{"id":"29d705f7-8751-44f9-a45e-fc5ab7490106","arxiv_id":"2501.12787","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fixed-effects regression model with a tree for country clusters and a tree for nonlinear covariate effects is proposed and tested on European quality-of-life data.","lead":"This paper introduces a regression method with two decision trees, one that groups similar countries and one that finds subgroups of people based on how their health and income relate to quality of life. It then applies the method to survey data from 45,000 older Europeans and runs simulations to compare it with standard models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unit tree is built once on raw outcome means (Section 3, Step 1), so tr0 clusters can be artifacts of between-unit covariate composition in the SHARE application; the simulations never test this because covariates are iid across units.","rationale":"I read the paper in good faith. Its central claim is that Eq. (7) with two trees can be fitted validly and yields country clusters that are interpretable as shared intercepts. For the algorithm to deliver that, the Section 3 Step 1 ordering by outcome means must be a valid proxy for intercept ordering. The reader's weakest_assumption is the same concern, and I agree with it. The perturbation that would break it is not exotic: countries in SHARE differ by construction in age, chronic disease, income, and employment (Table 1), all of which enter tr(x), and in simulation scenarios 3 and 4 the covariates are generated with the same distribution in every unit, so the problematic regime is never tried. The failure mode is not internal inconsistency; the algorithm does what it says, but the result can be an artifact of the ordering. I do not see a stronger objection: no formal verification is claimed, the comparison to Berger and Tutz (2018) and RE-EM trees is reasonably executed, and the novelty claim is credible apart from the Gottard et al. (2023) discussion, which uses random effects rather than clustered fixed effects. A null simulation test is cheap and decisive. Since the reader's CONDITIONAL verdict already reflects this uncertainty, my read leaves the verdict unchanged.","tokens_in":23657,"tokens_out":6422,"duration_ms":67729,"concrete_test":"Run the null simulation from Scenario 4 with tr0(i) constant across all units and one informative covariate having unit-specific means, e.g., X1ij ~ N(\\mu_i, 1) with \\mu_i increasing with i, \\beta_1 = 0.8, and all other DGP details as in Section 5.6. Apply TTSC with the same settings (n=20, ni=50, nmb=100, 10-fold CV with 1SE rule) over 100 replications and record the number of identified tr0 clusters and RMSEI. If TTSC systematically identifies more than one intercept cluster when the true number is one, the outcome-ordering step manufactures intercept clusters from covariate composition, invalidating the SHARE cluster interpretation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3, Step 1 orders units by their outcome means \\bar{y}_i before any covariate adjustment, and the subsequent unit tree tr0 is restricted to splits into intervals of that fixed ordering. For tr0 to recover shared intercept clusters, \\bar{y}_i must preserve the ordering of the true intercepts. This fails when units differ in covariate composition or covariate effects: E[y_ij | i] = tr0(i) + E[tr(x_ij) | i], so the ordering can be dominated by the covariate term. The SHARE data are exactly such a case: countries differ markedly in age, chronic disease count, income distribution, and employment (Table 1). In simulation scenarios 3 and 4 (Sections 5.5 and 5.6), all covariates are drawn iid within and across units, so the intercept ordering and the outcome-mean ordering coincide and the distortion is never exercised. If the ordering is wrong, the algorithm cannot recover true clusters because unit splits are intervals of a mistaken order; the Figure 2 country clusters and the cross-national interpretation then partly reflect covariate composition rather than country-specific intercepts. This is the main risk to the paper's central claim that the method provides sparse clustering of unit-specific intercepts together with tree-structured covariate effects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tree-structured fixed-effects regression model for clustered data in which the predictor η(x_ij) = tr0(i) + tr(x_ij) combines a tree-structured unit-specific intercept term with a tree-structured covariate effect. The fitting procedure orders units by their outcome means, treats the unit index as an ordinal splitting variable, greedily selects splits by deviance, and prunes by cross-validated predictive log-likelihood with a one-standard-error rule. The method is applied to SHARE data on quality of life in older adults, yielding five country clusters and covariate subgroups defined by chronic disease count, income, and age. A simulation study compares the proposed TTSC model to LMM, LMMP, RE-EM, LMMT, LTSC, LTSCB, Null, and Perfect models across four DGPs and six settings.","tokens_in":23904,"tokens_out":5059,"duration_ms":51557,"significance":"If the claims hold, the paper fills a genuine gap by combining sparse fixed-effects clustering of units with tree-structured covariate effects, and it does so with an algorithm that is implemented in the