{"id":"1bb27bf1-34dc-4963-879d-fc71a930aff6","arxiv_id":"2507.12033","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Age-structured Bayesian spatiotemporal models outperform indirect standardization for suicide-call data and reveal that younger age groups experienced the steepest risk increase from 2021 onward.","lead":"This paper compares the standard method for calculating disease risk by age group, indirect standardization, with more flexible Bayesian models that let age patterns vary across places and years. Applied to suicide-related emergency calls in the Valencian Community, the flexible models fit better and suggest that younger age groups saw the largest risk increase after 2020.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Model 8's time-age effect is driven by a single post-2020 shift; the paper offers no uncertainty or validation that it reflects suicide risk rather than reporting behavior.","rationale":"The reader's verdict is CONDITIONAL (high confidence), with correctness risk medium. The reader's weakest assumption is exactly the outcome-validity concern: suicide-related calls are treated as a direct proxy for suicide risk without accounting for differential help-seeking or call behavior. I agree with that as the most load-bearing concern. It is load-bearing because the paper's strongest claim is about a hidden age-time pattern in risk, and every inferential step depends on the outcome measure being a valid proxy for risk. The paper provides no discussion of this issue, no validation against mortality data, and no uncertainty quantification for the specific random effects that carry the headline finding. My second concern (no credible intervals on Figure 2/5 effects) is concrete and testable but secondary to the proxy-validity issue; the recommended test combines both: compute credible intervals for the time-age effect, and ideally compare against official mortality data. Under either failure, the conclusion would need to be downgraded to a statement about calls, not risk. Since the mathematical machinery of the modeling is sound and the model comparison is thorough, the verdict should remain CONDITIONAL rather than REJECT, but the conditions should include demonstrating statistical uncertainty and addressing outcome validity. The reader focused on the proxy issue and uncertainty; my recommended concrete test adds specificity and a direct falsification path.","tokens_in":20204,"tokens_out":1635,"duration_ms":17145,"concrete_test":"Re-fit Model 8 and compute posterior credible intervals (e.g., 95% intervals) for the time-age interaction exp(ζ3_jk) at, say, age groups [15,20) and [40,45), and compare the 2019 vs 2021 posterior distributions; if the intervals overlap substantially, the paper's claimed post-2020 youth increase is not statistically supported. Additionally, if possible, validate by re-estimating the same age-time model on official suicide mortality counts for the Valencian Community for 2017–2022; if the youth mortality trend does not show a similar post-2020 increase, the call-data pattern is likely reporting-driven rather than a risk pattern.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that the best model (Model 8) reveals a rising temporal trend and stronger risk increases among younger individuals, with an increase in the time effect in younger age groups starting in 2020. The load-bearing inference is that this pattern reflects underlying suicide risk. Two issues make this insecure. First, the paper offers no measure of uncertainty for the plotted posterior means in Figures 2 and 5; the summary statistics in Table 4 give posterior intervals only for hyperparameters, not for the age-group and time-age random effects. The concluding claims about a post-2020 increase in younger age groups therefore have no stated precision. Second, and more fundamentally, the data are suicide-related emergency calls, not suicide deaths or attempts. If help-seeking or call behavior changed by age or by year — for example, post-2020 awareness campaigns or differential willingness to call among younger people — the estimated time-age interaction would capture reporting behavior, not risk. The paper explicitly treats calls as a proxy for suicide risk but never discusses this ascertainment issue, despite citing literature that calls are influenced by neighborhood and social factors. Figure 5's pattern is visually striking, but a visual pattern with no uncertainty bands and no external validation against national suicide statistics (e.g., INE mortality data) cannot support the strong conclusion about a youth-specific post-COVID increase. The model-selection exercise (520 WAIC comparisons) is extensive, but WAIC selects among models that all share the same proxy assumption; it does not validate the outcome measure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that indirect standardization in disease mapping relies on a proportionality assumption that is often violated, and proposes age-structured hierarchical Bayesian spatiotemporal models as an alternative. Using suicide-related emergency calls in the Valencian Community for 2017–2023, the authors fit 520 model variants with INLA, including space-time, space-age, and time-age interactions, and select the best model by WAIC. The best model (Model 8) has an unstructured time effect, an RW1 age effect, and Type II, I, and IV interactions; it reportedly reveals a rising temporal trend, a nonlinear age pattern, and a post-2020 increase in the time-age effect for younger age groups. The methodological claim is that such age-specific temporal patterns would be missed under indirect standardization.","tokens_in":20472,"tokens_out":5001,"duration_ms":53026,"significance":"If the empirical results were supported by uncertainty quantification and validation, the paper would make a useful applied contribution by demonstrating the practical