{"id":"c63d6903-9280-4e23-922f-372dfdf7a7b2","arxiv_id":"2508.14882","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A residual-based knockoff generator for distribution-free, mixed-type data and a local-derivative importance score (MALD) are proposed, supported by simulations and a mouse methylation case study.","lead":"This statistics paper proposes a new way to generate 'knockoff' twin variables for mixed data types without assuming their distribution, plus a new importance score that works with nonlinear models. Simulations claim better false-discovery control and higher power than existing methods, and a mouse DNA methylation analysis finds CpG sites tied to aging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FDR guarantee rests on exact exchangeability of residual-based knockoffs; estimation from the same data breaks this, so the abstract's 'principled FDR control' may be only asymptotic.","rationale":"The reader's weakest assumption exactly identifies the same load-bearing concern: finite-sample exchangeability of residual-estimated knockoffs. My review agrees that the abstract overclaims 'principled FDR control' without stating that the guarantee is asymptotic or approximation-dependent. Since the reader rendered a CONDITIONAL verdict with low confidence, and my concern is arguably a sharper form of the same issue but does not move the verdict, UNCHANGED is appropriate. I could not verify the actual proof or simulations because the supplied text is corrupted, but the concern is grounded in the algorithmic structure described in the abstract and legible fragments. If the paper contains a rigorous finite-sample guarantee under explicit regularity conditions, the test would reveal that the concern does not land; if not, the abstract should be revised to state the mode of convergence.","tokens_in":18422,"tokens_out":4087,"duration_ms":55314,"concrete_test":"Use the authors' rangerKnockoff package on simulated data with p=100 independent null features drawn from a nonlinear heteroscedastic model (e.g., X_j = Z_j U_j with Z_j ~ N(0,1), U_j ~ |N(0,1)|), Y independent of X. Generate knockoffs with default residual fitting on the same n=200 sample, run the knockoff filter at q=0.1 over 500 replications, and compute empirical FDR. Also run an oracle version using the true conditional distribution. If the residual-based FDR exceeds 0.12 while the oracle stays near 0.10, the exchangeability violation is real and the abstract's FDR claim is unsupported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the method achieves better false discovery rate control than existing methods, which requires the knockoff property: for every null feature j, the pair (X_j, Xtilde_j) is exchangeable conditional on X_{-j}. The described residual-based construction estimates the conditional residual function and noise distribution from the same sample used to compute knockoff statistics. Under standard model-X knockoff theory, exact finite-sample FDR control requires the conditional distribution of Xtilde_j given X_{-j} to equal that of X_j. When the residual model is estimated from data, this equality fails in finite samples because the estimated parameters depend on all observations, including X_j; swapping X_j with Xtilde_j changes the joint law. Sample splitting mitigates but does not eliminate the mismatch: if the fitted conditional mean or residual variance differs from the true ones, the marginal distributions differ and exchangeability is only approximate. Thus the guarantee is at best asymptotic or consistency-dependent, while the abstract states 'principled FDR control' without stating regularity conditions or mode of convergence. If residual hyperparameters are tuned on the same data used for selection, overfitting can further inflate FDR. The real-data CpG findings and MALD are secondary; the load-bearing assertion is the nominal FDR guarantee, which rests on this unstated approximate-exchangeability premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a knockoff-generation framework based on conditional residuals, intended to handle heterogeneous, distribution-free data without assuming known feature distributions, and a variable-importance measure called the Mean Absolute Local Derivative (MALD) for nonlinear outcome models. The abstract reports that simulations demonstrate better false discovery rate control and higher power than existing methods, and a DNA methylation application identifies CpG sites associated with aging. However, the supplied full text is a severely corrupted text extraction: the body contains no readable equations, theorems, simulation descriptions, tables, or software links. Consequently, the central methodological claims and their supporting evidence cannot be verified from the submitted material.","tokens_in":18623,"tokens_out":2584,"duration_ms":32206,"significance":"If the claims were established, the paper would make a useful contribution: extending model-X knockoffs beyond known distributions to mixed data types, and offering a nonlinear-compatible importance score. The stated availability of R and Python software is a practical strength. Nevertheless, the load-bearing assertions—principled FDR control and simulation superiority—rest on technical details and experimental evidence that are not accessible in the submitted manuscript. The paper cannot currently be evaluated on its merits, and the lack of any stated regularity conditions for the FDR guarantee is a substantive gap independent of the formatting problem.","major_comments":[{"comment":"The