{"id":"8d0cdbc4-066c-45a9-bb0e-4dc124e25358","arxiv_id":"2501.00583","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RobustPALMRT generalizes PALMRT to shift-invariant fitting procedures, enabling finite-sample valid tests for robust and quantile regression and for dispersion effects.","lead":"A group of statisticians extends a recent permutation-based test so it works with robust and quantile regression, including data-driven scale estimation, while keeping a finite-sample guarantee on false positives. The new tests reveal dispersion differences in Long-COVID immune profiles that standard mean-based tests miss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DispersionPALRMT is never shown to satisfy the Conditions of Theorem 1, so the advertised finite-sample guarantee for this quantile-regression extension is unproven.","rationale":"The reader's conditional verdict is essentially right: the paper's main theorem and its proof appear sound, and the Huber-Huber procedure is verified in Appendix A, but the advertised DispersionPALRMT extension lacks an explicit verification of Conditions 1 and 2. I diverge from the reader's specific wording, because permutation-equivariant quantile regression residuals plus group-wise sorted summaries are plausibly invariant under joint row permutation; the issue is that the paper never proves this, and the tie-breaking/non-uniqueness convention for quantile regression is only asserted to preserve Condition 1. That makes the concern a real gap in support for one of the three headline contributions, but not a demonstrated counterexample. The secondary issue flagged by the reader, the abstract's unqualified 'controls type I error rate' versus the theorem's 2α bound, is a genuine presentation overstatement but does not threaten the theorem itself; I treat it as part of the recommended revision rather than the primary load-bearing concern. A direct finite-n check of the two conditions would settle whether the gap is purely expository or reflects a substantive flaw in DispersionPALRMT.","tokens_in":22391,"tokens_out":29684,"duration_ms":281877,"concrete_test":"Implement Algorithm 2 exactly as specified, with n=6 and p=1 (intercept plus one demographic covariate), and enumerate all σ∈S_n. For random Y and X, compare M(Yσ, Xσ, [Z,Zπ]σ) with M(Y, X, [Z,Zπ]) up to numerical tolerance, and separately check Condition 1 by adding [Z,Zπ]γ to Y for random γ. If any σ or γ fails, Theorem 1 does not cover DispersionPALRMT and its type I error should be re-simulated under H0 at α=0.05 with at least 10^4 replicates. If all checks pass, the gap is a missing proof rather than a false claim, and a short lemma verifying the two conditions would resolve it.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central guarantee is Theorem 1, which requires the model-fitting procedure M to satisfy Condition 1 (shift invariance under adding [Z,Zπ]γ to Y) and Condition 2 (joint row-permutation invariance). Appendix A verifies both for the Huber-Huber estimator via Algorithms A1/A2. No analogous verification is given for the quantile-regression DispersionPALRMT procedure of Section 4.3 / Algorithm 2. The formal M at the start of Section 4.3 returns grouped residual summaries (r_{qLow,Control}, r_{qHigh,Case}, etc.) without specifying how the two quantile fits are combined or how non-unique quantile-regression solutions are resolved; the paper only asserts in a general technical note that standard software conventions preserve Condition 1. Because the case/control labels X are used both as covariates and as grouping variables, Condition 2 is not immediate: jointly permuting (Y,X,[Z,Zπ]) changes which rows fall in each group, and the proof requires M to be exactly invariant under this relabeling. If either Condition 1 or Condition 2 fails for the implemented DispersionPALRMT, Theorem 1 does not apply, and the abstract's claim that the RobustPALMRT framework 'controls type I error rate' for this dispersion extension is unsupported. This is the load-bearing gap in the paper's advertised scope; the Huber-Huber location test itself appears to satisfy the conditions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RobustPALMRT, a permutation-based framework for testing whether a covariate X has an effect on a response Y after adjusting for control covariates Z. The main theoretical result, Theorem 1, states that if the model-fitting procedure