{"id":"df29617b-fbdf-4e2f-ae6b-508808c82737","arxiv_id":"2411.14849","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"US ischemic heart disease mortality declined until 2014, flattened, then rose slightly after 2019, with clear regional and rural-urban disparities according to fast partitioned Bayesian models.","lead":"This paper applies fast Bayesian spatio-temporal models to infer ischemic heart disease mortality trends across US counties from 1999 to 2021, imputing suppressed small counts and splitting the country into state-based partitions to speed computation. The results point to a slowdown in the long decline after 2014 and a possible uptick after 2019, with rural counties in the West, Midwest, and South carrying higher risk than urban ones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Imputation of suppressed counts can create the reported post-2014 deceleration and post-2019 uptick, so the central change-point claim is insecure until the analysis is redone under a censoring-appropriate model.","rationale":"The reader's weakest assumption correctly identifies Section 2.2 imputation as the key vulnerability. My stress-test sharpens it: the issue is not merely lost uncertainty from single imputation, but a time-varying bias that can manufacture the exact shape of the reported national trend. As mortality falls, more county-years fall below the suppression threshold, so an imputation procedure that systematically over- or under-fills suppressed counts will distort later years more than earlier years. Since the main public-health message is the timing of the slowdown (after 2014) and the post-2019 uptick, this is load-bearing. A censored-likelihood or proper multiple-imputation reanalysis is a concrete, feasible check because the authors have released code and data, and the final model is fast enough to refit. The paper is otherwise honest about limitations, and the computational contribution (partitioned INLA models) is genuinely useful. I therefore do not move the verdict: CONDITIONAL remains appropriate until the imputation robustness check is performed.","tokens_in":15074,"tokens_out":4332,"duration_ms":46356,"concrete_test":"Using the public repository, refit the selected model (ICAR + RW1 + Type II interaction) on a dataset where suppressed counts are not imputed by Section 2.2 but treated as interval-censored, e.g. Y_it in {0,...,9} for suppressed cells, or multiply imputed from a model fitted with the censoring rule. Then recompute the national average risk and Figures 6-7, and report posterior distributions of average annual log-risk slopes for 2009-2014, 2014-2019, and 2019-2021. The imputation concern is neutralized only if the 2014-2019 slope remains significantly less steep than 2009-2014 (or if the credible intervals exclude a continued decline) under the censored analysis; if the flattening and uptick move into the null, the paper's headline claim is an artifact of the imputation procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — deceleration after 2014 and uptick after 2019 — is identified only after the Section 2.2 imputation, and that imputation can generate exactly such a pattern. Suppressed cells are not missing at random: a cell is missing exactly when the true count is ≤9, and the fraction of suppressed cells rises over time as mortality falls. Replacing these cells by round(e_it rhat_it), capped at 9, while treating the filled values as observed introduces a time-varying systematic error: posterior-median shrinkage tends to pull imputed values down, while spatial borrowing and the 9-cap bias them toward the upper end of the suppressed range. Because the final model (Eq. 2) conditions on these synthetic counts, the model-selection criteria, the national trend, and the regional/rural curves in Figures 6-7 all inherit this bias, and no imputation uncertainty reaches the reported credible intervals. The paper's own rural-observed-SMR sanity check (Section 3) is at a high aggregation level and does not target the national post-2014 slope; the Discussion concedes sensitivity to this method. Thus the headline epidemiological conclusion is not yet supported by a censoring-appropriate analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript analyzes county-level ischaemic heart disease mortality in the 48 contiguous US states over 1999-2021 using CDC WONDER data, in which counts below 10 are suppressed. Missing counts are imputed by fitting a year-specific BYM2 spatial model and replacing each censored count with the rounded posterior median predicted count truncated at 9. The imputed data are then modeled with a partition-based 'divide and conquer' spatio-temporal model (bigDM) with ICAR spatial prior, RW1 