{"id":"0de785f6-6377-4548-bfa7-ce99512ca876","arxiv_id":"2412.19242","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A likelihood-based functional SEM with Gaussian process latent variables is developed, with EM estimation, penalized smoothing, and goodness-of-fit indices for sparse longitudinal data.","lead":"This paper introduces a family of functional structural equation models in which both the latent factors and their observed indicators are random curves modeled as Gaussian processes. The method estimates time-varying factor loadings and regression coefficients from sparse or irregularly sampled longitudinal data using an EM algorithm.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's stated identification constraint (first loading fixed to 1) is not enforced: all loadings are estimated freely with no latent-scale fix, leaving the model unidentified.","rationale":"The reader's verdict is CONDITIONAL, with W_eta in Section 3.2 identified as the weakest assumption. I agree that W_eta is underspecified, but I consider the identification issue more load-bearing because it affects every reported estimate and the central claim of accurate estimation. The paper explicitly states an identification constraint (first loading fixed to 1) and then violates it in the simulation and application: all lambda_j are estimated freely, and no alternative constraint (e.g., fixing the latent variance) is described. Because the model is invariant to scaling the latent variable, the likelihood has a flat direction, so the parameters are not identifiable in the real-data analysis. This is a correctable flaw, but it must be fixed before the reported curves and significance statements can be interpreted. The W_eta issue is also real and should be clarified, but it is a matter of implementation specification rather than statistical identifiability. The conditional verdict remains appropriate because the paper can be revised to impose the identification constraint and specify W_eta fully.","tokens_in":24883,"tokens_out":10770,"duration_ms":102382,"concrete_test":"Re-fit the HRS model (16)-(17) under the stated constraint lambda_1(t) = 1 (the constraint from Section 2.2), or equivalently fix the variance of GFP_i(t) to 1, and compare the estimated loading curves and regression coefficients with Figures 4-5. If the constrained estimates are proportional to the unconstrained ones with identical relative shapes, the unconstrained model is unidentified. Additionally, compute the profile log-likelihood as a function of a common scaling factor c applied to (eta, gamma, and 1/lambda); a flat profile confirms non-identifiability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 2.2, the authors state that Model (1) is identified only if either the first factor loading is fixed to 1 or the latent variance constraint K_zeta_m(t,t)=1 is imposed, and they commit to fixing the first loading. However, Section 6.1 (Simulation 1) estimates lambda_1, lambda_2, and lambda_3 freely (Table 1), the structural model in Section 2.3 includes no fixed loading, and the HRS application (Eq. 16) estimates all five loadings lambda_1,...,lambda_5 freely (Figure 4). No alternative constraint on the latent scale is imposed. The measurement model is scale-invariant: replacing eta by c*eta and all loadings lambda_j by lambda_j/c leaves the distribution of z unchanged; in the structural model eta = Gamma_x x + zeta, the same transformation scales Gamma_x and the covariance of zeta. The likelihood is therefore flat along this scaling direction. The reported MSEs, confidence bands, and regression-coefficient curves are not for uniquely defined quantities unless the data-generating process fixes the latent scale (as in the simulations, where K_eta is known) or a constraint is imposed during estimation. The real-data analysis does neither, so the magnitudes in Figures 4-5 and the significance statements about 'sex' and 'cancer' are not well defined. This is more fundamental than the unspecified W_eta in Section 3.2, which affects implementation details; the identification gap affects every estimated parameter.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a family of Functional Structural Equation Models (FSEMs) in which latent variables are Gaussian processes and indicators are observed sparsely or irregularly over a domain. The inferential framework is an EM algorithm with penalized likelihood and cross-validated smoothing, supplemented by bootstrap-based confidence bands and functional goodness-of-fit indices. The method is evaluated in two simulation studies and applied to Health and Retirement Study data to estimate a General Factor of Personality latent curve and its association with sex and cancer status.","tokens_in":25230,"tokens_out":6322,"duration_ms":64780,"significance":"If the identification and specification gaps are resolved, this would be a useful contribution to functional latent variable modeling: the framework is ambitious, handles sparse and irregular longitudinal data, provides uncertainty quantification, and the authors supply code for reproduction. The simulation studies cover several data designs and report coverage rates. However, the current manuscript does not yet support the central claims because the identification constraint stated in Section 2.2 is not enforced in the simulations or the application, leaving the reported parameter estimates and significance statements not well-defined. The abstract also claims a restricted maximum likelihood approach that is not implemented in the body of the paper.","major_comments":[{"comment":"The identification constraint announced in Section 2.2 is not enforced anywhere in the estimation. Section 2.2 states that Model (1) is identified only if the first factor loading is fixed to 1 or the latent