M satisfies two invariance conditions (shift invariance under adding [Z,Zπ]γ to Y, and joint row-permutation invariance) and the errors are exchangeable, then the resulting p-value satisfies P(pvalue ≤ α) ≤ 2α. The authors verify these conditions for a Huber regression with MAD scale estimation (Appendix A) and use the framework to develop Huber-Huber RobustPALMRT. They also propose DispersionPALRMT, a quantile-regression-based test for differences in dispersion between case and control groups, and apply it to Long-COVID immune profiling data.","tokens_in":22730,"tokens_out":13475,"duration_ms":131178,"significance":"If the central guarantee is correct, Theorem 1 substantially extends the PALMRT framework of Guan (2024) to robust regression with estimated scale, quantile regression, and multiple model fits, while retaining a finite-sample type I error bound under only exchangeability. The separation of model fitting from model evaluation is a useful contribution, and the simulations suggest meaningful power gains under heavy-tailed and skewed errors. The paper provides a full proof of the theorem in Appendix C and a detailed verification for the Huber-Huber procedure, with code and reproducibility information mentioned. However, the advertised scope includes DispersionPALRMT, and the formal verification for that procedure is missing; this is a load-bearing gap that needs to be addressed before the paper can be accepted.","major_comments":[{"comment":"The formal definition of M for DispersionPALRMT does not match the structure required by the framework. The displayed M packs residuals from both fits Y ∼ X+Z+Zπ and Y ∼ Xπ+Z+Zπ into one tuple, whereas the framework comparison (6) evaluates M separately at (Y,X,[Z,Zπ]) and (Y,Xπ,[Z,Zπ]). With the displayed definition, M(Y,Xπ,[Z,Zπ]) would fit Y ∼ Xπ+Z+Zπ and Y ∼ Xππ+Z+Zπ, which is not the intended permuted fit. Algorithm 2 implements the intended comparison with a scalar function of a single fit, but the paper never states M in that form nor verifies Conditions 1 and 2 for it. A proof of shift invariance and joint row-permutation invariance for quantile-regression residual summaries, including a tie-breaking convention for non-unique quantile solutions, is needed before Theorem 1 can be invoked for DispersionPALRMT. Without it, the abstract's finite-sample type I error claim for the dispersion extension is unsupported.","section":"Section 4.3 / Algorithm 2"},{"comment":"The paragraph on non-unique minimizers states that 'standard software packages do this [preserve Condition 1]' without proof. For Theorem 1 to apply to the implemented algorithms, the tie-breaking convention must satisfy both Condition 1 and Condition 2, not merely Condition 1. A convention that selects, for instance, the first optimum in a fixed row order would break joint row-permutation invariance. The authors should specify the exact tie-breaking rule used in Algorithms 1 and 2 (e.g., the R or Python implementation) and verify that it preserves both invariance conditions.","section":"Section 4.2, non-unique minimizers"}],"minor_comments":[{"comment":"The tuple defining M lists 'rqLow,Control' twice and omits 'rqHigh,Control'; it should list both quantiles for both groups, e.g., (r_qLow,Control, r_qHigh,Control, r_qLow,Case, r_qHigh,Case).","section":"Section 4.3, M definition"},{"comment":"The notation 1/|Xτ=j| is ambiguous; it should be 1/n_j with n_j = |{i : Xτ_i = j}|, and the set notation should be used consistently.","section":"Section 4.3 / Algorithm 2"},{"comment":"In the statement of Theorem 2, the condition on ̃ω is written as '̃ω(M1, M2) + ̃ω(M2, M2) = 1'; it should be '̃ω(M1, M2) + ̃ω(M2, M1) = 1'.","section":"Appendix C, Theorem 2"},{"comment":"The claim that this is 'the first test for robust regression parameters with scale estimation that is valid in a distribution-free setting without moment assumptions in finite-samples' is strong; the authors should verify this against recent literature (e.g., invariance-based and permutation-based tests) and, if it stands, provide a precise comparison.","section":"Section 2.3"},{"comment":"When reporting the 'highly conservative' type I error of DispersionPALRMT (e.g., 0.004 in the normal-errors setting with n=100), the paper should report Monte Carlo standard errors or confidence intervals so that readers can assess the precision of these empirical rates.","section":"Section 