Improved Guarantees for Heterogeneous Treatment-Effect Estimation via Matrix Completion

Anay Mehrotra; Manolis Zampetakis; Phuc Tran; Van H. Vu

arxiv: 2605.30319 · v1 · pith:OPAI72OLnew · submitted 2026-05-28 · 📊 stat.ML · cs.AI· cs.DS· cs.LG· math.ST· stat.TH

Improved Guarantees for Heterogeneous Treatment-Effect Estimation via Matrix Completion

Anay Mehrotra , Phuc Tran , Van H. Vu , Manolis Zampetakis This is my paper

Pith reviewed 2026-06-29 05:17 UTC · model grok-4.3

classification 📊 stat.ML cs.AIcs.DScs.LGmath.STstat.TH

keywords heterogeneous treatment effectsmatrix completioncausal inferencepanel datalow-rank approximationrow-wise error boundsperturbation theory

0 comments

The pith

A matrix completion estimator achieves row-wise ℓ₂ error Õ(√(1/n + n/m²)) for heterogeneous treatment effects without knowing propensities.

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

The paper shows how to estimate each unit's individual treatment effect over time by completing a low-rank matrix of unit-time outcomes under unknown assignment probabilities. Existing matrix-completion results only control average errors across all units, which is too coarse for heterogeneous effects. The new estimator is simple and efficient, and its analysis rests on a sharp new bound that controls error separately for each row. A sympathetic reader would care because many policy or medical questions require knowing how an intervention affects specific groups rather than the population average. If the bound holds, practitioners can obtain per-unit guarantees that scale with the number of units and observation times in a concrete way.

Core claim

Under standard low-rankness and regularity assumptions on the treatment-effect matrix and observation pattern, there exists a computationally efficient estimator that recovers each row's average treatment effect to row-wise ℓ₂ error Õ(√(1/n + n/m²)) without any knowledge of the propensities; the analysis supplies the first sharp row-wise ℓ₂ perturbation bound for low-rank approximation.

What carries the argument

The sharp row-wise ℓ₂-perturbation bound for low-rank matrices, which directly controls the per-unit error in the completed treatment-effect matrix.

If this is right

Per-unit treatment-effect estimates become reliable enough for individualized policy decisions rather than only aggregate averages.
The estimator runs in polynomial time and requires no propensity estimation step.
The same row-wise bound can be plugged into any downstream causal quantity that depends on individual rows.
Existing spectral and Frobenius perturbation results are complemented by a matching row-wise guarantee.

Where Pith is reading between the lines

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

The same row-wise control may improve guarantees in other panel-data problems such as demand estimation or sensor calibration where accuracy per entity matters.
When m grows faster than n the dominant term becomes 1/√n, suggesting that adding more time periods yields diminishing but still positive returns for per-unit precision.
If the low-rank assumption is only approximate, the error bound could be used as a diagnostic to detect when additional regularization or model expansion is needed.

Load-bearing premise

The treatment-effect matrix must be low-rank and the observation pattern plus propensities must satisfy standard regularity conditions.

What would settle it

A concrete low-rank matrix and observation pattern obeying the regularity conditions where the row-wise ℓ₂ error of the estimator exceeds Õ(√(1/n + n/m²)) by more than a constant factor.

read the original abstract

A central goal of modern causal inference is estimating heterogeneous treatment effects to answer questions like "how does an intervention affect each unit," rather than only on average. We study this problem with panel-data where we observe $n$ units across $m$ times under unknown, non-uniform treatment assignments. The data in this setting is naturally represented as a matrix of all unit--time treatment effects. Estimating heterogeneous treatment effects can then be expressed as obtaining a good estimation of each row's average in this matrix. This allows us to formulate the problem as matrix completion, which can be solved under natural low-rankness assumptions. However, existing matrix-completion guarantees are not powerful enough to get meaningful bounds for the per-row guarantee required for estimating the heterogeneous treatment effect; roughly speaking, they are only useful for estimating average treatment effect bounds, as also illustrated in a recent line of work. We give a simple, computationally efficient estimator that, without knowledge of the propensities and under standard low-rankness and regularity assumptions, achieves a row-wise $\ell_2$ error of $\tilde{O}(\sqrt{\frac{1}{n} + \frac{n}{m^2}})$. Technically, our analysis establishes the first sharp row-wise $\ell_2$-perturbation bound for low-rank approximation, complementing existing spectral-, Frobenius-, and entrywise perturbation theory.