variance constraint is imposed, and it commits to fixing the first loading. Yet Section 6.1 reports MSEs for lambda_1, lambda_2, and lambda_3 (Table 1), Section 6.2 reports MSEs for lambda_1, lambda_2, and lambda_3 (Table 3), and Section 7 fits all five loadings lambda_1,...,lambda_5 freely (Eq. (16), Figure 4). No alternative constraint on the latent scale is described. The measurement and structural models are scale-invariant: replacing eta by c*eta, lambda_j by lambda_j/c, and scaling Gamma_x and the covariance of zeta accordingly leaves the distribution of the observed data unchanged, so the likelihood is flat along this direction. In the simulations the data-generating process fixes the latent scale through the known covariance of eta, but the estimation method does not appear to use that knowledge, so the reported MSEs are not meaningful unless the estimated scale is aligned with the true scale. In the real-data analysis, no such alignment exists, so the loadings in Figure 4 and the significance statements about sex and cancer in Figure 5 are not for uniquely defined quantities. The authors must either impose the fixed-loading constraint (or the variance constraint) in the estimation, or explicitly show how the EM algorithm selects a unique point on the equivalence class.","section":"§2.2, Eq. (1); §6.1, Table 1; §7, Eq. (16)"},{"comment":"The abstract and introduction state that the inferential framework is 'based on a restricted maximum likelihood approach', but Section 3.1 presents an ordinary maximum-likelihood EM algorithm maximizing a complete-data log-likelihood, and Section 8 repeats that the paper proposes a maximum likelihood framework. There is no restricted/residual likelihood, no adjustment for estimation of fixed effects, and no discussion of REML in the estimation equations. Either the manuscript should implement and describe an actual REML procedure, or the claims in the abstract and introduction should be revised to say maximum likelihood. This is a load-bearing mismatch because the abstract explicitly names REML as the basis of the approach.","section":"Abstract; §3.1; §8"},{"comment":"The weight matrix W_eta is left unspecified. Section 3.2 defines the structural model as eta = Gamma_eta W_eta eta + Gamma_x x + zeta and derives the joint covariance Sigma_eta = (I - Gamma_eta W_eta)^{-1} diag{...} (I - Gamma_eta W_eta)^{-T}, then states that the selection of W_eta is 'quite crucial' and 'depends on the specific model, with the flexibility to vary on a case-by-case basis'. No rule or example is given for choosing W_eta in a recursive model, even though different choices change the implied joint distribution of the latent factors and can break the recursive structure on which the identification discussion relies. In the paper's own simulations and application q=1, so W_eta has no effect, but the general framework claims to cover q>1 recursive systems; without a concrete definition of W_eta, the general model is incomplete.","section":"§3.2, Eq. (11) and following"},{"comment":"The relationship between the fixed-effect terms f'_{ijm} and f_{ijm} is unclear and appears to duplicate the loading for the fixed effect. For the fixed effect, Appendix B gives f'_{ijm} = E^T eta_im and f_{ijm} = (I_M ⊗ eta_im)^T omega lambda_jm, with a'_{jm} + a_{jm} taking either 0 or 1. This suggests that f' encodes the first loading fixed to 1 and f encodes the free loading, but the notation is never explicitly tied to the identification constraint. The simulations in Section 6, however, estimate lambda_1 freely, so the intended meaning of f' versus f must be clarified, and the constraint used in estimation must be stated unambiguously.","section":"§2.3, Eq. (4)-(6)"}],"minor_comments":[{"comment":"The text says the MSEs are computed 'using an MCMC algorithm with 200 iterations', but the paper describes a Monte Carlo EM algorithm and no MCMC sampler is specified. Please clarify the sampling scheme or correct the terminology.","section":"§6.1"},{"comment":"The columns labelled phi_11, nu_11, phi_21, etc. are not clearly identified in the text; the reader must infer that these are estimates of the first eigenfunction and eigenvalue of the residual covariance operators. A brief explanation in the text or table caption would improve readability.","section":"Table 1"},{"comment":"The degrees of freedom df = p(p+1)/2 - k for the fit indices counts k estimated parameters, but for functional parameters the number of estimated coefficients depends on the basis truncation and is not defined here. Please specify how k is computed for the functional FSEM.","section":"§5"},{"comment":"The HRS application uses J=6 basis functions with only eight time points, but no sensitivity analysis or discussion of the choice of J is provided. Since the results may depend on this truncation, a brief robustness check would strengthen the application.","section":"§7"},{"comment":"The text introduces K_zeta_m(s,t) = exp(-2m|t-s|) for m=1,2,3 although the simulation has q=1 latent factor; it is likely that these are intended to be the covariance kernels of the residual functions or of the unique factors, but the notation is inconsistent and should be corrected.","section":"§6.2"}],"recommendation":"major_revision","confidential_remarks":"The identification gap is the principal obstacle. If the authors can impose a clear scale constraint in the estimation, re-run the simulations and the HRS analysis under that constraint, and either implement REML or revise the abstract, the paper would be a candidate for publication. The framework is original and the software availability is a positive feature, but the current version does not support the central quantitative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is the first FSEM paper I know that puts latent variables modeled as GPs into the structural equations, with fixed/concurrent/historical loadings and sparse longitudinal data. That gap is real, and the authors fill it with an explicit penalized EM framework, simulation evidence, an HRS application, and extensions of SEM fit indices. The core idea is worth engaging. But the paper has a load-bearing identification problem that the authors state and then do not actually enforce.