5.2"},{"comment":"The text 'Figure 1a shows a normal QQ plot ... we see that there is shows strong right-skewness' contains a typo ('there is shows'); it should read 'we see that there is strong right-skewness'.","section":"Section 1, Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The central theorem for the Huber-Huber procedure appears sound and the verification in Appendix A is careful. The main blocker is the DispersionPALRMT section: the formal M is mis-specified, and Conditions 1 and 2 are not verified for the implemented quantile-regression procedure. This is fixable within the paper's scope: either add a rigorous verification for DispersionPALRMT (including tie-breaking) or explicitly restrict the theoretical claim to the Huber-Huber method and present DispersionPALRMT as an empirically validated extension. I recommend major revision rather than rejection because the core framework is valuable and the gap is localized."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that the main contribution is solid, but the advertised scope goes beyond what the theory actually covers. The extension of PALMRT to robust regression with scale estimation (the Huber-Huber procedure) is real, useful, and appears to satisfy the conditions of the main theorem. The DispersionPALRMT quantile-regression extension, however, is presented as if it were covered by the same theorem, and it is not. No verification is given that Condition 2 (joint row-permutation invariance) holds for the grouped residual summaries, and on inspection it fails: the grouping is by the same X that gets permuted, so the output changes under a joint row permutation. The finite-sample type I error guarantee therefore does not apply to that procedure.\n\nWhat is genuinely new: the shift-invariance condition replacing least-squares projection, the proof that Huber regression with MAD scale estimation fits the conditions (Appendix A, Lemma 1), and the finding that separating model fitting from evaluation—specifically using a robust loss in the evaluation step—can substantially improve power. The simulations are thorough, the proof of Theorem 1 is detailed and builds on Guan's framework, and the code is available. The Long-COVID application is a nice demonstration of the location test.\n\nWhere I would push back: the abstract says the framework \"controls type I error rate for finite-samples\" without mentioning the 2α factor. The theorem gives a 2α bound, and the paper itself notes the factor can be necessary in extreme cases. That is a minor overstatement, though the text is honest about it. More seriously, the DispersionPALRMT gap is load-bearing for one of the three advertised contributions. The authors should either prove the conditions for a dispersion-type procedure (perhaps with within-group permutations or a different invariant) or clearly state that the theoretical guarantee does not cover this extension. As it stands, a reader could walk away thinking the type I error of their quantile-regression dispersion test is provably controlled when it is not.\n\nWho is this for: people working on permutation tests, robust regression, and heavy-tailed biological data. The Huber-Huber location test is ready to use; the dispersion test is promising and empirically conservative but needs a theoretical fix or a caveat. I would send this to peer review—it deserves careful referees—but I would expect a major revision before publication.","headline":"Robust location extension of PALMRT is solid; the dispersion extension is overclaimed because the theorem's conditions are not verified.","tokens_in":23185,"tokens_out":4658,"would_cite":true,"duration_ms":47092,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62J05","62G35"],"pacs":[],"model":"deepseek-v4-flash","headline":"RobustPALMRT gives a finite-sample type I error bound for testing a covariate in a linear model, under only exchangeable errors, while allowing robust and quantile regression fits with estimated scale.","keywords":["permutation test","robust regression","quantile regression","finite-sample type I error","exchangeability","partial correlation","dispersion test","Long-COVID"],"falsifier":"Simulate the null model $Y = Z\\theta + \\epsilon$ with exchangeable, symmetric, heavy-tailed errors and an independent case/control label, run DispersionPALRMT at nominal level $\\alpha$, and check whether its rejection rate exceeds $2\\alpha$; if it does, the grouped residual summaries break the permutation-invariance condition and Theorem 1 does not apply to that procedure.","tokens_in":22155,"feed_emoji":"📊","tokens_out":7736,"duration_ms":70855,"temperature":0.7,"pith_summary":"This paper introduces RobustPALMRT, a permutation-based test for whether a covariate of interest has any effect on a response after adjusting for control covariates in a linear model. The central theoretical claim is that the test controls the type I error rate in finite samples, with $P(\\mathrm{pvalue} \\le \\alpha) \\le 2\\alpha$, assuming only that the errors are exchangeable and that the chosen fitting method is shift-invariant and permutation-equivariant. This matters because heavy-tailed, skewed, or outlier-laden noise, common in biological data, inflates false positives for classical tests like the F-test. The paper shows that robust regression with scale estimation fits inside the framework, and that separating model fitting from model evaluation can improve power. It also introduces DispersionPALRMT, a quantile-regression variant that tests for differences in dispersion and finds new immune differences in Long-COVID patients.","feed_headline":"Robust permutation test tames heavy-tailed regression noise","feed_subtitle":"Extending PALMRT to robust and quantile fits gives finite-sample error control and more power under skew.","key_machinery":"The load-bearing object is the pair-of-permutations comparison array $T(\\pi_1,\\pi_2;\\epsilon) = \\omega(M(\\epsilon, X_{\\pi_2}, [Z_{\\pi_1}, Z_{\\pi_2}]))$, whose defining symmetry $T(\\sigma\\circ\\pi_1,\\sigma\\circ\\pi_2;\\epsilon_\\sigma) = T(\\pi_1,\\pi_2;\\epsilon)$ transfers exchangeability of the errors into exchangeability of columns of the $B\\times B$ comparison matrix $A_{\\pi_1,\\pi_2} = I(T(\\pi_1,\\pi_2;\\epsilon)\\ge T(\\pi_2,\\pi_1;\\epsilon))$. A weighted-column tournament argument then bounds the probability that the identity column's average is small, giving the $2\\alpha$ guarantee. Conditions 1 and 2 on the fitting method $M$ -- shift invariance in the $[Z, Z_\\pi]$ directions and joint row-permutation equivariance -- are what let $M(Y, X, [Z, Z_\\pi])$ be replaced by $M(\\epsilon, X, [Z, Z_\\pi])$ under the null, so the proof goes through for robust and quantile regressions rather than only OLS projections.","core_discovery":"The core discovery is an extension of the permutation-augmented linear regression test (PALMRT) to any fitting method that satisfies two verifiable conditions: adding any shift of the control covariates must not change the output, and permuting the data rows must permute the output in the same way. Under the null model $Y = Z\\theta + f(X,\\epsilon)$ with exchangeable $\\epsilon$, the paper proves that the permutation p-value defined by comparing fits to $(Y, X, [Z, Z_\\pi])$ and $(Y, X_\\pi, [Z, Z_\\pi])$ satisfies $P_{H_0}(\\mathrm{pvalue} \\le \\alpha) \\le 2\\alpha$. Because robust regression with a Huber loss and MAD scale estimation satisfies the conditions, the result covers robust and quantile regression, including hyper-parameter tuning and scale estimation. The paper further shows empirically that evaluating models with a robust loss improves power under heavy tails and skew even when the model is fit by OLS, and that the quantile-based DispersionPALRMT detects scale differences in the Long-COVID immune data that location-focused tests miss.","pith_inferences":["Editorial inference: the $2\\alpha$ bound is loose in simulations; an adaptive recalibration using the empirical null distribution of p-values, or an exact correction based on the comparison matrix's column sums, could close the gap and restore nominal $\\alpha$ without losing the finite-sample guarantee.","Editorial inference: DispersionPALRMT's grouped case/control summaries appear to violate the row-permutation condition, so its guarantee is not covered by Theorem 1; a symmetric dispersion statistic based on all pairwise absolute residual differences might recover the finite-sample control while keeping the power advantage.","Editorial inference: the separation of fit and evaluation suggests a cheap robustness upgrade for existing permutation tests: replace the $L_2$ score with a robust loss, which simulations show can raise power substantially without changing how the model is estimated.","Editorial inference: because the theory only