{"id":"b4ddf7b4-ee04-4a99-a859-3e1ea333bc45","arxiv_id":"1908.05885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using mcsimex with a misclassification matrix estimated from the fitted clustering model reduces bias in regression coefficients when outcomes are regressed on unsupervised cluster labels.","lead":"This paper applies an existing bias-correction method, mcsimex, to regression models whose predictor is a cluster label obtained from unsupervised clustering, and shows in simulations that correcting for label misclassification reduces bias. It offers biomarker researchers a practical workflow for re-analyzing studies that regress survival or treatment response on automatically discovered subgroups.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Application's bootstrap-based misclassification matrix estimates clustering instability, not the label error P(H*|H) that mcsimex requires; the reported survival correction may target the wrong quantity.","rationale":"The reader identified the consistency of the misclassification matrix as the weakest assumption, which is close to my concern. My attack sharpens it: even when the GMM is correctly specified, the bootstrap estimator used in Section 4 does not estimate the quantity in eq. (7); it measures label instability across resamples, with the full-data GMM labels serving as a proxy for truth. This is distinct from the reader's framing that the GMM is merely an approximation to hierarchical clustering. The distinction matters because it points to a concrete fix: use a misclassification matrix defined against an external or clinically meaningful reference, or provide a theoretical justification for the bootstrap estimator. I do not see a flaw in the logistic-GMM simulations; they convincingly show bias reduction under the ideal setting. The Cox extension is a useful contribution, but its simulation uses known misclassification without a clustering step, so it does not validate the full two-stage workflow for survival outcomes. The application is the only demonstration on real data, and its methodological deviation makes the reported hazard ratio uncertain. These issues are addressable through additional simulations and a re-analysis, consistent with a conditional verdict rather than rejection.","tokens_in":10512,"tokens_out":14858,"duration_ms":146584,"concrete_test":"Re-analyze the multiple myeloma data using the confusion matrix in Table 4, normalized by column to form P(GMM class | Zhan class), as the misclassification matrix for mcsimex in the Cox model, instead of the bootstrap disagreement matrix. If the resulting hazard ratio and confidence interval differ materially from Table 5, the application's correction is sensitive to the choice of misclassification estimator, confirming that the bootstrap matrix does not target the theoretical P(H*|H).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The simulations support the mcsimex correction only when the misclassification matrix is a consistent estimate of P(H*|H) as defined in eq. (3) and estimated by eq. (7). In the survival application (Section 4), the matrix is not estimated this way. The bootstrap procedure compares the predicted class from the in-bag GMM fit to the class from the full-data GMM fit for out-of-bag observations, thereby estimating P(hat_h_boot(Z)=i | hat_h_full(Z)=j). This quantifies instability of the clustering rule under parameter re-estimation, not the misclassification probability of the full-data GMM labels relative to the latent biological subtype H. The full-data GMM labels are themselves error-prone surrogates; Table 4 shows they agree with the original Zhan hierarchical classes in only 90% of cases. If the true label error is larger than the reported bootstrap rates (0.07 and 0.13), the mcsimex correction will be incomplete. Consequently, the increase in hazard ratio from 2.55 (naive) to 3.15 (mcsimex) in Table 5 cannot be clearly attributed to correction for misclassification; it may partly reflect correction for sampling variability of the clustering fit. This gap is load-bearing because the consistency argument in Section 2.2 does not cover the estimation strategy actually used in the application, and Section 3.2 provides no simulation of the full two-stage workflow for Cox models.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies regression of an outcome on class labels obtained by unsupervised clustering, treating the clusters as misclassified surrogates of latent classes. The authors specify a Gaussian mixture model for the covariates (Section 2.1), estimate the misclassification matrix by plug-in integration over the fitted mixture (Eq. 7), and propose the mcsimex algorithm to reduce bias in a subsequent GLM or Cox regression. Simulation studies for logistic regression (Section 3.1) and Cox regression (Section 3.2) are reported, together with an application to multiple myeloma survival using GMM labels approximating the hierarchical clusters of Zhan et al. (2006) (Section 4). The paper also describes an extension of the R simex package to coxph objects.","tokens_in":10791,"tokens_out":8857,"duration_ms":84586,"significance":"If the claims hold, the paper provides a practical and general two-stage correction that preserves the common workflow of first clustering and then regressing. The extension of mcsimex to Cox models and the public code are concrete contributions. The simulation evidence supports the claim that, when the