate this change in the revised manuscript.

read point-by-point responses

Referee: The central claim that absorption produces a bad-control bias for the ATT is load-bearing and follows from the joint estimation of factors/loadings on the full panel under the linear factor structure for heterogeneity. The manuscript should present the explicit model equations and the resulting bias expression (likely in the theoretical section following the abstract) so that readers can verify the absorption mechanism without relying solely on the verbal description.

Authors: We agree that the absorption mechanism is central and that explicit equations will allow readers to verify the bias directly. In the revision we will add, immediately after the abstract or in the opening theoretical section, the full data-generating process (including the linear factor structure imposed on treatment-effect heterogeneity), the IFE estimator applied to the full panel, and the closed-form bias expression for the ATT that results from absorption. This addition makes the bad-control problem and the role of joint factor estimation transparent without altering any of the paper's conclusions or requiring new assumptions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained from model equations

full rationale

The paper derives a conditional bias result for the IFE estimator under an explicit assumption that treatment-effect heterogeneity follows a linear factor structure. This follows directly from the joint estimation of factors/loadings on the full panel (including treated units post-treatment), producing absorption into the interactive fixed effects and a bad-control issue for the ATT. No step reduces a prediction to a fitted input by construction, invokes a self-citation as the sole justification for a uniqueness claim, or renames an empirical pattern. The argument is presented as a possibility under stated conditions rather than a universal result, and the logic is internal to the model setup without external load-bearing citations required for the core claim.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The claim rests on the standard interactive fixed effects model for outcomes plus the additional assumption that treatment-effect heterogeneity itself follows a linear factor structure; no free parameters or new entities are introduced.

axioms (2)

domain assumption The outcome equation follows an interactive fixed effects factor model.
Standard setup for IFE estimators, invoked to define the estimator whose properties are analyzed.
domain assumption Treatment-effect heterogeneity admits a linear factor structure.
The key condition under which the absorption and multicollinearity problems are shown to arise.

pith-pipeline@v0.9.0 · 5398 in / 1340 out tokens · 54197 ms · 2026-05-07T10:21:59.809634+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

8 extracted references · 2 canonical work pages

[1]

Abadie, A., Diamond, A., Hainmueller, J., 2015, Comparative politics and the synthetic control method, American Journal of Political Science 59(2), 495–510

Abadie, A., Diamond, A., Hainmueller, J., 2010, Synthetic control methods for compar- ative case studies: Estimating the effect of California’s tobacco control program, Journal of the American Statistical Association 105(490), 493–505. Abadie, A., Diamond, A., Hainmueller, J., 2015, Comparative politics and the synthetic control method, American Journal o...

work page arXiv 2010
[2]

Borusyak, K., Jaravel, X., Spiess, J., 2024, Revisiting event-study designs: robust and efficient estimation, Review of Economic Studies 91(6), 3253–3285

Ben-Michael, E., Feller, A., Rothstein, J., 2021, The augmented synthetic control method, Journal of the American Statistical Association 116(536), 1789–1803. Borusyak, K., Jaravel, X., Spiess, J., 2024, Revisiting event-study designs: robust and efficient estimation, Review of Economic Studies 91(6), 3253–3285. Callaway, B., Sant’Anna, P. H., 2021, Diffe...

work page arXiv 2021
[3]

Firpo, S., Possebom, V., 2018, Synthetic control method: Inference, sensitivity analysis and confidence sets, Journal of Causal Inference 6(2), 20160026

Ferman, B., Pinto, C., 2021, Synthetic controls with imperfect pretreatment fit, Quanti- tative Economics 12(4), 1197–1221. Firpo, S., Possebom, V., 2018, Synthetic control method: Inference, sensitivity analysis and confidence sets, Journal of Causal Inference 6(2), 20160026. 17 Gobillon, L., Magnac, T., 2016, Regional policy evaluation: Interactive fixe...

2021
[4]

Lemma A.1.Consider the PO model in(1), withk 0 known and fixed

Recall that SSE(α,F) = PN i=1(Y i −αD i)′M F (Y i −αD i) and thatM F =I T −P F = I T −F(F ′F) −1F ′ =I T −F F ′/T. Lemma A.1.Consider the PO model in(1), withk 0 known and fixed. It then follows uniformly that SSE(α,F) = NX i=1 D′ iAiM F AiDi + NX i=1 λ′ 0,iF ′ 0M F F 0λ0,i + 2 NX i=1 D′ iA′ iM F F 0λ0,i + NX i=1 e′ iM F ei +o p(N T) whereA i = diag(αi)−α...

2009
[5]

The first result shows thatS N T(α,F) converges uniformly to ˜SN T(α,F) for some bounded set of (α,F), whereas the last two demonstrate that the minimum of ˜SN T(α,F) is at (γ, ˘F). Lemma A.2.The difference betweenS N T(α,F)and ˜SN T(α,F) = 1 N T NX i=1 D′ iAiM F AiDi +λ ′ 0,iF ′ 0M F F 0λ0,i + 2D′ iAiM F F 0λ0,i is uniformly small in probability: i.e.,S ...

2009
[6]

As such, there exists a rotation matrixHsuch thatF ∗ = ˘F Hgiven that ˘D(F ∗)>0

(a) Ifβ ∗ = 0, thenθ ∗ = vec(M F∗ ˘F) must equal zero to ensure that ˜SN T(β∗,F ∗) = 0, which would imply imply that ˘Fbelongs to the linear subspace spanned byF ∗. As such, there exists a rotation matrixHsuch thatF ∗ = ˘F Hgiven that ˘D(F ∗)>0. (b) Assume, by contradiction, that ˘D(F ∗) = 0 and ˘D( ˘F)>0. Letd 1:T = (d1, d2, . . . , dT )′ and consider th...

2009
[7]

We drawF α,t andδ α,i from independent standard normal distributions, holding their values fixed acrossR= 10,000 replications

More specifically, we setλ α,i =δ α,i − ¯δα + (−1,1) ′ for treated units 23 (i∈ T) andF α,t =f α,t − ¯f α + (1,2) ′ for all post-treatment periods (t > T 0), where ¯δα = 1 N1 P i∈T δα,i, and ¯F α = 1 T1 P t>T0 F α,t. We drawF α,t andδ α,i from independent standard normal distributions, holding their values fixed acrossR= 10,000 replications. It then follo...

2017
[8]

Figure A.1 illustrates the finite-sample behavior of the estimators under these DPGs in settings withk α ∈ {1,2,3}

and idiosyncratic errors drawn from mutually independent standard normal distributions. Figure A.1 illustrates the finite-sample behavior of the estimators under these DPGs in settings withk α ∈ {1,2,3}. In these DGPs, the IFE estimator of ¯αexhibits multimodal distributions. The problem is more severe for realizations of the data in whichD( ˆF)≈0, meanin...

2017