consequences of the proportionality assumption and by providing a transparent comparison of 520 interaction structures. The manuscript has notable strengths: the supplementary table reports all 520 WAIC values, the identifiability constraints for the interaction models are explicitly tabulated, and prior sensitivity is examined with non-informative priors. However, the central empirical claims—especially the youth-specific post-2020 increase—are currently based on posterior means of random effects without credible intervals, and the outcome variable (emergency calls) is treated as a direct proxy for suicide risk without discussion of ascertainment. These gaps make the headline findings not yet convincing.","major_comments":[{"comment":"The plotted posterior means of exp(δ), exp(γ), and exp(ζ3) in Figures 2 and 5 are presented without credible intervals, and Table 4 reports posterior intervals only for hyperparameters. Because the abstract's claims of a rising temporal trend and stronger risk increases among younger individuals are statements about these random effects, the absence of uncertainty bands means the reader cannot judge whether the 2020–2023 increase in the time-age interaction for ages 10–35 is distinguishable from noise. I request pointwise credible intervals or posterior probability statements for the differences that define the conclusion.","section":"Section 4, Figures 2 and 5, Table 4"},{"comment":"The outcome is suicide-related emergency calls, not verified suicide deaths or attempts. The paper explicitly uses these calls as a proxy for suicide risk but does not discuss ascertainment bias: if younger people became more likely to call after 2020 (e.g., due to awareness campaigns) or if call rates are influenced by neighborhood and social factors (as cited from Marco et al. 2024), the estimated time-age interaction would reflect reporting behavior rather than underlying risk. This is load-bearing for the empirical conclusion and should be addressed, ideally by validating against mortality data or by discussing the direction and possible magnitude of the bias.","section":"Section 4, data description"},{"comment":"Model selection relies solely on WAIC over the 520 fitted models, with no hold-out validation or cross-validation checks. The WAIC improvement of Model 8 over Model 7 is only 174.87 units on a scale where differences between some models exceed 1,000 units; without validation, the superiority of Model 8 and the robustness of its interaction structure are not established. A small simulation study, a temporal hold-out, or bootstrap-based model selection would strengthen the claim.","section":"Section 3.2 and Table 3"},{"comment":"The proportionality assumption is tested only through Loess-smoothed scatterplots for a selected set of municipality-years, which is a subjective graphical procedure. Since the motivation for the modeling approach rests on the claim that proportionality is violated in these data, I recommend a formal statistical test or a quantitative summary of the deviations across all S×T units.","section":"Section 2 and Figure 1"}],"minor_comments":[{"comment":"The age effect is defined as γ = (γ1, . . . , γT)′ but should be indexed by K age groups, not T time periods.","section":"Section 3, after Eq. (3)"},{"comment":"'the reduction in the reduction of the suicide-related emergency calls rate' appears to be a typo and should read 'a reduction in the suicide-related emergency calls rate'.","section":"Section 4, paragraph on Figure 7"},{"comment":"The text cites 'Morris, 2017' and 'Wakefield et al., 2017', but the reference list contains Morris (2001) and Wakefield et al. (2001); the citations and the reference list need to be reconciled.","section":"References"},{"comment":"The text refers to 'Figure 7 shows the spatio-temporal effect' immediately after Figure 2, but Figure 7 in the supplementary material is the full space-age effect; the in-text figure cross-references are inconsistent.","section":"Figure numbering"},{"comment":"The abstract states the temporal trend rises over 2017–2022, while the text of Section 4 describes a stabilization between 2022 and 2023; please reconcile these statements.","section":"Abstract and Section 4"},{"comment":"The statement that indirect standardization assumes 'age-specific risks are homogeneous across regions and time periods' is imprecise; the assumption in Eq. (2) is that the age-specific rates qk are common up to a multiplicative spatiotemporal risk θij, so proportionality, not homogeneity, is the issue.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's applied-statistics scope. The authors should be encouraged to provide the INLA code and data, as reproducibility would increase the value of the 520-model comparison. The main empirical claim is suggestive but currently over-sold relative to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid, workmanlike application of age-space-time CAR models to suicide-related emergency calls in the Valencian Community. The methodology itself is not new—Knorr-Held (2000) interactions, Goicoa et al. (2016) CAR, INLA—and the \"overcoming standardization\" pitch has been made before by Congdon and Perez-Panades. What is new is the specific empirical application and the systematic comparison of 520 model specifications, which is a real piece of effort and genuinely useful as a template for model selection in this subfield.\n\nThe good stuff: the model comparison is thorough, with identifiability constraints handled correctly following Goicoa et al. (2018). The prior sensitivity check (Table 4, PC vs non-informative) is honest and shows the main results are not driven by prior choice. The data description and the exploratory proportionality check via the Wakefield-style plots are transparent. I believe the analysis is internally coherent.