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's main advance is a claimed sharp row-wise ℓ₂ perturbation bound that yields per-unit error rates for heterogeneous treatment effects under low-rank panel data.

read the letter

The core takeaway is that this work supplies a row-wise ℓ₂ error bound of Õ(√(1/n + n/m²)) for the treatment-effect matrix via a simple estimator that avoids needing propensity scores. This moves past the average-treatment-effect style guarantees common in prior matrix-completion papers on causal inference.

The paper does a clean job of stating the low-rank plus regularity assumptions up front and framing the technical step as the first sharp row-wise ℓ₂ perturbation result, sitting alongside existing spectral, Frobenius, and entrywise bounds. The estimator itself is described as computationally efficient, which is a practical plus if the analysis holds.

Soft spots are mostly about scope rather than internal contradictions. The n/m² term means the bound tightens only when the number of time periods grows faster than the number of units; in balanced or short panels the guarantee may stay loose. The low-rank assumption on the treatment-effect matrix is standard but still strong, and the paper does not appear to provide extensive checks on how sensitive the result is when that assumption is mildly violated. The abstract positions the bound as new relative to the cited literature, but any referee would want to see the full proof to confirm sharpness and the absence of hidden constants that grow with dimension.

This is a paper for researchers who already work with matrix methods in causal inference or who care about row-wise rather than aggregate error bounds. A reader who needs per-unit guarantees in panel settings could extract value from the rate and the estimator form. The work looks grounded enough on its own terms to merit sending out for serious refereeing rather than a desk reject.

Referee Report

0 major / 1 minor

Summary. The manuscript addresses heterogeneous treatment effect estimation from panel data with n units observed over m time periods under unknown, non-uniform propensities. It formulates the problem as low-rank matrix completion and proposes a simple, computationally efficient estimator that achieves a row-wise ℓ₂ error of Õ(√(1/n + n/m²)) without requiring knowledge of the propensities, under standard low-rankness and regularity assumptions on the observation pattern. As a technical contribution, the analysis establishes the first sharp row-wise ℓ₂-perturbation bound for low-rank approximation, complementing existing spectral, Frobenius, and entrywise results.

Significance. If the stated bound and its sharpness hold, the work is significant for supplying the first per-row guarantees tailored to heterogeneous (rather than average) treatment effects. This directly addresses a limitation of prior matrix-completion theory, which the abstract notes is mainly useful for average-effect bounds. The estimator's computational simplicity and independence from propensity knowledge are practical strengths, and the row-wise perturbation result is a clean technical advance with potential use beyond causal inference.

minor comments (1)

The abstract uses Õ notation for the error bound; the full manuscript should explicitly state the logarithmic factors hidden by the Õ and confirm they do not depend on unknown quantities such as the rank or condition number.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their thoughtful summary and for recognizing the significance of the row-wise guarantees and the new perturbation bound. The recommendation of 'uncertain' is noted, but no specific major comments were provided in the report for us to address point by point.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central claim is a new row-wise ℓ₂ perturbation bound for low-rank matrix approximation, derived under explicitly stated low-rankness and regularity assumptions on the treatment-effect matrix, observation pattern, and propensities. The abstract positions this as complementing (rather than depending on) prior spectral/Frobenius/entrywise theory, with the estimator presented as computationally efficient and not relying on propensity knowledge. No load-bearing step reduces by construction to a fitted parameter, self-citation chain, or renamed input; the derivation chain is self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review based solely on abstract; full paper may contain additional assumptions. The central claim rests on low-rankness of the treatment-effect matrix and regularity conditions on observations and propensities.

axioms (2)

domain assumption The matrix of unit-time treatment effects is low-rank
Invoked to enable matrix completion; stated as standard assumption in abstract.
domain assumption Regularity assumptions on data, observation pattern, and propensities hold
Required for the row-wise bound; referenced but not detailed in abstract.

pith-pipeline@v0.9.1-grok · 5796 in / 1326 out tokens · 23693 ms · 2026-06-29T05:17:12.928839+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 2 canonical work pages

[1]

Synthetic Difference-in-Differences

[AAHI+21] Dmitry Arkhangelsky, Susan Athey, David A. Hirshberg, Guido W. Imbens, and Stefan Wager. “Synthetic Difference-in-Differences”. In: American Economic Review 111.12 (2021), pp. 4088–4118 (cit. on pp. iii, vi). [ABDI+21] Susan Athey, Mohsen Bayati, Nikolay Doudchenko, Guido Imbens, and Khashayar Khosravi. “Matrix Completion Methods for Causal Pane...