body of the manuscript is unreadable: it consists of repeated mojibake placeholders, with no recoverable equations, theorems, section headings, simulation tables, or algorithm descriptions. Since every central claim in the abstract depends on this content, the submitted version is not reviewable. The authors must provide a clean, correctly encoded PDF or source file so that the technical content can be assessed.","section":"Full text (throughout)"},{"comment":"The abstract claims 'principled control of the false discovery rate' through conditional-residual knockoffs. Standard model-X FDR control requires, for each null feature j, exchangeability of (X_j, Xtilde_j) conditional on X_{-j} under the true joint distribution. If the conditional residual function and noise distribution are estimated from the same sample used for selection, finite-sample exact exchangeability fails; at best one obtains an approximate or asymptotic guarantee. The manuscript appears to state no regularity conditions, convergence mode, or sample-splitting adjustment. This is a load-bearing gap and must be addressed explicitly.","section":"Abstract and knockoff construction (unreadable full text)"},{"comment":"The abstract asserts that 'simulations show that our approach achieves better false discovery rate control and higher power than existing methods.' No simulation setup, replicate count, error bars, baseline methods, or data-generating processes are recoverable from the supplied text. Since these simulations are the primary evidence for the headline claim, the authors must report full experimental details and uncertainty quantification, or the claim should be substantially weakened.","section":"Simulation claims (abstract; no readable simulation section)"}],"minor_comments":[{"comment":"The abstract mentions software availability in R and Python but provides no repository URLs, package names beyond 'rangerKnockoff' and 'MALDimportance', or version identifiers. Please include persistent links or references.","section":"Abstract"},{"comment":"The term 'heterogeneous data' is used without definition. It could mean mixed variable types (continuous, categorical, count) or non-identically distributed observations. Please define the intended scope precisely.","section":"Abstract / introduction"}],"recommendation":"major_revision","confidential_remarks":"The submitted file appears to be a corrupted text extraction; no technical content is readable. In practice this is a desk-reject-level problem. I chose major_revision rather than reject because the issues are verifiability and missing theoretical/simulation detail rather than an identified internal contradiction, and the idea may be salvageable with a clean resubmission. The editor should require a clean PDF before any substantive review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper may be a real contribution, but right now it's impossible to check. The full text I received is mostly replacement characters with a stray header from another arXiv paper (2508.14883). I can read the abstract and a few fragments, nothing more. So my assessment is provisional.\n\nWhat seems genuinely new is the conditional-residual construction for knockoffs on mixed data without distributional assumptions, and the MALD importance measure for nonlinear outcome models. That addresses a real gap: most knockoff methods assume homogeneous data or known feature distributions. The DNA methylation aging application and shipping code (rangerKnockoff, MALDimportance) are good practice.\n\nThe main concern you'd need to resolve is the FDR guarantee. The abstract says \"principled FDR control,\" but if the residual model is estimated from the same sample used for selection, exact exchangeability fails in finite samples. That's a standard critique of estimated-model knockoffs, and the paper needs to state the regularity conditions and whether the guarantee is exact, asymptotic, or approximate. Without that, the headline claim is oversold. This isn't fatal by itself—many knockoff methods live in this asymptotic regime—but it should be stated plainly.\n\nI can't speak to the simulations or baselines because I can't read them. The abstract claims better FDR control and power, but with no error bars or replicate details visible. That's a matter of the corrupted file more than the science, but it means I'm not willing to vouch for the numbers.\n\nBottom line: this deserves a serious referee if a readable version is available. The idea is timely and the application is plausible. I'd want to see the exchangeability conditions addressed and the simulations fully specified before trusting the results. For me, I wouldn't cite it yet—I'd wait for a clean version. But I'd bring it to a reading group to discuss the residual construction, because the concept is interesting even if the details are currently beyond reach.