{"id":"21c061e6-d294-4607-b93f-3812d284f2f1","arxiv_id":"1908.07908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Mixed-SCGLR extends supervised component-based generalised linear regression to generalised linear mixed models, estimating interpretable orthogonal components that jointly explain multiple responses with grouped data.","lead":"The paper presents a method for fitting complex ecological models where many overlapping measurements predict multiple species counts, while accounting for shared habitat structure. It combines component-based dimension reduction with random effects for grouped data, and tests the method on simulated and tree-genus data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 6 over-claims for non-Gaussian GLMMs: Algorithm 1 depends on Eq. (7)'s Gaussian working-variable assumption, but Section 4 tests only the Gaussian case.","rationale":"The reader's weakest assumption correctly identifies Eq. (7)'s Gaussian working-variable approximation as load-bearing. I agree: the methodological derivation is internally coherent, but the central performance claim is broader than the evidence. The simulation section validates only the LMM case, so the Poisson real-data example is the only test of the generalized setting, and it lacks any comparison to a benchmark or a gold-standard fit. This is not a reason to reject the algorithm, but it is a reason to require a revision that either tests the non-Gaussian regime directly or qualifies the conclusion. I therefore do not change the reader's CONDITIONAL verdict. The concrete test above would settle whether the concern lands by directly assessing the adequacy of Eq. (7) in the Poisson setting where the paper's real-data application lives.","tokens_in":8549,"tokens_out":5176,"duration_ms":60925,"concrete_test":"Run the Section 4 simulation design with Poisson responses (e.g., y_k ~ Poisson(exp(eta_k)), 10 groups of 10 units, tau = 0.5, correlated X bundles) over 500 replicates. Fit Mixed-SCGLR and a gold-standard GLMM fit by adaptive Gauss-Hermite quadrature or MCMC. Compare fixed-effect estimates, random-effect variance estimates, and out-of-sample predictive RMSE. If the relative difference in estimated variance components or predictive RMSE exceeds 10%, or if coverage of the fixed effects falls below the nominal 90%, Eq. (7)'s Gaussian approximation is inadequate and the Section 6 claim should be restricted to linear mixed models.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The algorithm's core is Schall's pseudo-likelihood. Equation (7) asserts that the working variables z_xi^k, conditional on gamma, delta, xi, are Gaussian with mean (Xu)gamma_k + T delta_k + U xi_k and covariance W^{-1}_{xi,k}. This is exact only for linear mixed models with normal errors and identity link; for Poisson or binomial responses it is an approximation that can be poor with small groups and small counts. Algorithm 1 uses Eq. (7) in every step: Step 1's psi_T is a cos^2 fit under the W_xi metric, Step 2's Henderson system is the normal posterior from Lemma 1, and Step 3's variance update is Schall's REML formula derived from that normal hierarchy. Yet Section 4 simulates only LMMs (normal errors, identity link), where Eq. (7) holds exactly, so Table 2 cannot validate the 'Generalized Linear' part of the claim. The real-data Section 5 is Poisson, but it reports no comparison with Ridge/Lasso or with a gold-standard GLMM fit, and tuning choices (l = 4, s = 0.5) are selected for interpretability before K is cross-validated. Thus the strongest claim—'performs well on grouped data, and provides robust predictive models'—has no evidence in the regime where the approximation is non-trivial.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Mixed-SCGLR, an extension of Supervised Component-based Generalised Linear Regression to multivariate GLMMs. The method constructs orthogonal components f = Xu of the redundant X block by maximizing [φ(u)]^s [ψ_T(u)]^{1-s}, where φ measures structural relevance and ψ_T is a goodness-of-fit term built from Schall's working variables and their conditional covariance; parameters are then updated by solving Henderson systems, with variance components updated by Schall's REML-type formula. A simulation study in the Gaussian LMM setting compares Mixed-SCGLR with Ridge and Lasso regularizations, and a Poisson real-data application is presented.","tokens_in":8799,"tokens_out":4581,"duration_ms":46249,"significance":"If the empirical claims