{"id":"915f640a-96f4-402f-ae73-fe104e4b61b3","arxiv_id":"2504.12750","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid model that estimates spatial dependence by maximum likelihood and then feeds spatially adjusted functional and scalar inputs into a deep neural network predicts spatially dependent outcomes with lower test error than linear spatial and non-spatial deep learning baselines.","lead":"This paper proposes a spatial functional deep neural network that adds a spatial autoregressive correction to functional deep learning predictions. The authors claim it predicts spatially dependent functional outcomes, such as COVID-19 death counts, more accurately than two existing alternatives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The simulation DGP in Section 5 contains no link function and even applies an extra (I−ρW)^{-1} to ε, so the claimed nonlinear predictive advantage of SFDNN is untested.","rationale":"The reader's weakest assumption is that the simulations never generate data from the nonlinear model (2.4) because no link function appears in the data-generating equations. Reading Section 5 confirms this: the DGP is a linear SSoFRM, and the only scenario variation is the error distribution. Even the phrase 'under the exponential and sigmoidal link functions' refers to scenarios that are never implemented. This is load-bearing because the manuscript's claim is about a 'non-linear modeling framework' whose practical value is demonstrated almost entirely by the Monte Carlo results; if the simulations do not exercise the nonlinear regime, the evidence cannot support the strongest claim. My review also flags a secondary issue the reader did not emphasize: the written DGP includes an additional (I−ρW)^{-1} on ε, which is inconsistent with the iid-error likelihood used to estimate ρ in Section 3 and may explain the inflated ML MSPE at high ρ. This makes the simulation evidence doubly unreliable as written. The application section contains additional reporting inconsistencies (MSPE compared with RMSE; 'forecast the average number of deaths for 2024' despite testing in 2022), but the simulation mismatch is the primary reason the empirical claim is not credible. Because these problems undermine the central evidence rather than merely weakening it, the reader's REJECT verdict remains appropriate; no verdict adjustment is needed.","tokens_in":19402,"tokens_out":7307,"duration_ms":73544,"concrete_test":"Re-run the Monte Carlo with responses drawn from the model SFDNN actually targets, e.g., Y = g((I−ρW)^{-1}(β0 1n + ∫Xβ du + ZΓ)) + ε for g(v)=v, exp(v), and logistic(v), with ρ=0.5,0.9 and t3/exp errors, and recompute Tables 1–3 MSPE. If SFDNN's advantage shrinks or reverses under nonlinear g, the central claim fails. Separately, inspect the GitHub code: if it literally contains the extra (I−ρW)^{-1} before ε, rerun with ε inside the inverse exactly once and check whether the ML baseline's ρ=0.9 MSPE drops toward SFDNN's.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that SFDNN significantly outperforms ML and FDNN, especially under nonlinear relationships and strong spatial dependence—rests on Monte Carlo evidence that never simulates from the proposed NSSoFRM in Eq (2.4). Section 5 states that the response is generated as Y=(I−ρW)^{-1}{β0 1n + ∫Xβ du + ZΓ + (I−ρW)^{-1}ε}, which has no link function g(·) on the conditional mean. The only differences across Tables 1–3 are the error distributions (N(0,1), t3, exp(1)); the mean function remains linear in the spatially transformed predictors. The text says FDNN and SFDNN are expected to outperform 'particularly when nonlinear transformations are present (i.e., under the exponential and sigmoidal link functions)', but those scenarios are never run. In addition, the written DGP applies (I−ρW)^{-1} to ε inside the braces, creating an error covariance different from the iid ε assumed in the ML likelihood of Section 3; the very large ML MSPE at ρ=0.9 could therefore reflect an error-covariance misspecification rather than a genuine advantage of SFDNN. No consistency argument is given for plugging the linear-model ML \\hatρ into a nonlinear network. Without simulations generated from Eq (2.4), the headline outperformance claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatial functional deep neural network (SFDNN) for predicting a scalar response from spatially dependent functional and scalar predictors. The model is defined in Eq. (2.4) as a nonlinear scalar-on-function regression with a spatial autoregressive component, estimated in two steps: first the spatial autocorrelation parameter ρ is estimated by maximum likelihood under an underlying linear spatial scalar-on-function model (Section 