\n\nThe soft spots are concentrated in the inference from the best model. First, the central empirical claim—a post-2020 increase in the time-age effect concentrated in younger groups—is presented with posterior means only, no credible intervals. Figure 5 is visually striking, but without intervals or a formal comparison (e.g., is the 2020-2021 shift in the 10-25 groups distinguishable from noise?), the claim is under-supported. Second, and more fundamental, suicide-related emergency calls are a proxy for risk, not risk itself. The authors themselves cite work showing calls are influenced by neighborhood and social factors; if help-seeking or willingness to call changed differentially by age after 2020, the time-age interaction would capture reporting behavior, not suicide risk. The paper should acknowledge this directly and, ideally, cross-check against mortality data (e.g., INE). Third, WAIC-based selection among 520 models with no holdout validation is common in this literature, but it means the \"best\" model's superiority on fit does not translate directly into predictive or explanatory validity. The Loess-based proportionality test is exploratory; fine as motivation, but it is not a formal test.\n\nNone of these are fatal. The empirical pattern is plausible and worth taking seriously. The paper deserves a serious referee. I would send it to review, and I'd ask for the uncertainty quantification and a frank discussion of the proxy issue before accepting. I would not cite it myself, but I'd bring it to the reading group if we were doing disease-mapping or small-area health work.","headline":"A competent, extensive age-space-time CAR model comparison applied to suicide calls; the youth post-2020 finding is plausible but needs uncertainty intervals and a frank handling of the calls-as-proxy assumption.","tokens_in":21095,"tokens_out":3083,"would_cite":false,"duration_ms":34758,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P10","62M30","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that indirect standardization hides a post-2020 rise in suicide-related emergency calls among younger people, which an age-structured Bayesian spatiotemporal model reveals.","keywords":["indirect standardization","disease mapping","age-structured hierarchical Bayesian models","spatiotemporal interactions","INLA","suicide-related emergency calls","proportionality assumption","WAIC"],"falsifier":"Re-run the best model on suicide mortality records for the same region and years: if the post-2020 increase in younger age groups disappears or reverses while the model structure is unchanged, the paper's central pattern is an artifact of emergency-call reporting rather than a property of suicide risk.","tokens_in":20084,"feed_emoji":"📈","tokens_out":7109,"duration_ms":71440,"temperature":0.7,"pith_summary":"Indirect standardization, the default way to filter out age when mapping disease risk, assumes that age effects are the same in every place and every year; the paper shows this assumption breaks down for suicide-related emergency calls in the Valencian Community between 2017 and 2023. The authors replace the standardized analysis with age-structured hierarchical Bayesian models that include spatial, temporal, and age random effects plus space-time, space-age, and time-age interactions. Fitting 520 such models, they find that adding age effects improves fit substantially, and the best model reveals a rising temporal trend, a nonlinear age pattern, and a post-2020 increase in risk concentrated among younger age groups. If this is right, conventional standardized estimates would have hidden exactly the age-specific shift that matters for prevention policy.","feed_headline":"Age-standardized maps miss post-2020 rise in youth suicide calls","feed_subtitle":"Keeping age in the model, not standardizing it away, reveals which groups drive the recent increase.","key_machinery":"The load-bearing object is the age-structured hierarchical Bayesian spatiotemporal model with log-linear predictor, where the risk $\\Theta_{ijk}$ in area $i$, year $j$, and age group $k$ is decomposed into an intercept, spatial $\\phi$, temporal $\\delta$, and age $\\gamma$ main effects, plus additive pairwise interactions $\\zeta^1$ (space-time), $\\zeta^2$ (space-age), and $\\zeta^3$ (time-age). The interaction terms are built by Kronecker products of the precision matrices of the corresponding main effects, following the Type I-IV typology of Knorr-Held (2000), with identifiability constraints from Goicoa et al. (2018). This machinery lets the age pattern vary across space and time rather than forcing a common multiplicative age schedule, which is exactly the assumption that indirect standardization imposes and that the paper argues biases conventional estimates. INLA makes fitting 520 candidate models computationally feasible, and WAIC selects the final model.","core_discovery":"On the paper's own terms, the discovery is that the proportionality assumption behind indirect standardization, $E(O_{ijk}/N_{ijk}) = \\theta_{ij} q_k$, is violated in this setting, and that a hierarchical Bayesian model with age in the linear predictor captures what standardization misses. The best-fitting model takes $\\log(\\Theta_{ijk}) = \\alpha + \\phi_i + \\delta_j + \\gamma_k + \\zeta^1_{ij} + \\zeta^2_{ik} + \\zeta^3_{jk}$, with a Leroux conditional autoregressive spatial effect, an unstructured time effect, a first-order random walk age effect, and Type II space-time, Type I space-age, and Type IV time-age interactions. Its posterior shows a temporal trend that rises from 2017 through 2022 and then stabilizes, a nonlinear age profile with elevated contributions from the 15-55 age range and declines in the oldest groups, and a time-age interaction in which the youngest groups shift from below-average to above-average contribution around 2021 while middle and older age groups move the opposite way. The paper reads this as evidence