held, Mixed-SCGLR would offer a useful alternative to penalized methods for multivariate GLMMs with many redundant regressors, with interpretable components and few tuning parameters. The paper's formal contribution is clear: Lemma 1 gives a compact Bayesian justification of the Henderson systems, the PING algorithm in the appendix provides a concrete constrained-optimization procedure, and the extension to higher-rank components is carefully specified. However, the current evidence is weakened by the simulation metric and by the absence of a validated non-Gaussian case, so the contribution is not yet established.","major_comments":[{"comment":"The MLRE criterion takes, for each simulation, the minimum relative error over the two responses for each method. A method that fits only one of the two responses well and the other poorly will obtain the same low score as a method that fits both moderately, so Table 2 does not support the claim that Mixed-SCGLR gives more precise fixed-effect estimates than Ridge or Lasso. Please report per-response relative errors, with standard errors or quantiles over the 500 simulations, and use a joint or average criterion that rewards balanced multivariate performance.","section":"Section 4, MLRE definition and Table 2"},{"comment":"Equation (7) states that Schall's working variables z_ξ^k are exactly Gaussian with mean (Xu)γ_k + Tδ_k + Uξ_k and covariance W^{-1}_{ξ,k}. For non-Gaussian responses this is an approximation; nevertheless every step of Algorithm 1 (Step 1's ψ_T, Step 2's Henderson system, Step 3's variance update) is derived from this Gaussian hierarchy. Section 4 simulates only linear mixed models with normal errors and identity link, where Eq. (7) holds exactly, and the Poisson analysis in Section 5 is not benchmarked against Ridge, Lasso, or a standard GLMM fit. The paper therefore provides no empirical support for the claimed performance in the generalized-linear regime.","section":"Sections 3.1 and 5, Eq. (7)"},{"comment":"The tuning parameters l=4 and s=0.5 are selected 'among a series of parameter choices' for interpretability before K is cross-validated. This order of selection can bias the reported AveNRMSE, because the same data are used to choose l and s and then to evaluate models. Please describe how l and s were chosen (e.g., on a separate tuning set) or provide a sensitivity analysis showing that the AveNRMSE conclusions are stable to reasonable changes in l and s.","section":"Section 5, tuning of l and s"},{"comment":"Algorithm 1 is an alternating procedure, but the stopping rule is only described as 'repeated until stability of u and parameters'; no convergence criterion or assessment is given. Since the objective changes at each iteration through the working variables, it would be useful to specify a precise stopping tolerance and to report convergence behavior or failures in the simulations.","section":"Section 3.1 and Algorithm 1"}],"minor_comments":[{"comment":"There are several typos in the introduction, for example 'developped' should be 'developed' and 'unknow' should be 'unknown'.","section":"Section 1"},{"comment":"The notation 'Mixed-SC(G)LR' is inconsistent with 'Mixed-SCGLR' used in the rest of the paper; please unify the notation.","section":"Tables 1 and 2"},{"comment":"In the definition of D_h = X' W F_h, the metric W is not specified; it should be stated whether W is the same as in Eq. (2) or a response-weighted metric.","section":"Section 3.2, Eq. (9)"},{"comment":"Table 1 reports optimal parameter values but not their variability over the 500 simulations; adding standard deviations or an interquartile range would strengthen the comparison.","section":"Section 4, Table 1"},{"comment":"The caption refers to 'X-part linear predictors' as supplementary variables, but this term is not defined before the figure; please clarify.","section":"Figure 2 caption"},{"comment":"The text contains 'developped plots' and 'parcimonious'; these should be 'developed plots' and 'parsimonious'.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and the method is well motivated. My main concern is that the current empirical evaluation is not sufficient to support the paper's central claims; the MLRE metric and the lack of a non-Gaussian validation are the key issues. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper does what it says on the tin. It extends SCGLR to grouped data by embedding the component criterion in Schall's pseudo-likelihood, and the derivation is coherent. The Henderson systems, the variance updates, and the PING algorithm in the appendix all hang together. That is real work, and it is a natural extension not in the cited literature.\n\nWhat is new: the combination of SCGLR's component search with Schall's working variables and REML-type variance updates. The criterion [φ(u)]^s [ψ_T(u)]^{1-s} is inherited from earlier SCGLR papers, but using the Schall working variables in ψ_T is the novel step. The paper also gives a clear algorithmic recipe (Algorithm 1) and a useful constrained optimization appendix. The real-data application to tree genus counts is a nice demonstration of interpretable components from a Poisson GLMM.\n\nNow the soft spots, in proportion. The stress-test is right: Section 4 is the only simulation study and it is strictly Gaussian (LMM, identity link), where Equation (7)'s normal working-variable assumption holds exactly. So Table 2 can support a claim about LMMs but not about the \"Generalized\" part of GLMMs. The conclusion's claim that Mixed-SCGLR \"performs well on grouped data\" is therefore too strong. This is a load-bearing gap, not a minor one.\n\nThe simulation metric is also weak. MLRE takes the minimum relative error over the two responses for each method. That lets a method that fits one response well and ignores the other score highly. For a multivariate method, a joint error metric (e.g., sum or average, or a multivariate loss) is needed. There are also no error bars despite 500 simulations; standard errors would be trivial to add.\n\nThe real-data section has no comparison with ridge, lasso, or a standard GLMM fit, so we learn only that the method runs and gives interpretable components. Tuning is also done with some hindsight: l and s are chosen for interpretability before K is cross-validated. That is not fatal, but it should be reported as a sensitivity analysis, not as a clean selection procedure.\n\nNo code or data are provided. For an algorithmic paper in 2019, that is a real deficiency, especially because SCGLR has an R package and this extension could be implemented on top of it.\n\nOn circularity: the reader's circularity burden score of 2 is right. The estimation is not circular; components are built from current working responses and parameters are updated from the resulting Henderson systems. Cross-validation for K and s is model selection, not circular reasoning.\n\nWho this is for: applied statisticians in ecology and biometrics who want regularized, interpretable multivariate GLMMs. They will get a useful algorithm, but they should not trust the performance claims until the non-Gaussian simulations and a fair metric show up.\n\nRecommendation: this deserves peer review. Send it to a serious referee. The method is sound enough to be worth the referee's time, but the revision must add non-Gaussian simulations, a proper multivariate error metric with uncertainty, and ideally a public implementation.\n\nMy verdict: conditional accept, in spirit.","headline":"A coherent and genuinely new extension of SCGLR to GLMMs via Schall's algorithm, but the empirical case only covers the Gaussian regime and the simulation metric flatters single-response fits.","tokens_in":9376,"tokens_out":1794,"would_cite":false,"duration_ms":22302,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J12","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper extends the SCGLR component-based regularisation criterion to generalized linear mixed models, claiming lower mean relative error than Ridge or Lasso on grouped data with redundant predictors and interpretable components on a…","keywords":["component-based regularisation","multivariate GLMM","random effects","structural relevance","SCGLR","Schall's algorithm","Henderson equations","Ridge-Lasso comparison"],"falsifier":"Replicate the Section 4 simulation with Poisson or binomial responses, small groups (e.g., N=10 groups of size 5), and known coefficient vectors; compute the mean relative error for Mixed-SCGLR, LMM-Ridge and (G)LMM-Lasso with cross-validated tuning. If for $\\tau\\ge0.3$ Mixed-SCGLR is not at least as accurate as the best penalised alternative, the claim that it performs well on grouped data would be contradicted; a second check is whether the maximised criterion in Equation (3) increases monotonically along the PING iterations, since a non-monotone path indicates possible local-maximum stops.","tokens_in":8338,"feed_emoji":"📊","tokens_out":10175,"duration_ms":98352,"temperature":0.7,"pith_summary":"Mixed-SCGLR