2021
[2]

Entrywise eigenvector analysis of random matrices with low expected rank

Proceedings of Machine Learning Research. PMLR, 2023, pp. 3821–3826. url: https://proceedings. mlr.press/v195/agarwal23c.html (cit. on pp. iii–v). [AFWZ20] Emmanuel Abbe, Jianqing Fan, Kaizheng Wang, and Yiqiao Zhong. “Entrywise eigenvector analysis of random matrices with low expected rank”. In: Annals of statistics 48.3 (2020), p. 1452 (cit. on p. v). [...

2023
[3]

Always Valid Inference: Continuous Monitoring of A/B Tests

url: https:// books.google.co.in/books?id=FPkN0AEACAAJ (cit. on pp. vi, vii). [IR15] Guido W. Imbens and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press, 2015 (cit. on pp. ii, vi, vii). [JKPW22] Ramesh Johari, Pete Koomen, Leonid Pekelis, and David Walsh. “Always Valid Inferenc...

work page doi:10.1145/3716496 2015
[4]

User-Friendly Tail Bounds for Sums of Random Matrices

Proceedings of Machine Learning Research. PMLR, 2016, pp. 1670–1679 (cit. on p. ii). [Tro12] Joel A. Tropp. “User-Friendly Tail Bounds for Sums of Random Matrices”. In: Foundations of Computational Mathematics 12.4 (2012), pp. 389–434 (cit. on p. vii). [Tro15] Joel A. Tropp. “An Introduction to Matrix Concentration Inequalities”. In: Foundations and Trend...

work page doi:10.1142/s021919972550035x 2016

[1] [1]

Synthetic Difference-in-Differences

[AAHI+21] Dmitry Arkhangelsky, Susan Athey, David A. Hirshberg, Guido W. Imbens, and Stefan Wager. “Synthetic Difference-in-Differences”. In: American Economic Review 111.12 (2021), pp. 4088–4118 (cit. on pp. iii, vi). [ABDI+21] Susan Athey, Mohsen Bayati, Nikolay Doudchenko, Guido Imbens, and Khashayar Khosravi. “Matrix Completion Methods for Causal Pane...

2021

[2] [2]

Entrywise eigenvector analysis of random matrices with low expected rank

Proceedings of Machine Learning Research. PMLR, 2023, pp. 3821–3826. url: https://proceedings. mlr.press/v195/agarwal23c.html (cit. on pp. iii–v). [AFWZ20] Emmanuel Abbe, Jianqing Fan, Kaizheng Wang, and Yiqiao Zhong. “Entrywise eigenvector analysis of random matrices with low expected rank”. In: Annals of statistics 48.3 (2020), p. 1452 (cit. on p. v). [...

2023

[3] [3]

Always Valid Inference: Continuous Monitoring of A/B Tests

url: https:// books.google.co.in/books?id=FPkN0AEACAAJ (cit. on pp. vi, vii). [IR15] Guido W. Imbens and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press, 2015 (cit. on pp. ii, vi, vii). [JKPW22] Ramesh Johari, Pete Koomen, Leonid Pekelis, and David Walsh. “Always Valid Inferenc...

work page doi:10.1145/3716496 2015

[4] [4]

User-Friendly Tail Bounds for Sums of Random Matrices

Proceedings of Machine Learning Research. PMLR, 2016, pp. 1670–1679 (cit. on p. ii). [Tro12] Joel A. Tropp. “User-Friendly Tail Bounds for Sums of Random Matrices”. In: Foundations of Computational Mathematics 12.4 (2012), pp. 389–434 (cit. on p. vii). [Tro15] Joel A. Tropp. “An Introduction to Matrix Concentration Inequalities”. In: Foundations and Trend...

work page doi:10.1142/s021919972550035x 2016