that the age-specific post-2020 increase would have been invisible under indirect standardization.","pith_inferences":["The same modeling template could be applied to suicide mortality or to other rare outcomes such as substance abuse or gender-based violence, where age patterns plausibly vary by place and time; the authors gesture at these applications but do not test them.","A direct comparison of municipality rankings from standardized ratios versus posterior risk from the full model would quantify how often policy prioritization changes when the proportionality assumption is dropped.","The post-2020 youth increase is an observational pattern in emergency-call data; without adjusting for changes in help-seeking or reporting behavior, it should not be read as direct evidence about the pandemic's causal effect on suicide risk.","Repeating the analysis separately by sex or by call outcome, such as attempt versus ideation, would show whether the time-age pattern is driven by a particular subgroup or severity category."],"forward_implications":["If the proportionality assumption is violated, indirect standardization can misreport both the level and the direction of spatiotemporal risk trends, so age-disaggregated modeling should replace it in disease mapping.","The post-2020 increase in suicide-related emergency calls is not uniform across age: younger groups carry the increase, so prevention resources directed only at historically high-risk older groups would miss the emerging pattern.","Adding age effects improves model fit by thousands of WAIC units relative to models without age, meaning age is not a nuisance to standardize out but an informative dimension of risk.","The time-age interaction shows a reversal around 2020-2021, with younger groups increasing and older groups decreasing, a divergence that a summary standardized rate would smooth away.","The computational approach scales to large space-time-age datasets because INLA avoids expensive MCMC simulation."],"supporting_citations":[{"why":"Supplies the exploratory graphical test and the argument that indirect standardization assumes proportionality of stratum-specific risks across space and time.","marker":"Wakefield et al., 2017"},{"why":"Documents that violating the proportionality assumption can produce incorrect summarized relative risks, the central failure the paper targets.","marker":"Morris, 2017"},{"why":"Provides the Type I-IV interaction framework for space-time effects, which the paper extends to space-age and time-age interactions.","marker":"Knorr-Held (2000)"},{"why":"Supplies the conditional autoregressive prior used for the spatial random effect.","marker":"Leroux et al. (2000)"},{"why":"Introduces INLA, the approximate Bayesian inference method that makes fitting 520 models computationally feasible.","marker":"Rue et al. (2009)"},{"why":"Shows that identifiability constraints on interaction effects are needed, and the paper applies these constraints.","marker":"Goicoa et al. (2018)"},{"why":"Establishes WAIC, the criterion used to select the best model.","marker":"Watanabe (2010)"},{"why":"Provides prior spatiotemporal evidence on suicide-related emergency calls and age-gender patterns that motivates the case study.","marker":"Marco et al., 2024"}],"fun_headline_variants":["Standardization hides youth suicide rise; Bayesian age model reveals it","Age-standardized maps miss post-2020 youth suicide jump","Hidden youth suicide trend exposed when age isn't standardized away","Standardization masks stronger risk rise in younger suicide callers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that emergency calls related to suicide measure suicide risk the same way across age groups, regions, and years; if help-seeking or call behavior changed over time or differed by age, the estimated age-time patterns would describe reporting rather than risk.","fun_headline_variants_meta":{"raw":{"variants":["Standardization hides youth suicide rise; Bayesian age model reveals it","Age-standardized maps miss post-2020 youth suicide jump","Hidden youth suicide trend exposed when age isn't standardized away","Standardization masks stronger risk rise in younger suicide callers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000537,"raw_usage":{"total_tokens":2546,"prompt_tokens":883,"completion_tokens":1663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":1594}},"tokens_in":499,"tokens_out":1663,"duration_ms":14012,"temperature":1.0,"reasoning_tokens":1594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:55:19.872456+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the best model on suicide mortality records for the same region and years: if the post-2020 increase in younger age groups disappears or reverses while the model structure is unchanged, the paper's central pattern is an artifact of emergency-call reporting rather than a property of suicide risk.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes WAIC, the criterion used to select the best model."},{"cited_title":"and Gracia, E","cited_arxiv_id":null,"evidence_quote":"Provides prior spatiotemporal evidence on suicide-related emergency calls and age-gender patterns that motivates the case study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Type I-IV interaction framework for space-time effects, which the paper extends to space-age and time-age interactions."},{"cited_title":"G., Lei, X","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional autoregressive prior used for the spatial random effect."},{"cited_title":"and Chopin, N","cited_arxiv_id":null,"evidence_quote":"Introduces INLA, the approximate Bayesian inference method that makes fitting 520 models computationally feasible."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that identifiability constraints on interaction effects are needed, and the paper applies these constraints."}],"review_version":1}