CausalSmith — CSWP-2026-028 · ← back to the paper · arXiv · GitHub

AI referee report

on “Private Estimation and Honest Inference for Conditional Treatment Effects with Rough Nuisance Functions”

Overall score 7.8/10 5 findings — 4 minor · 1 nit

The paper delivers a substantial, carefully scoped private minimax characterization with honest inference, while several presentation and cross-reference defects require editorial repair.

Referee report of 1 Oct 2026. This report was written for an earlier draft; the paper has been revised since, so some findings may already be addressed. One automated referee’s estimate — not peer review. Produced by a single automated referee in the pipeline's review stage. It can be wrong, and it is not peer review.

Summary

For a precisely specified scalar binary observational model, the paper characterizes pure-private pointwise absolute-error risk and uniformly honest connected-interval expected length by the same three-term benchmark. Its substantive contributions are density-robust occupancy-weighted attainment and matching causal testing comparisons across privacy regimes. Accepting the verified statements, the surrounding claims faithfully preserve their scope, and the closest literature receives a careful comparison. The paper merits publication after focused improvements to organization, wording, and references.

Strengths

The estimation and interval conclusions cover every admissible measurable design density and the full binary arm-mean range.
The occupancy-weighted ratio provides a concrete connection between nuisance approximation, usable pairs, replacement sensitivity, and privacy noise.
The lower-bound comparisons apply to all admissible private randomized procedures, with an explicit simultaneous-containment argument for honest interval length.
The manuscript distinguishes the closest nonprivate comparison's experiment and attainment conditions from its own model.
The limitations candidly quantify the extremely conservative interval constants and distinguish exact-real guarantees from implementation certification.

Findings

minor Introduction; Causal records and observational sampling

The introduction references sec:related-work, but the supplied Related work section has no corresponding label. The sentence saying that coordinate conventions are 'recorded in Appendix A' uses a manually numbered reference and points away from the conventions actually presented immediately afterward in the main text.

Suggested fix

Add \label{sec:related-work} to the Related work section. Correct the coordinate-convention pointer to its actual location, using an exact structural label and cref, or remove the pointer after consolidating the definitions. Check all reader-facing structural references for the same requirements.

minor Observation model, identification, and decision criteria

The observation model repeatedly introduces the same record coordinates through an opening paragraph, separate outcome, treatment, and covariate definitions, and another full-coordinate definition. This interrupts the progression from observational sampling to identification and the decision problems.

Suggested fix

Present the full causal record and observed marginal together in one concise main-text passage. Relocate the standalone coordinate conventions and elementary probability-space definitions to an appendix while preserving their anchored environments and mathematical content.

minor Discussion and limitations, before Limitations and future work

The sentences 'These are order comparisons for the specified model, rather than finite-sample design recommendations' and 'This argument explains the common frontier without reducing honest inference to a point-estimation slogan' use contrastive non-coverage framing outside the explicitly titled limitations subsection. The latter also introduces an unnecessarily dismissive description of an alternative argument.

Suggested fix

Write affirmatively that the regime comparisons characterize optimal dependence on sample size and privacy budget, and that simultaneous containment establishes the interval converse directly. Place any qualification about practical design recommendations in Limitations and future work.

minor Discussion and limitations

Two consecutive paragraphs explain essentially the same interval mechanism: transferred coverage produces simultaneous containment, and connectedness converts separation into length. This repeats explanations already supplied in the introduction and main-results section.

Suggested fix

Merge the two discussion paragraphs into one concise account connecting the interval argument to the paper's broader statistical interpretation. Retain the detailed explanation in the main-results section.

nit A private release based on within-cell pairs; Population pair moments and occupancy

The phrases 'following frozen algorithm' and 'following frozen statement' introduce unexplained editorial terminology into the mathematical exposition.

Suggested fix

Remove 'frozen' and state directly that b denotes the public error-envelope bias component in the inversion algorithm and the cellwise bias certificate in the occupancy lemma.

Review history

Round	Date	Score	Findings
1	1 Oct 2026	6.5	7
2	1 Oct 2026	7.6	7
3 · current	1 Oct 2026	7.8	5

Every referee draw is kept: scores are noisy, so the history is shown rather than only the last or the best pass.

Read against paper.tex sha256 a27b2894a769.