misclassification matrix is a consistent estimate of the label error mechanism, mcsimex reduces bias and improves coverage for logistic regression; the Cox simulations with known misclassification rates show improvement as well. The central limitation is that the application in Section 4 uses a bootstrap-derived matrix that estimates a different quantity, so the applied correction is not covered by the consistency argument of Section 2.2. With revision to close that gap and to report the missing k-means simex results, the paper would be a useful addition.","major_comments":[{"comment":"The misclassification matrix used in the data application is not the estimator of Eq. (7) and does not estimate P(H*|H) as defined in Eq. (3). The bootstrap procedure compares the class predicted by the in-bag GMM fit with the class obtained from the full-data GMM fit for out-of-bag observations, so it measures the instability of the fitted clustering rule under parameter re-estimation rather than the probability that the full-data GMM label differs from the latent biological subtype. Since the full-data GMM labels themselves only agree with the original hierarchical classes in 90% of cases (Table 4), the reported rates 0.07 and 0.13 may understate the true label error. The increase in hazard ratio from 2.55 (naive) to 3.15 (mcsimex) can therefore not be unambiguously attributed to correction for misclassification. Section 3.2 also does not simulate this full two-stage workflow: its Cox simulations contaminate labels with a known misclassification matrix and do not fit a clustering model first.","section":"Section 4, Table 5"},{"comment":"The displayed results do not contain the k-means simex results that the text appears to reference. The Kmeans panels have columns True and Naive only, while the Simex columns appear only under GMM. The statements that using the simex approach improves bias and coverage and that for k-means the simex approach cannot fully correct for the added noise are therefore not supported by any visible simulation. If the k-means simex results were omitted from the tables, they need to be shown or the corresponding claims should be removed. The text also says results are summarized by bias, mean-square error, and coverage, but no MSE values appear in Tables 1 or 2.","section":"Section 3.1, Tables 1 and 2"},{"comment":"The definition G(λ)=β(Π^(1+λ)) with extrapolation to λ=-1 is inconsistent with the algorithm's base point λ0=1. If λ0=1 is used as the naive estimator, then G(1) would equal β(Π^2), not β(Π), under Eq. (8); alternatively, if the algorithm's λ is shifted by one relative to Eq. (8), the extrapolation target should be λ=0, not λ=-1. The manuscript should state the convention unambiguously, because the simulation levels and extrapolation point determine the estimator.","section":"Section 2.3, Eq. (8) and Algorithm"}],"minor_comments":[{"comment":"There is a typo in 'the the mcsimex method'; it should be 'the mcsimex method'.","section":"Introduction, last paragraph"},{"comment":"The word 'binominal' should be 'binomial'.","section":"Section 3.2"},{"comment":"The word 'accomodate' should be 'accommodate'.","section":"Section 6"},{"comment":"The spelling 'Kmeans' is used inconsistently; the text should consistently use 'k-means'.","section":"Throughout"},{"comment":"The description of the 'average misclassification matrix' from 1000 bootstraps is brief; it should specify whether the average is taken entrywise over the bootstrap matrices or is obtained from a single averaged model, and how the reported 0.07 and 0.13 rates are derived from the 7-class confusion structure after collapsing to two risk groups.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The omission of the k-means simex columns appears to be an oversight, but the application issue is more fundamental. The authors should either simulate the actual workflow used in the multiple myeloma application (GMM fit, bootstrap-based misclassification matrix, mcsimex) or use a matrix estimator that matches Eq. (7), and they should reconcile the λ convention in Section 2.3 before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuine contribution is narrow and real: this paper takes the existing mcsimex estimator and points it at the common workflow where a biologist clusters first and then regresses an outcome on the cluster labels, with the misclassification matrix estimated from the fitted Gaussian mixture via eq. (7). The logistic simulations support the claim that this redresses bias and recovers coverage when the clustering model is a GMM. Extending the simex R package to coxph and shipping the code and repository is also concrete good. This is the first mcsimex application to unsupervised labels I know of, and it fills a software gap. The citation pattern is appropriate; no sign of self-citation inflation.\n\nSoft spots, in proportion:\n\n1. Tables 1 and 2 promise k-means with simex, but no k-means simex column appears. The text says k-means simex is poorer, but the numbers are missing. That needs fixing.\n\n2. Section 3.1 says results are summarized by bias, MSE, and coverage, but the tables show only bias and coverage. Either delete the MSE promise or report it.\n\n3. The nondifferential misclassification assumption — label errors independent of outcome — is never stated. For mcsimex this is load-bearing, and a reader needs it explicit.