3), then a functional deep neural network is trained on spatially adjusted covariates (Section 4). The paper presents Monte Carlo simulations comparing SFDNN with an ML-based linear model and a functional deep neural network (FDNN) under Gaussian, t3, and exponential errors at three levels of spatial dependence, and applies the method to Brazilian COVID-19 data. The central claim is that SFDNN significantly outperforms both ML and FDNN in predictive accuracy, especially under strong spatial dependence and non-Gaussian errors.","tokens_in":19749,"tokens_out":5607,"duration_ms":58981,"significance":"If the claimed predictive advantage were valid, SFDNN would be a practically useful addition to the spatial-functional regression literature, combining flexible nonlinear modeling with autoregressive spatial dependence. The paper provides a reproducible-code link and a real-data application, which are strengths. However, the significance is currently limited because the simulation study does not generate data from the proposed nonlinear model in Eq. (2.4), and the reported ML baseline is misspecified in the data-generating process. The empirical evidence for the headline claim is therefore not established; the contribution is a plausible algorithmic idea rather than a validated method.","major_comments":[{"comment":"The data-generating process is Y = (I_n - ρW)^{-1}{β0 1_n + ∫ X(u)β(u) du + ZΓ + (I_n - ρW)^{-1}ε}, which contains no link function g(·) on the conditional mean. This is not a draw from the proposed NSSoFRM in Eq. (2.4), where the conditional mean is g((I_n - ρ̂W)^{-1}(...)). Consequently, none of the Monte Carlo scenarios evaluates the nonlinear component that is claimed to be SFDNN's key advantage. The text at the end of Section 5 explicitly expects superiority 'particularly when nonlinear transformations are present (i.e., under the exponential and sigmoidal link functions)', but such scenarios are never run; only the error distribution changes across Tables 1–3. This is a load-bearing gap: the headline claim of 'significantly outperforms ... particularly in the presence of high spatial dependence and non-Gaussian errors' is supported only by simulations that never exercise a nonlinear regression function.","section":"Section 5, simulation DGP (equation after 'The response variable ... is generated according to')"},{"comment":"The error term enters the simulation as (I_n - ρW)^{-1}ε inside the braces, so the conditional error covariance is σ^2 (I_n - ρW)^{-1}(I_n - ρW)^{-T}, not σ^2 I_n. However, the ML estimation in Section 3 is derived under the assumption ε ~ N(0, σ^2 I_n), and the likelihood (Section 3) uses residuals Y - ρWY - X_c θ, which under the simulation DGP equal X_c θ + (I_n - ρW)^{-1}ε. The ML baseline is therefore fitted under a misspecified error covariance, and the very large ML MSPE at ρ = 0.9 (e.g., 37.165 in Table 1 for n_train=100) may reflect this covariance misspecification rather than an inherent advantage of SFDNN. The comparison is unfair to ML. The authors should simulate from the linear SSoFRM (2.3) with iid ε, and separately from the nonlinear model (2.4) with specified link functions, to obtain interpretable comparisons.","section":"Section 5, simulation DGP and Section 3 likelihood"},{"comment":"No consistency or identifiability argument is given for plugging the MLE ρ̂ from the linear SSoFRM into the nonlinear model (2.4). The paper asserts in Section 2 that applying (I_n - ρ̂W)^{-1} before the activation 'ensures that the spatial dependence structure is preserved', but this is not demonstrated. If the true conditional mean is nonlinear, the linear-model MLE of ρ need not be consistent for the autoregressive parameter of the nonlinear process, and the architecture in Eq. (4.6) is not guaranteed to represent E[Y | X, Z]. At minimum, a simulation generated from Eq. (2.4) with known g (e.g., exponential, sigmoidal, or identity) is needed to test whether the two-step estimator recovers the spatial dependence and yields valid predictions. Without this, the estimation procedure rests on an untested assumption that is load-bearing for the method's validity.","section":"Section 4.1 and Eq. (2.4)"},{"comment":"The application section contains reporting inconsistencies that obscure the model comparison. The text states SFDNN achieves an MSPE of 0.253, while 'ML and FDNN models exhibit higher RMSE values of 0.708 and 1.106', but MSPE and RMSE are not comparable on the same scale (sqrt(0.253) ≈ 0.503). The text also says the fitted models use 2022 predictors to forecast 'the average number of deaths for 2024', yet the testing period is later described as 2022 and Figure 7 labels the testing data as 2022. The prediction target