\n\nSection 2.2 says the model is identified by fixing the first loading to 1; Section 6.1 and the HRS analysis estimate all loadings freely, and Table 1 reports MSE for lambda_1, which only makes sense if lambda_1 was estimated. Since the measurement model is invariant to scaling eta and all loadings, and the latent covariance is also estimated, the likelihood is flat along that direction. In the simulations the true covariance is known, but nothing in the estimation enforces the scale; in the real data there is no constraint at all, so the magnitudes and the significant sex effect statements are not well defined. This is more serious than the underspecified weight matrix W_eta in Section 3.2, which is a gap in the implementation details; the identification gap affects every loading and coefficient curve in the paper. This needs to be fixed before the results can be interpreted.\n\nOther soft spots are smaller. The abstract says restricted maximum likelihood, but the EM section and discussion say maximum likelihood; that is an inconsistency worth clarifying. The simulations lack Monte Carlo standard errors and any comparison to simpler alternatives. The W_eta matrix is described as crucial but no selection rule is given. The code repository is mentioned; I did not run it.\n\nThe strengths are real: the derivation is careful, the sparse and irregular setting is handled concretely, the HRS analysis is sensible, and the fit-index extension is useful. This is not a paper to desk reject. It deserves a rigorous referee who focuses on identification, then on W_eta and the REML/ML question. If the identification issue is resolved, I would expect it to make a solid contribution to functional SEMs.","headline":"Genuinely new latent-variable FSEM with GP factors and sparse-data machinery, but the stated identification constraint is not enforced, leaving loadings and regression effects scale-unidentified until fixed.","tokens_in":25750,"tokens_out":3453,"would_cite":false,"duration_ms":38363,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Functional structural equation models with Gaussian-process latent variables can be estimated from sparse and irregularly sampled longitudinal data using an EM algorithm with penalized smoothing.","keywords":["Functional structural equation model","Latent variables","Gaussian process","Sparse longitudinal data","EM algorithm","Functional factor loadings","Confidence bands","Goodness of fit indices"],"falsifier":"Simulate data with two latent curves that feed back on each other, fit the proposed EM with a recursive $W_{\\eta}$ that excludes the feedback, and check whether the factor-loading estimates stay unbiased and whether the goodness-of-fit indices detect the misspecification.","tokens_in":24694,"feed_emoji":"📈","tokens_out":10161,"duration_ms":91460,"temperature":0.7,"pith_summary":"This paper develops a family of functional structural equation models in which the latent variables and their indicators are random curves, modeled as Gaussian processes, so that the measurement and structural parts of the SEM are functional regressions. The authors aim to show that such models can be fitted to data where each curve is observed only at a few, possibly irregularly spaced time points, with missing observations, using a Monte Carlo EM algorithm and penalized-likelihood smoothing. If the framework works as claimed, applied researchers can study how a latent trait such as the General Factor of Personality relates to its indicators over time, without first imputing or smoothing the curves. The simulations report low mean squared errors and confidence-band coverage near the nominal level, and the authors note that the current implementation is limited to designs with a few hundred observations and tens of factors.","feed_headline":"Latent-curve SEMs can now be fit to sparse longitudinal data","feed_subtitle":"It estimates loadings and coefficients from irregularly timed curves, avoiding imputation and pre-smoothing.","key_machinery":"The load-bearing machinery is the conversion of functional latent-variable models into multivariate Gaussian models through basis expansions. Latent curves and residual curves are represented by coefficients in a common basis, and their covariance operators are encoded by Karhunen-Loève eigenfunctions and eigenvalues, so the measurement and structural equations become linear regressions with block matrices that depend on the observed design points. This makes the complete-data log-likelihood Gaussian, so the E-step reduces to computing conditional means and covariances of the latent curves, and the M-step reduces to ridge-type penalized least squares with derivative-based penalty matrices. The recursive structural assumption and the weight matrix $W_{\\eta}$, which encodes which latent factors influence which, tie the latent curves together and determine their joint covariance; the paper presents the choice of $W_{\\eta}$ as crucial and case-specific.","core_discovery":"On its own terms, the paper's central claim is that a recursive functional SEM with Gaussian-process latent factors admits a tractable likelihood-based estimation procedure. The measurement model expresses each observed indicator curve as a functional intercept plus fixed, concurrent, or historical effects of latent curves, and the structural model expresses each latent