\n\n4. The Cox simulations in Table 3 use true binary labels corrupted by known noise, not labels produced by unsupervised clustering. They validate mcsimex in Cox models, but not the full two-stage workflow.\n\n5. The application's bootstrap matrix is the real weak point. It compares bootstrap GMM labels to full-data GMM labels, which estimates clustering instability under parameter re-estimation, not P(H*|H) relative to the latent biological subtype. The full-data GMM labels themselves only agree with Zhan's classes in 90% of cases; if the true label error is larger than the reported 0.07 and 0.13, the correction is incomplete. The hazard ratio shift from 2.55 to 3.15 may partly reflect correcting for sampling variability of the clustering fit rather than misclassification. The consistency argument in Section 2.2 does not cover this estimation strategy.\n\nThat said, the central argument for the GMM-label logistic setting holds up. These issues are addressable, not fatal. The paper is honestly written and reproducible.\n\nWho is it for: biomarker methodologists doing clustering-then-regression, and measurement-error readers who want a competent case study. It deserves a serious referee, but the referee should send it back for the missing k-means simex results and MSE table, and should ask the authors to either fix the application's misclassification-matrix estimator or reframe the survival result as sensitivity analysis rather than correction.","headline":"Useful, honest application of mcsimex to regression on unsupervised cluster labels, with a real gap between the consistency argument and the bootstrap matrix used in the survival application.","tokens_in":799,"tokens_out":939,"would_cite":true,"duration_ms":32955,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62J12","62N02"],"pacs":[],"model":"deepseek-v4-flash","headline":"Regressing on cluster labels quietly biases effect estimates; a simulation-extrapolation step corrects the bias.","keywords":["misclassification","unsupervised clustering","simulation-extrapolation","mcsimex","measurement error","logistic regression","Cox proportional hazards","Gaussian mixture model"],"falsifier":"Simulate data where the true class labels are generated by a mechanism different from the fitted Gaussian mixture model, such as hierarchical clustering on non-Gaussian or heavy-tailed distributions, and check whether the mcsimex-corrected estimates still achieve nominal coverage; the paper's k-means imbalance results suggest coverage will degrade, providing a concrete test of the assumption.","tokens_in":10311,"feed_emoji":"🎯","tokens_out":2518,"duration_ms":25208,"temperature":0.7,"pith_summary":"This paper argues that the common practice of regressing an outcome on class labels obtained from unsupervised clustering is statistically flawed because the labels carry misclassification error, and the error is usually ignored. The authors propose to treat the clusters as misclassified categorical covariates and to correct the regression using the misclassification simulation and extrapolation method (mcsimex). They show, via simulation, that the mcsimex correction with a misclassification matrix estimated from the fitted Gaussian mixture model reduces bias and improves confidence-interval coverage for logistic regression and Cox proportional hazards models. The method is then applied to a multiple myeloma survival study, where the corrected hazard ratio is larger than the naive estimate, although confidence intervals are wide and overlapping.","feed_headline":"Regressing on cluster labels is biased; here is the fix","feed_subtitle":"A simulation-extrapolation method restores unbiased effect estimates and coverage when outcome is regressed on labels from unsupervised…","key_machinery":"The key machinery is the misclassification matrix $\\Pi$ and its estimator $\\hat{\\Pi}$ given in equation (9), where each entry is the probability that a covariate drawn from cluster $j$ is assigned by the hard clustering rule $\\hat{h}(z)$ to cluster $i$. This matrix, together with the mcsimex extrapolation algorithm developed by Kuchenhoff and colleagues, carries the argument: it quantifies the label noise and simulates increasingly misclassified labels, then extrapolates back to the zero-noise setting. The validity of the whole workflow rests on $\\hat{\\Pi}$ being a consistent estimate of the true misclassification mechanism.","core_discovery":"The central claim is that when an outcome is regressed on class labels inferred by unsupervised clustering, the naive estimator of the effect parameter is biased, and the bias does not vanish with increasing sample size. The paper establishes that the bias can be redressed by applying mcsimex, with the misclassification matrix consistently estimated from the fitted clustering model via equation (7). In simulations with Gaussian mixture models, the mcsimex correction reduces bias from roughly 0.7 to about 0.15 for the class effect and restores coverage from near zero to around 90 percent in both logistic and Cox regression settings. The paper also documents that the correction is less effective when the clustering model is misspecified, as with k-means on imbalanced clusters, and that the jackknife variance estimator in the simex package underestimates uncertainty, prompting a bootstrap alternative.","pith_inferences":["A testable extension would be to apply the same mcsimex correction to other clustering algorithms, such as hierarchical clustering with a data-driven dissimilarity, by estimating