\n\nSend it to peer review, but ask the authors to resubmit with a clean PDF first.","headline":"Promising knockoff extension for heterogeneous data, but the current text is unreadable and the FDR guarantee needs a stated regularity condition.","tokens_in":19172,"tokens_out":2831,"would_cite":false,"duration_ms":31726,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Residual knockoffs control false discoveries on any mixed data type.","keywords":["knockoff variables","false discovery rate","conditional residuals","variable importance","mixed data types","nonlinear models","DNA methylation","model-X knockoffs"],"falsifier":"Run a simulation with independent features and known null variables, apply the paper's residual-based knockoff construction at a nominal FDR, and repeat across many datasets; if the observed false discovery rate consistently exceeds the nominal level when the residual model is misspecified, the finite-sample exchangeability claim fails.","tokens_in":18208,"feed_emoji":"🎯","tokens_out":2883,"duration_ms":37703,"temperature":0.7,"pith_summary":"The paper proposes a knockoff generation method that works on heterogeneous, distribution-free data by building knockoffs from conditional residuals rather than known probability models. The authors claim this preserves false discovery rate control while extending knockoff selection to mixed continuous, categorical, and other feature types. They also introduce MALD, an importance measure based on the average absolute local derivative of a fitted outcome model, which ranks variables without requiring a linear outcome assumption. If correct, the framework makes principled FDR-controlled selection available for real-world biomedical datasets with nonlinear relationships, such as the DNA methylation aging study used as a demonstration.","feed_headline":"Residual knockoffs control false discoveries on any data type","feed_subtitle":"No distributional assumptions needed; importance scores work with random forests and neural nets.","key_machinery":"Conditional-residual knockoff generation: for each feature, a model of the feature given the remaining features produces residuals whose conditional structure is used to fabricate a knockoff copy, avoiding the need for a fully specified joint distribution. Alongside it sits the Mean Absolute Local Derivative (MALD), which averages the magnitude of the fitted outcome function's derivative with respect to a feature over the observed data, providing a nonlinear-compatible importance score.","core_discovery":"The central claim is that exchangeable knockoff variables can be generated without knowing the joint distribution of the features, provided one can estimate each feature's conditional distribution given the others and use the resulting residuals to form the knockoff. The counterfactual variables are constructed so that, together with the original features, they satisfy the exchangeability property required by the knockoff filter. The same paper introduces MALD, the mean absolute local derivative of the outcome with respect to each feature, as an interpretable variable importance score that works with nonlinear outcome models including random forests and neural networks. Simulations are prese","pith_inferences":["The conditional-residual principle is a general template: replacing the internal residual model with any good conditional distribution estimator could extend knockoff selection to spatial, temporal, or graph-structured features, though the paper does not claim this.","Because MALD is derivative-based, it could be used for local or subpopulation-level importance inside deep models, a natural next step left implicit by the authors.","A direct stress test would compare the method's empirical FDR against oracle knockoffs built from the true joint distribution on synthetic data; the gap quantifies how much of the finite-sample guarantee depends on residual-model accuracy."],"forward_implications":["Variable selection can be run on datasets that mix continuous, categorical, count, and genomic features without fitting one global generative model.","FDR control no longer requires a linear or known outcome model, so selected sets from random forests or neural networks can carry error-rate guarantees.","MALD offers a common importance scale across nonlinear model classes, allowing direct comparison of predictors in complex biological systems.","The mouse DNA methylation demonstration suggests the method can nominate candidate CpG sites for aging studies with a stated false discovery bound."],"supporting_citations":[],"fun_headline_variants":["Residual knockoffs control FDR without distributional assumptions","MALD importance metric works with nonlinear models for variable selection","Knockoffs via conditional residuals handle mixed data types","Residual knockoffs and MALD: flexible selection for heterogeneous data"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes that conditional-residual models fitted on the analyzed sample yield residuals whose knockoffs remain exchangeable enough with the true features to give nominal false discovery control in finite samples.","fun_headline_variants_meta":{"raw":{"variants":["Residual knockoffs control FDR without distributional assumptions","MALD importance metric works with nonlinear models for variable selection","Knockoffs via conditional residuals handle mixed data types","Residual knockoffs and MALD: flexible selection for heterogeneous data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":2779,"prompt_tokens":677,"completion_tokens":2102,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":2034}},"tokens_in":421,"tokens_out":2102,"duration_ms":20074,"temperature":1.0,"reasoning_tokens":2034,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:13:02.107463+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a simulation with independent features and known null variables, apply the paper's residual-based knockoff construction at a nominal FDR, and repeat across many datasets; if the observed false discovery rate consistently exceeds the nominal level when the residual model is misspecified, the finite-sample exchangeability claim fails.","supporting_citations":[],"review_version":1}