extends the Supervised Component-based Generalised Linear Regression (SCGLR) to generalized linear mixed models, so that data with grouped or correlated units and many redundant predictor variables can be analysed without collapsing the dependence structure. The method alternates between building an orthogonal component $f=Xu$ that balances fit to the responses (through Schall's working variables) with proximity to the strongest directions in $X$, and estimating fixed and random effects from Henderson systems. The paper's simulations indicate that, when the predictors contain genuine latent bundles, this component-based regularisation has lower mean relative error than Ridge or Lasso-regularised mixed models; on a Poisson count dataset of tree genera, it selects components that separate rainfall/wind and photosynthesis signals. A sympathetic reader would take the contribution to be a practical, interpretable alternative to penalty-based regularisation for multivariate GLMMs.","feed_headline":"Component regression beats ridge and lasso on grouped data","feed_subtitle":"Mixed-SCGLR adds random effects to supervised components, reducing error and giving interpretable axes on grouped data.","key_machinery":"The optimization problem at the heart of the method is $$\\max_{u'Au=1}\\,[\\varphi(u)]^s[\\psi_T(u)]^{1-s},$$ where $f=Xu$ is a component, $\\varphi(u)$ measures how close $Xu$ is to the strongest structures in $X$ (a sum over variables of powered correlations), and $\\psi_T(u)$ measures how well $Xu$ plus the extra covariates $T$ explains Schall's working variables $z_{\\xi}^k$ in the weighted metric $W_{\\xi,k}$. Conditional on $u$, the fixed effects $\\gamma_k$, $\\delta_k$ and random effects $\\xi_k$ are obtained by solving the Henderson system (5), and the variance parameters $\\sigma_k^2$ are updated by Schall's formula. The Projected Iterated Normed Gradient (PING) algorithm maximizes the criterion on the sphere $u'Au=1$ under orthogonality constraints when building higher-rank components.","core_discovery":"The central claim is that the SCGLR trade-off criterion, $$\\max_{u'Au=1}[\\varphi(u)]^s[\\psi_T(u)]^{1-s},$$ can be embedded into Schall's algorithm for GLMMs, with $\\psi_T$ redefined using the conditional working-variable variance $W_{\\xi,k}$, and with fixed and random effects solved by Henderson systems. The paper argues that this yields a valid estimation scheme because, conditionally on the random effects, the working variables form a normal hierarchy whose posterior mode is exactly the Henderson solution, and the variance components can be updated by Schall's formula. In the Gaussian simulation, the resulting estimates are more precise than LMM-Ridge and (G)LMM-Lasso whenever the explanatory variables form strong bundles ($\\tau\\ge0.3$); in the Poisson example, the method produces 12 components that can be read as interpretable environmental bundles while maintaining predictive accuracy.","pith_inferences":["If the Gaussian working-variable approximation holds poorly for small groups, the Henderson step may yield biased random-effect and variance estimates; a residual or bootstrap diagnostic of Equation (7) would be a natural next check, but the paper does not provide one.","The same alternating scheme could accommodate spatial, nested or crossed random effects by changing $U$ and $A_k$; the component-building step would be unchanged.","For count data with overdispersion, adding an extra dispersion parameter would likely improve predictive intervals and may alter the selected number of components; this is an extension the paper leaves implicit.","For nearly unstructured predictors, component-based regularisation offers little advantage over a lasso, so a data-dependent choice of $s$ could recover accuracy in that regime."],"forward_implications":["For grouped data with many redundant predictors, users can obtain a regularised multivariate GLMM without choosing a shrinkage parameter; instead they choose the number of components $K$ and the balance $s$ by cross-validation.","Because the component is common to all responses, the method borrows information across $q$ outcomes when detecting structure, unlike separate ridge or lasso fits.","The estimated components provide a low-dimensional projection of $X$ that can be inspected through correlation scatterplots, revealing which original variables form each bundle.","On the Genus data, 12 components suffice