and the reported error metrics should be clarified and made consistent, otherwise the empirical advantage of SFDNN is not quantitatively interpretable.","section":"Section 6, Brazilian COVID-19 application"}],"minor_comments":[{"comment":"The column header 'MPSE' in Tables 1–3 should be 'MSPE' to match the text and standard terminology.","section":"Section 5, Tables 1–3"},{"comment":"The description 't3 (t-distribution with non-centrality parameter' is incomplete and imprecise; the authors should specify that this is a Student t distribution with 3 degrees of freedom and zero noncentrality (or state the noncentrality explicitly).","section":"Section 5, error distributions"},{"comment":"In the description of the adaptive bandwidth, the text says 'h closest neighbors are identified', then defines H_i as the largest distance among 'these four nearest neighbors'; the 'four' appears to be a leftover and should be 'h' for consistency.","section":"Section 6, spatial weight construction"},{"comment":"Reference 'Reiss, P. T., Goldsmith, J., Shang, H. L. and Odgen, R. T. (2017)' and 'Reiss, P. T., Goldsmith, J., Shang, H. L. and Ogden, R. T. (2017)' are duplicates with a typo in the first author name; one should be removed.","section":"References"},{"comment":"The abbreviation 'SSFDNN' appears in the paragraph 'The proposed SSFDNN allows us to flexibly model...' and should be 'SFDNN' for consistency with the rest of the paper.","section":"Section 6, terminology"},{"comment":"The paper repeatedly claims 'significantly outperforms' without reporting any statistical significance tests for the differences in MSPE across Monte Carlo simulations; adding confidence intervals or tests would strengthen the claim.","section":"Abstract and Section 7"}],"recommendation":"major_revision","confidential_remarks":"The simulation flaw in Section 5 is central: the Monte Carlo evidence never generates data from the proposed nonlinear model (2.4), and the ML baseline is fitted under a misspecified error covariance. If the authors cannot provide corrected simulations that (a) generate from the nonlinear model with explicit link functions and (b) use an error structure consistent with the likelihood, the paper's empirical claims are unsupported and the manuscript should not be published. A major revision is warranted only if those corrections are feasible and the results still favor SFDNN."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the SFDNN idea is sensible and clearly written, but the simulation evidence for its headline nonlinear advantage doesn't actually test the model in Eq (2.4).\n\nWhat's new is the specific integration: estimate rho by ML under a linear SSoFRM, then feed the spatially adjusted functional and scalar covariates into a functional deep neural network. That plug-in architecture appears to be original, and the algorithmic description in Section 4 is understandable. The authors also ship code on GitHub and use a real application with a sensible spatial weighting scheme. Those are genuine merits.\n\nThe soft spot is load-bearing. The data generating process in Section 5 is Y = (I-rhoW)^-1 { beta0 + sum X beta du + Z Gamma + (I-rhoW)^-1 epsilon }. That is not the NSSoFRM in Eq (2.4), because there is no link function g on the conditional mean. The promised \"exponential and sigmoidal link functions\" never appear; the three scenarios differ only by error distribution. So the paper's central claim about capturing nonlinear relationships is unsupported by its own Monte Carlo design. The extra (I-rhoW)^-1 on epsilon also misspecifies the ML baseline's error covariance, which likely inflates ML's MSPE at rho=0.9 and makes SFDNN look better than it should. That is a real problem, not a quibble.\n\nMinor issues: the application text mixes MSPE and RMSE in one sentence (\"lowest MSPE of 0.253\" followed by \"higher RMSE values of 0.708 and 1.106\") and should be made consistent. Also, no consistency argument is given for plugging a linear-model ML rho into a nonlinear network; the authors treat it as a heuristic, which is fine if stated as such, but the simulations don't probe how robust that choice is.\n\nWhat survives the critique: the evidence that SFDNN handles strong spatial dependence better than plain FDNN is real, even under the misspecified DGP. The nonlinear advantage is untested, not disproven.