the misclassification matrix from the empirical confusion between the fitted clusters and a validation set with known labels.","The bias documented here likely generalizes to any two-stage pipeline where an unsupervised model is fitted first and its discrete outputs are used as covariates, including cluster-then-predict approaches in personalized medicine, so the correction could have broad applicability beyond the examples shown.","The paper's reliance on a correctly specified Gaussian mixture model suggests that the method's performance degrades gracefully but not uniformly; the k-means imbalance result hints that misspecification of the clustering model is the dominant source of residual bias after correction.","If the misclassification matrix is estimated from the same data that produced the clusters, the bootstrap procedure used in the application properly accounts for the extra uncertainty, but it doubles the computational cost; an asymptotic variance formula for Cox models would be a valuable future contribution."],"forward_implications":["If the central claim is correct, then prior biomarker studies that regressed survival or response on unsupervised cluster labels may have reported biased effect estimates and overly narrow confidence intervals, and those results should be revisited.","The mcsimex workflow offers a general correction that applies to any regression model supported by the simex framework, now extended to Cox proportional hazards models, making it usable across a wide range of clinical and genomic applications.","The corrected hazard ratio in the multiple myeloma application (roughly 3.1 vs 2.5) suggests that the true association between the high-risk cluster and survival may be stronger than the naive analysis indicated.","For clustering algorithms other than Gaussian mixture models, the user must ensure that a reliable misclassification matrix can be estimated; the paper shows k-means can fail when clusters are imbalanced.","Researchers should prefer a bootstrap or asymptotic variance estimate over the jackknife, since the jackknife underestimates uncertainty and leads to coverage below the nominal level."],"supporting_citations":[{"why":"Supplies the mcsimex method, the algorithm that the paper adapts for correcting misclassified categorical covariates.","marker":"Küchenhoff et al. (2006)"},{"why":"Provides the general framework of measurement error in covariates and the catalogue of correction methods that motivates treating cluster labels as misclassified.","marker":"Carroll et al. (2006)"},{"why":"Establishes that maximum likelihood estimates under the naive misspecified model converge, giving the limit $\\beta(\\Pi)$ that mcsimex extrapolates from.","marker":"White (1982)"},{"why":"Supplies the multiple myeloma gene expression data and the original hierarchical clustering analysis whose risk groups the paper re-analyzes.","marker":"Zhan et al. (2006)"},{"why":"Provides the mclust package used to fit the Gaussian mixture models in the simulations and the application.","marker":"Scrucca et al. (2016)"},{"why":"Documents that the jackknife variance estimator underestimates uncertainty in simex, which the paper confirms and addresses with bootstrap.","marker":"Küchenhoff et al. (2007)"}],"fun_headline_variants":["Cluster labels bias regressions — fix it","Simulation-extrapolation repairs cluster-label regression bias","Unsupervised cluster labels cause bias; mcsimex fixes it","Bias from cluster labels: a simulation-extrapolation cure","Regressing on cluster labels? Expect bias and fix it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The correction works only if the misclassification matrix estimated from the fitted clustering model matches how the labels were actually misclassified.","fun_headline_variants_meta":{"raw":{"variants":["Cluster labels bias regressions — fix it","Simulation-extrapolation repairs cluster-label regression bias","Unsupervised cluster labels cause bias; mcsimex fixes it","Bias from cluster labels: a simulation-extrapolation cure","Regressing on cluster labels? Expect bias and fix it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1527,"prompt_tokens":798,"completion_tokens":729,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":648}},"tokens_in":414,"tokens_out":729,"duration_ms":6702,"temperature":1.0,"reasoning_tokens":648,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:01:46.597774+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data where the true class labels are generated by a mechanism different from the fitted Gaussian mixture model, such as hierarchical clustering on non-Gaussian or heavy-tailed distributions, and check whether the mcsimex-corrected estimates still achieve nominal coverage; the paper's k-means imbalance results suggest coverage will degrade, providing a concrete test of the assumption.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the general framework of measurement error in covariates and the catalogue of correction methods that motivates treating cluster labels as misclassified."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that maximum likelihood estimates under the naive misspecified model converge, giving the limit $\\beta(\\Pi)$ that mcsimex extrapolates from."},{"cited_title":"B., and Raftery, A","cited_arxiv_id":null,"evidence_quote":"Provides the mclust package used to fit the Gaussian mixture models in the simulations and the application."}],"review_version":1}