to approximate the abundance of 94 tree genera, showing the method scales to many responses and dozens of predictors.","The inclusion of unregularised covariates $T$ means known relevant variables can be forced into the model while $X$ is still regularised by components."],"supporting_citations":[{"why":"Supplies the SCGLR criterion and the Fisher-scoring adaptation that the mixed version extends.","marker":"[2]"},{"why":"Defines the GLMM framework and exponential-family conditional distributions the model assumes.","marker":"[7]"},{"why":"Supplies Schall's algorithm, including the working variables and variance-parameter updates used in each iteration.","marker":"[9]"},{"why":"Provides Henderson's mixed-model equations used to estimate fixed and random effects jointly.","marker":"[5]"},{"why":"Motivates the Bayesian normal-hierarchy justification and the choice of noninformative priors for fixed effects.","marker":"[10]"},{"why":"Provides the LMM-Ridge baseline against which Mixed-SCGLR is compared in simulations.","marker":"[3]"},{"why":"Provides the L1-penalised GLMM baseline (G)LMM-Lasso used in the comparison.","marker":"[4]"}],"fun_headline_variants":["Mixed-SCGLR beats ridge and lasso on grouped data","Component-based GLMM beats ridge and lasso","SCGLR for mixed models outperforms ridge and lasso","Schall-powered components beat ridge and lasso","Mixed-model SCGLR edges out ridge and lasso"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Schall's working variables are exactly Gaussian, with the covariance the algorithm uses, conditional on the random effects; for Poisson or binomial responses with small groups this is an approximation, and every parameter update in Algorithm 1 depends on it.","fun_headline_variants_meta":{"raw":{"variants":["Mixed-SCGLR beats ridge and lasso on grouped data","Component-based GLMM beats ridge and lasso","SCGLR for mixed models outperforms ridge and lasso","Schall-powered components beat ridge and lasso","Mixed-model SCGLR edges out ridge and lasso"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000708,"raw_usage":{"total_tokens":3160,"prompt_tokens":887,"completion_tokens":2273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":2193}},"tokens_in":503,"tokens_out":2273,"duration_ms":17764,"temperature":1.0,"reasoning_tokens":2193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:17:05.384568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replicate the Section 4 simulation with Poisson or binomial responses, small groups (e.g., N=10 groups of size 5), and known coefficient vectors; compute the mean relative error for Mixed-SCGLR, LMM-Ridge and (G)LMM-Lasso with cross-validated tuning. If for $\\tau\\ge0.3$ Mixed-SCGLR is not at least as accurate as the best penalised alternative, the claim that it performs well on grouped data would be contradicted; a second check is whether the maximised criterion in Equation (3) increases monotonically along the PING iterations, since a non-monotone path indicates possible local-maximum stops.","supporting_citations":[{"cited_title":"and Mortier, F","cited_arxiv_id":null,"evidence_quote":"Supplies the SCGLR criterion and the Fisher-scoring adaptation that the mixed version extends."},{"cited_title":"and Searle, S.R (2001) Generalized, Linear, and Mixed Models","cited_arxiv_id":null,"evidence_quote":"Defines the GLMM framework and exponential-family conditional distributions the model assumes."},{"cited_title":"(1991) Estimation in generalized linear models with random effects","cited_arxiv_id":null,"evidence_quote":"Supplies Schall's algorithm, including the working variables and variance-parameter updates used in each iteration."},{"cited_title":"(1975) Best linear unbiaised estimators and prediction under a se- lection model","cited_arxiv_id":null,"evidence_quote":"Provides Henderson's mixed-model equations used to estimate fixed and random effects jointly."},{"cited_title":"(1984) Random-Effects Models for Serial Observations with Binary Response","cited_arxiv_id":null,"evidence_quote":"Motivates the Bayesian normal-hierarchy justification and the choice of noninformative priors for fixed effects."},{"cited_title":"and Foulkes, A.S","cited_arxiv_id":null,"evidence_quote":"Provides the LMM-Ridge baseline against which Mixed-SCGLR is compared in simulations."},{"cited_title":"and Tutz, G","cited_arxiv_id":null,"evidence_quote":"Provides the L1-penalised GLMM baseline (G)LMM-Lasso used in the comparison."}],"review_version":1}