\n\nWho gets value: researchers working on spatial functional prediction who want a practical deep-learning baseline and are willing to treat the methodology as a heuristic rather than a theory. It deserves a serious referee, but with major revision: the simulations should be regenerated from Eq (2.4) with actual link functions, the ML baseline should be correctly specified for the error covariance, and the application reporting cleaned up. I would not cite the results in their current form, but I'd read a revision.","headline":"A reasonable SAR-plus-FDNN hybrid whose main empirical claim about nonlinear gains is undermined by a simulation DGP that never generates from the proposed nonlinear model.","tokens_in":20242,"tokens_out":1641,"would_cite":false,"duration_ms":20084,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A spatial deep network for functional data claims better prediction when locations interact and errors are non-Gaussian.","keywords":["spatial functional data","functional deep neural network","spatial autoregressive model","scalar-on-function regression","functional principal component analysis","non-Gaussian errors","predictive accuracy","COVID-19 mortality"],"falsifier":"A simulation where the response is generated from the nonlinear model in Eq (2.4) with a known link function, known $\\rho$, and a spatial weight matrix, and then compared across ML, FDNN, and SFDNN on out-of-sample prediction error; if SFDNN's advantage disappears or the ML estimate of $\\rho$ becomes badly biased as the nonlinearity strengthens, the central claim would be undermined.","tokens_in":19219,"feed_emoji":"📈","tokens_out":2282,"duration_ms":25196,"temperature":0.7,"pith_summary":"This paper tries to establish that predictions from spatially dependent functional data improve when a spatial autoregressive term is folded into a deep neural network. It proposes a spatial functional deep neural network (SFDNN) that first estimates the spatial dependence parameter by maximum likelihood and then feeds spatially adjusted functional and scalar covariates through a deep network. The authors argue this hybrid preserves the spatial structure while allowing nonlinear regression, and they support the claim with Monte Carlo simulations covering Gaussian, heavy-tailed, and skewed errors under weak to strong spatial dependence. The paper also reports an application to Brazilian COVID-19 data where the proposed model obtained the lowest mean squared prediction error in the testing phase. A sympathetic reader would take the central claim to be that explicitly modeling spatial dependence inside a functional deep learner yields practically better out-of-sample predictions than either linear spatial functional regression or a functional deep network without spatial adjustment.","feed_headline":"Spatial deep learning predicts functional data with lower error","feed_subtitle":"A new model folds spatial dependence into a deep network, beating linear and non-spatial rivals in simulations and COVID-19 forecasting.","key_machinery":"The central object is the neuron activation in Eq (4.6), where the first hidden layer applies the spatial inverse matrix $(I_n-\\hat{\\rho}W)^{-1}$ to the functional principal component integrals before the nonlinear activation $g(\\cdot)$ is applied. The spatial parameter $\\hat{\\rho}$ is estimated once by maximum likelihood from a linear spatial scalar-on-function regression, then treated as fixed; the functional weights are represented as B-spline basis expansions, and Adam optimization trains the network. This two-step construction is what lets the model combine a spatial autoregressive mechanism with a flexible nonlinear functional regression.","core_discovery":"The paper's central claim is that the proposed SFDNN model, built from a nonlinear spatial scalar-on-function regression framework, significantly improves predictive accuracy for spatially dependent functional data compared with maximum-likelihood-based spatial functional linear regression (ML) and a functional deep neural network (FDNN) that ignores spatial structure. The improvement is said to be largest when spatial autocorrelation is strong (rho = 0.9) and when errors are heavy-tailed t3 or skewed exponential, conditions under which the comparators degrade sharply. In the COVID-19 application, all models produced similar training mean squared errors, but SFDNN achieved the lowest mean squared prediction error in the test period, which the authors interpret as evidence of superior generalization.","pith_inferences":["Editorial inference: The paper's simulations never generate responses from the nonlinear model in Eq (2.4), because no link function appears in the data-generating equations; if the true process is nonlinear, the maximum-likelihood estimate of $\\rho$ may become inconsistent, so the reported advantage could depend on the spatial parameter staying valid under misspecification.","Editorial inference: A natural testable extension is to replace the fixed ML estimate of $\\rho$ with an alternating or jointly estimated spatial parameter inside