existing TSVC R package. The simulation evidence supports competitiveness for nonlinear covariate effects and for settings with few unit clusters, and the SHARE application demonstrates interpretability and reproduces plausible findings on quality of life. However, the central claim of recovering shared intercept clusters rests on the outcome-mean ordering of units, which has not been validated under between-unit covariate heterogeneity, and the application lacks any uncertainty assessment. The contribution is therefore of moderate significance and would be strengthened considerably by additional validation.","major_comments":[{"comment":"The unit tree tr0(i) is restricted to splits into intervals of the fixed ordering of units by outcome means ȳ_i. For the true cluster structure to be recoverable, this ordering must preserve the ordering of the true intercepts tr0(i). In general E[y_ij | i] = tr0(i) + E[tr(x_ij) | i], so between-unit differences in covariate distributions or covariate effects can dominate the outcome-mean ordering. In the SHARE application (Table 1; Figures 2 and 3), countries differ markedly in age, chronic disease count, income, and employment, so the identified country clusters may partly reflect covariate composition rather than shared country-specific intercepts. In simulation scenarios 3 and 4 (Sections 5.5 and 5.6), covariates are iid within and across units, so the outcome-mean ordering and the true intercept ordering always coincide and the distortion is never exercised. Please add simulation scenarios with between-unit covariate mean shifts or unit-varying covariate effects, and in the application assess the sensitivity of the country clusters to the ordering, for example by ordering units by covariate-adjusted residuals or by an alternative stratification.","section":"Section 3, Step 1 and Eq. (7)"},{"comment":"The substantive interpretation of the five country clusters and the reported subgroup effects (e.g., the -5.10 and +3.40 CASP score differences) is presented without any measure of uncertainty or stability. The greedy stepwise algorithm, the fixed ordering of units, and the 1SE cross-validation rule can produce unstable partitions, especially with only 28 units. A bootstrap or resampling-based assessment of cluster membership and of the adjusted coefficients, or at least a sensitivity analysis with respect to Smax, nmb, and dmax, is needed before drawing cross-national conclusions from Figures 2 and 3.","section":"Section 4, Figures 2 and 3"},{"comment":"The simulation study evaluates performance mainly through RMSEX, RMSEI, TPR, and FPR, but it does not report a direct measure of cluster recovery accuracy (e.g., adjusted Rand index or proportion of units assigned to the true cluster) for scenarios 3 and 4, where the true unit clusters are known. Since the paper's central claim concerns sparse clustering of unit-specific intercepts, RMSEI alone does not distinguish correct cluster recovery from a model that simply fits unit means. A direct cluster-recovery metric would provide the missing evidence that the outcome-mean ordering identifies the intended cluster structure.","section":"Section 5"}],"minor_comments":[{"comment":"The abbreviation \"DPG\" appears in the description of the simulation design and elsewhere (e.g., \"A DPG with predictor (2)\"); it should be \"DGP\".","section":"Section 5.1"},{"comment":"The final summary bullet states \"MRSE on the covariate- as well as the unit-level\"; the intended term is \"RMSE\".","section":"Section 5.6, summary bullet"},{"comment":"The adjusted cluster intercept β̃_0c = β_0c + γ̄ uses a global average of the covariate-tree coefficients over all individuals; if covariate distributions differ across clusters, this quantity is not the expected value for that cluster. Please clarify the interpretation or define a cluster-specific adjustment.","section":"Section 2.2, Eq. (11)"},{"comment":"The sentence \"the coefficients of the tree-structured model in (7) can only interpreted with regard to a reference node\" is missing the word \"be\" before \"interpreted\".","section":"Section 2.2"},{"comment":"The figure legend does not state whether the displayed node values are raw coefficients from the model in Eq. (7) or the adjusted coefficients from Eq. (11); please specify this in the legend or caption.","section":"Figure 3"},{"comment":"The claim that a method combining tree-structured covariate effects with clustered fixed effects \"has not been available so far\" is too strong given the discussion of Gottard et al. (2023) later in the same section; consider softening the statement, for example by saying that this combination has not been available within the fixed-effects framework used here.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is closely related to Berger and Tutz (2018) and relies heavily on the TSVC package, but the extension to a second tree for covariate effects is a clear methodological increment. The main risk is that the outcome-mean ordering of units may not identify latent intercept clusters under between-unit covariate heterogeneity, a scenario that is realistic for the SHARE data but absent from the simulations. If the authors can add a targeted simulation and a sensitivity analysis for the application, the paper would be suitable for publication in an applied statistics journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper does something genuinely new: it puts the unit tree from Berger and Tutz (2018) and the covariate tree from RE-EM-style methods into one fixed-effects model, with an algorithm that refits all parameters after each split. That combination is not in the cited literature, and the simulation study is broad enough to show the method is competitive, especially when covariate effects are nonlinear and the number of clusters is small. The SHARE application is interpretable and the results align with prior findings.