needs exchangeability, extending it to multivariate responses (as the paper notes) and to regularization paths for high-dimensional $X$ is a natural next step; high-dimensional $Z$ remains the known bottleneck."],"forward_implications":["Using a level $\\alpha/2$ cutoff gives strict finite-sample control at $\\alpha$ for any method satisfying the two conditions, even with estimated scale parameters.","Analysts can fit with OLS but evaluate with a robust loss and keep the guarantee while gaining power under heavy-tailed errors; the evaluation choice matters independently of the fit.","DispersionPALRMT provides a finite-sample permutation test for differences in spread between groups, not just location, after adjusting for covariates.","Any future fitting method verified to satisfy shift invariance and permutation equivariance inherits the same type I error bound, so the method is not tied to least squares."],"supporting_citations":[{"why":"Supplies the original PALMRT construction, the comparison-matrix symmetry argument, and the exchangeability proof that RobustPALMRT extends.","marker":"Guan (2024)"},{"why":"Provides the first finite-sample distribution-free test for linear models with exchangeable errors, the baseline that later permutation methods improve on.","marker":"Lei & Bickel (2021)"},{"why":"Contributes the jackknife-plus style comparison argument and the tournament bound used in the proof of the $2\\alpha$ type I error guarantee.","marker":"Barber et al. (2021)"},{"why":"Defines quantile regression, which underlies DispersionPALRMT's dispersion testing through conditional quantiles.","marker":"Koenker & Bassett (1978)"},{"why":"Supplies the robust M-estimation framework and iteratively reweighted least squares background used for robust model fitting.","marker":"Ronchetti & Huber (2009)"},{"why":"Gives the specific Huber regression with MAD scale estimation algorithm whose shift invariance and permutation equivariance are verified in the appendix.","marker":"Venables & Ripley (2013)"},{"why":"Contributes the Long-COVID immune profiling dataset that motivates the method and is reanalyzed with RobustPALMRT and DispersionPALRMT.","marker":"Klein et al. (2023)"},{"why":"Provides the studentized Breusch-Pagan test used as the comparison baseline for DispersionPALRMT in heteroskedasticity simulations.","marker":"Koenker (1981)"}],"fun_headline_variants":["RobustPALMRT: permutation tests for heavy-tailed and skewed data","Finite-sample error control for robust and quantile regression tests","Robust evaluation loss boosts power in permutation association tests","DispersionPALMRT detects scale shifts robustly in Long-COVID data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The test's guarantee rests on the errors being exchangeable after adjusting for controls, and on the chosen fitting method ignoring shifts in the control covariates and treating all rows symmetrically; the paper verifies these properties for robust regression but flags that the quantile-based dispersion procedure does not clearly satisfy the row-permutation condition.","fun_headline_variants_meta":{"raw":{"variants":["RobustPALMRT: permutation tests for heavy-tailed and skewed data","Finite-sample error control for robust and quantile regression tests","Robust evaluation loss boosts power in permutation association tests","DispersionPALMRT detects scale shifts robustly in Long-COVID data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1600,"prompt_tokens":985,"completion_tokens":615,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":538}},"tokens_in":601,"tokens_out":615,"duration_ms":7168,"temperature":1.0,"reasoning_tokens":538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:47:56.592334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the null model $Y = Z\\theta + \\epsilon$ with exchangeable, symmetric, heavy-tailed errors and an independent case/control label, run DispersionPALRMT at nominal level $\\alpha$, and check whether its rejection rate exceeds $2\\alpha$; if it does, the grouped residual summaries break the permutation-invariance condition and Theorem 1 does not apply to that procedure.","supporting_citations":[{"cited_title":"F., Candès, E","cited_arxiv_id":null,"evidence_quote":"Contributes the jackknife-plus style comparison argument and the tournament bound used in the proof of the $2\\alpha$ type I error guarantee."}],"review_version":1}