the network; the paper leaves open whether such coupling would improve predictions when spatial dependence and nonlinearity interact.","Editorial inference: Because the spatial weights in the application are built from k-nearest-neighbor bi-square kernels on great-circle distances, the method's practical performance likely varies with the choice of neighborhood size and weight kernel, which the paper does not systematically explore.","Editorial inference: The framework could be extended to spatiotemporal functional data by letting the spatial weight matrix and the network inputs evolve over time, a direction the paper explicitly lists as future work."],"forward_implications":["If SFDNN's reported advantage holds, practitioners with spatially indexed functional predictors can expect materially lower out-of-sample prediction error than linear spatial functional models, particularly under strong spatial autocorrelation.","The model extends spatial functional regression to multiple functional and scalar covariates simultaneously, which broadens its applicability beyond single-functional-predictor settings.","The reported robustness to heavy-tailed and skewed errors suggests SFDNN is a candidate default for spatial functional prediction problems where Gaussian assumptions are questionable.","The two-step estimation strategy, fixing the spatial parameter before network training, offers a computationally tractable way to inject spatial dependence into deep learning without re-estimating the spatial structure in every gradient step.","In epidemiological forecasting, the COVID-19 results imply that spatial spillover and nonlinear interactions among cases, vaccination, and population can be captured in a single prediction tool."],"supporting_citations":[{"why":"Supplies the functional deep neural network architecture with functional weights expanded in basis functions, which the proposed SFDNN extends with spatial adjustment.","marker":"Thind et al. (2023)"},{"why":"Provides the extended spatial scalar-on-function regression framework and data-generation setup that the simulations build on.","marker":"Beyaztas, Mandal and Shang (2025)"},{"why":"Provides the COVID-19 dataset used in the empirical application.","marker":"Guidotti and Ardia (2020)"},{"why":"Supplies the local Moran's I statistic used to diagnose and visualize spatial clustering in the COVID-19 mortality data.","marker":"Anselin (1995)"},{"why":"Underpins the B-spline basis expansion used to parameterize the infinite-dimensional functional weights in the network.","marker":"Boor (2001)"},{"why":"Provides the Adam optimization algorithm used to train the FDNN and SFDNN models.","marker":"Kingma and Ba (2014)"}],"fun_headline_variants":["Spatial deep net trims prediction error for functional data","Deep learning with spatial dependence beats linear rivals","Spatial functional nets: lower error in COVID forecasts","New spatial deep model improves functional data predictions","Spatial-aware deep network outperforms in predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the spatial dependence parameter $\\rho$ estimated by maximum likelihood under a linear spatial functional model remains valid when the regression function is actually nonlinear, and that applying the inverse spatial matrix before the activation preserves the true spatial dependence.","fun_headline_variants_meta":{"raw":{"variants":["Spatial deep net trims prediction error for functional data","Deep learning with spatial dependence beats linear rivals","Spatial functional nets: lower error in COVID forecasts","New spatial deep model improves functional data predictions","Spatial-aware deep network outperforms in predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2324,"prompt_tokens":885,"completion_tokens":1439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1364}},"tokens_in":501,"tokens_out":1439,"duration_ms":8595,"temperature":1.0,"reasoning_tokens":1364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:24:35.415877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A simulation where the response is generated from the nonlinear model in Eq (2.4) with a known link function, known $\\rho$, and a spatial weight matrix, and then compared across ML, FDNN, and SFDNN on out-of-sample prediction error; if SFDNN's advantage disappears or the ML estimate of $\\rho$ becomes badly biased as the nonlinearity strengthens, the central claim would be undermined.","supporting_citations":[{"cited_title":"Spatial function-on-function regression","cited_arxiv_id":"2412.17327","evidence_quote":"Supplies the local Moran's I statistic used to diagnose and visualize spatial clustering in the COVID-19 mortality data."}],"review_version":1}