curve as a function of other latent curves and scalar or functional covariates. By expanding the curves in basis functions and using Karhunen-Loève expansions for the residual and latent covariance operators, the infinite-dimensional model is turned into a finite-dimensional Gaussian model whose complete-data likelihood is tractable. The EM algorithm draws the unobserved latent curves from Gaussian generator distributions and updates the functional coefficients, covariance operators, and measurement-error variances in closed form, with smoothing penalties selected by cross-validation. Simulation results show accurate factor loadings and regression coefficients, and confidence bands with coverage close to 95%, under regular, irregular, and missing-at-random designs.","pith_inferences":["A data-driven procedure for selecting the recursive structure $W_{\\eta}$—for example, cross-validating over candidate direction-of-effect matrices—would remove the main expert input the framework currently requires.","The stated scalability limit suggests that variational or stochastic EM approximations would be needed before the method reaches the large panel datasets common in biobanks and administrative records.","The same basis-expansion and EM machinery could plausibly be adapted to non-Gaussian functional indicators, such as binary symptom curves, by thresholding latent Gaussian processes as the authors mention as a future direction.","The HRS sex difference should be read as a model demonstration rather than a causal finding, since the analysis assumes a particular recursive structure and item-parceled indicators."],"forward_implications":["Sparse and irregularly sampled longitudinal data can enter a latent-variable SEM directly, without imputation or pre-smoothing, as long as the missingness pattern is compatible with the model.","Factor loadings can be time-varying or history-dependent, so the fitted measurement model can reveal how a latent trait relates to its indicators over time.","Confidence bands around functional loadings and regression coefficients let practitioners decide whether apparent time dynamics are real rather than noise.","The functional goodness-of-fit indices extend the standard SEM toolkit, allowing overall and pointwise assessment of model fit for curve-valued data.","In the Health and Retirement Study application, the model recovers the expected Big Five pattern for the General Factor of Personality and a higher score for women, illustrating a concrete use."],"supporting_citations":[{"why":"Supplies the functional regression and basis-expansion framework on which the FSEM is built.","marker":"Ramsay & Silverman (2005)"},{"why":"Provides the functional principal component approach used to estimate kernel functions and to handle sparse longitudinal observations.","marker":"Yao et al. (2005)"},{"why":"Defines the historical functional linear model that the paper adopts for historical effects of latent curves.","marker":"Malfait & Ramsay (2003)"},{"why":"Earlier functional structural equation model for observed variables, which the present framework extends to latent functional variables.","marker":"Lee & Li (2022)"},{"why":"Gives the geometric confidence-region construction used to build bands around estimated functional parameters.","marker":"Choi & Reimherr (2018)"},{"why":"Provides the longitudinal data for the GFP application and the missingness patterns the method handles.","marker":"Health and Retirement Study (2024)"}],"fun_headline_variants":["FSEM with GP latent variables handles sparse functional data","Latent curves as Gaussian processes: sparse SEMs solved","No imputation needed: new FSEM fits sparse curves","Functional SEM estimates latent curves from sparse observations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the analyst correctly specifying the recursive network among latent factors (the weight matrix $W_{\\eta}$), and the paper gives no data-driven rule for this choice even though it determines the joint covariance of the latent curves.","fun_headline_variants_meta":{"raw":{"variants":["FSEM with GP latent variables handles sparse functional data","Latent curves as Gaussian processes: sparse SEMs solved","No imputation needed: new FSEM fits sparse curves","Functional SEM estimates latent curves from sparse observations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1370,"prompt_tokens":983,"completion_tokens":387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":599,"tokens_out":387,"duration_ms":4546,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:48:02.958506+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data with two latent curves that feed back on each other, fit the proposed EM with a recursive $W_{\\eta}$ that excludes the feedback, and check whether the factor-loading estimates stay unbiased and whether the goodness-of-fit indices detect the misspecification.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the functional regression and basis-expansion framework on which the FSEM is built."},{"cited_title":"& Wang, J.-L","cited_arxiv_id":null,"evidence_quote":"Provides the functional principal component approach used to estimate kernel functions and to handle sparse longitudinal observations."},{"cited_title":"& Ramsay, J","cited_arxiv_id":null,"evidence_quote":"Defines the historical functional linear model that the paper adopts for historical effects of latent curves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier functional structural equation model for observed variables, which the present framework extends to latent functional variables."},{"cited_title":"& Reimherr, M","cited_arxiv_id":null,"evidence_quote":"Gives the geometric confidence-region construction used to build bands around estimated functional parameters."}],"review_version":1}