\n\nThe main soft spot is the unit-tree construction. Section 3 orders units by outcome means before any covariate adjustment, and all unit splits are intervals of that fixed ordering. For the clusters to be meaningful, the ordering must reflect the true intercepts. But E[y_ij | i] = tr0(i) + E[tr(x_ij) | i], so if units differ in covariate composition, the ordering can be dominated by the covariate term. The simulations never exercise this: in scenarios 3 and 4, covariates are iid across units, so the ordering coincides with the intercept ordering. In the SHARE application, countries differ systematically in age, chronic disease, income, and employment, so the identified country clusters in Figure 2 may partly reflect covariate composition rather than shared intercepts. That is a real threat to the paper's central claim, and it is fixable: add a DGP with unit-specific covariate means, or try iterative re-ordering within the fitting loop.\n\nThe other issues are minor. There is no uncertainty quantification for the tree structure or estimates—typical for tree methods, but worth stating. No simulation or analysis code is shipped. No theoretical properties, which is not surprising for a CART-style heuristic but should be acknowledged. Self-citation is not a problem here; the Berger-Tutz result is directly relevant and properly credited.\n\nWho is this for: practitioners and methodologists working on clustered or tree-based regression, especially with few clusters and many observations per cluster. It deserves a serious referee—the idea is sound, the gap is real, and the ordering concern is addressable in revision. I would accept it for review, with a request for the extra simulation and a more cautious interpretation of the clusters.","headline":"A genuinely useful combination of unit-tree and covariate-tree fixed effects, but the unit clustering rests on an outcome-mean ordering that the simulations never stress-test.","tokens_in":24432,"tokens_out":2994,"would_cite":true,"duration_ms":30753,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J02","62P25"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a regression model whose predictor is the sum of two trees—one that clusters units into shared intercepts and one that captures nonlinear covariate effects—and shows it on quality-of-life data from the SHARE survey.","keywords":["tree-structured regression","clustered data","fixed effects models","unit-specific intercepts","recursive partitioning","quality of life","CASP","SHARE"],"falsifier":"Simulate data with true clustered intercepts while letting covariate distributions differ strongly across units, and check whether recovery of the true clusters by the intercept tree degrades relative to a balanced setting; if the outcome-mean ordering is the weak point, cluster recovery should drop exactly in that scenario.","tokens_in":23425,"feed_emoji":"🌳","tokens_out":5991,"duration_ms":56936,"temperature":0.7,"pith_summary":"This paper develops a regression model for clustered data whose predictor is the sum of two trees: one tree gives each cluster of units (here countries) its own intercept, and the other tree captures how covariates affect the outcome, including nonlinearities and interactions. The authors aim to close a gap: earlier methods had either clustered fixed country effects with linear covariate terms, or tree-structured covariate effects with random intercepts, but not both. If the method works, analysts gain an interpretable tool for cross-national surveys that both identifies groups of similar countries and subgroups of individuals at risk, without imposing linearity or a normal distribution on the unit effects. The paper supports the claim with a simulation study and an application to quality of life in older adults from the SHARE survey.","feed_headline":"A two-tree model explains quality of life across Europe's countries","feed_subtitle":"It splits 28 countries into five clusters and links chronic disease, income, and age to wellbeing.","key_machinery":"The load-bearing object is the two-tree predictor $\\eta(x_{ij}) = tr_0(i) + tr(x_{ij})$ built inside a fixed-effects model. Units are first ordered by their outcome means and treated as an ordinal variable, so clustering becomes a sequence of binary splits on the unit index, keeping the search linear rather than exponential. At each step all candidate splits in either tree are evaluated, and the one with smallest deviance is chosen; then all parameters are refitted, so the two tree components mutually adjust for each other's splits. The resulting split sequence is pruned by ten-fold cross-validated predictive log-likelihood with the one-standard-error rule, and a coefficient adjustment $\\tilde{\\beta}_{0c} = \\beta_{0c} + \\bar{\\gamma}$, $\\tilde{\\gamma}_m = \\gamma_m - \\bar{\\gamma}$ converts reference-node coding into average cluster intercepts and subgroup deviations, which is what makes the output interpretable.","core_discovery":"The central claim is that the model $\\eta(x_{ij}) = tr_0(i) + tr(x_{ij})$—a fixed-effects predictor with a tree-structured unit intercept $tr_0(i)$ and a second tree-structured covariate effect $tr(x_{ij})$—can be fitted by a stepwise deviance-based algorithm and delivers sparse, interpretable estimates of both cross-unit differences and covariate subgroups. On the SHARE data the fitted model separates the 28 countries into five clusters with expected CASP scores from 32.26 to 40.41, and identifies eleven subgroups of individuals, with number of chronic diseases, income decile, and age as the splitting variables. The lowest-quality-of-life subgroup (more than four chronic diseases and income in the poorer half of one's country) is expected to score 5.10 points below the country expectation, while the highest subgroup (no chronic diseases, income in the wealthiest 40 percent) is expected to score 3.40 points above it.","pith_inferences":["Editorial: a covariate-adjusted version of the unit ordering could be tested; if implemented, it should improve cluster recovery when covariate distributions differ strongly across units.","Editorial: the same two-tree construction extends naturally to unit-specific slopes, with one unit tree per covariate effect, though the paper only sketches that extension.","Editorial: bagging the split sequence could yield prediction intervals and stabilize variable selection, which the paper does not pursue.","Editorial: the country clusters should be read descriptively; the observational SHARE design does not support a causal claim that wealth causes higher quality of life."],"forward_implications":["Clustered fixed-effects modeling becomes available for nonlinear and interactive covariate effects, not only linear ones.","The country clusters found in SHARE line up with welfare-regime and wealth patterns, giving a sparse description of cross-national quality-of-life differences.","The model selects variables during fitting, so subgroup discovery and variable selection happen in one pass.","Simulation results indicate the approach is most competitive when the number of units is small, observations per unit are many, and the true number of clusters is small.","Because fitting is likelihood-based, the same algorithm extends in principle to binary and discrete outcomes."],"supporting_citations":[{"why":"The tree-structured clustering fixed-effects model that this paper extends by adding a second tree for covariate effects.","marker":"Berger and Tutz (2018)"},{"why":"The RE-EM tree approach whose alternating estimation and cross-validated one-standard-error pruning are adapted here and used as a simulation competitor.","marker":"Sela and Simonoff (2012)"},{"why":"One of the two simultaneous proposals of mixed-effects regression trees, cited as the origin of combining tree-structured covariate effects with unit-specific intercepts.","marker":"Hajjem et al. (2011)"},{"why":"Supplies the recursive-partitioning software framework in which the proposed algorithm is implemented.","marker":"Berger et al. (2019)"},{"why":"Methodological reference for the SHARE survey data used in the application.","marker":"Börsch-Supan et al. (2013)"},{"why":"Documents the CASP-12 quality-of-life scale used as the outcome in the SHARE application.","marker":"Borrat-Besson et al. (2015)"},{"why":"The conditional-inference-tree-based mixed model used as a comparison method in the simulation study.","marker":"Fokkema et al. (2018)"}],"fun_headline_variants":["Tree model splits 28 countries into five quality-of-life clusters","Two trees reveal quality-of-life subgroups in ageing Europe","How a tree model maps wellbeing across Europe's nations","Chronic disease and income power a new tree model","Tree-based regression untangles country and individual effects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that ordering units by their average outcome recovers the true grouping structure for shared intercepts, so a noisy or confounded ordering could make the country clusters track the outcome ranking instead of genuine intercepts.","fun_headline_variants_meta":{"raw":{"variants":["Tree model splits 28 countries into five quality-of-life clusters","Two trees reveal quality-of-life subgroups in ageing Europe","How a tree model maps wellbeing across Europe's nations","Chronic disease and income power a new tree model","Tree-based regression untangles country and individual effects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2450,"prompt_tokens":905,"completion_tokens":1545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1479}},"tokens_in":521,"tokens_out":1545,"duration_ms":11293,"temperature":1.0,"reasoning_tokens":1479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:47:40.162769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data with true clustered intercepts while letting covariate distributions differ strongly across units, and check whether recovery of the true clusters by the intercept tree degrades relative to a balanced setting; if the outcome-mean ordering is the weak point, cluster recovery should drop exactly in that scenario.","supporting_citations":[],"review_version":1}