temporal prior, and Type II interaction. The main finding is that the US decline in IHD mortality slowed after 2014 and there was a slight increase after 2019, with regional and urban/rural disparities. The paper also reports model selection favoring the partitioned ICAR/Type II model over alternatives and provides reproducible code and data.","tokens_in":15315,"tokens_out":5399,"duration_ms":51937,"significance":"If correct, the paper offers a computationally practical pipeline for Bayesian spatio-temporal disease mapping on a large domain with censored counts, and its descriptive finding of a flattening IHD mortality trend is consistent with prior literature (Mehta et al.; Vaughan et al.). The full reproducibility (code and data on GitHub) and the comparison with global models are strengths. The main risk is that the central temporal claim is identified only after a single imputation of 12.5% of the observations, and the paper's own discussion acknowledges that the results 'may be sensitive to this method' and that the post-2019 rise cannot be separated from COVID-19 with only two years of data. Given the time-varying missingness, this is a load-bearing limitation rather than a presentation issue.","major_comments":[{"comment":"The imputation treats synthetic counts as observed, and no imputation uncertainty is propagated into the posterior credible intervals in Figures 5-7 or into the DIC/WAIC values in Table 2. Because suppression is deterministic (count < 10) and the fraction of suppressed cells increases over time as IHD mortality falls, the imputation error is time-varying and can distort the very slopes used to identify the post-2014 deceleration and post-2019 uptick. The rounded posterior medians truncated at 9 are not draws from the posterior predictive distribution, so the headline change-point claim could be an artifact of the imputation scheme. A censoring-appropriate reanalysis (e.g., multiple imputation, a censored likelihood as in Valeriano et al. 2021, or a sensitivity analysis excluding counties with any imputed counts) is needed before the central claim can be accepted.","section":"Section 2.2 and Eq. (2)"},{"comment":"The paper's sanity check for imputation effects, based on comparing model-based rural trends with observed-only SMRs, only checks aggregate levels; it does not target the post-2014 slope or the post-2019 increase. At the aggregation level used, a time-varying bias in imputed counts can cancel or be masked, so this check does not address whether the change-point pattern is an artifact. The check should be repeated for the slope estimates, for example by comparing the model's year-to-year trend with an observed-only SMR trend restricted to counties without any suppression.","section":"Section 3, Figure 7"},{"comment":"The claimed change-point dates are stated inconsistently. The abstract says 'slowdown in the decrease of IHD mortality in the US after 2014'; the text near Figure 6 says 'a slowdown in the decrease is observed after 2009 and a flattening in the trend is observed after 2014'; and the Discussion says 'our analysis suggests 2009 and 2019 as potential change points.' Because these dates are the central result, the paper needs to state a single, precisely defined claim and either justify the date or formally test for a change point rather than relying on 'mere visual inspection.'","section":"Abstract, Section 3, Discussion"}],"minor_comments":[{"comment":"The text 'from 1991 to 2021' should read 'from 1999 to 2021'.","section":"Section 3, Figure 4"},{"comment":"The phrase 'As suggested by the reviewers' is an artifact of the review process and should be removed or rephrased.","section":"Section 2.2"},{"comment":"The phrase 'differences exists' should be 'differences exist'.","section":"Abstract"},{"comment":"The expression 'the starting of the COVID-19 pandemic' should be 'the start of the COVID-19 pandemic', and the sentence 'From this year onwards, the average risks in rural areas is the highest' needs grammatical correction.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid computational contribution, but the central epidemiological claim is currently supported only by a single-imputation analysis. The authors have been transparent about the limitation, and a censoring-aware alternative is already sketched in the Discussion, so the required fix is within the scope of a revision. I see no evidence of circularity or missing-novelty issues; the main concern is that the headline result may not survive a sensitivity analysis. Given the journal's standards, major revision seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful, reproducible applied paper that demonstrates the bigDM divide-and-conquer workflow on 23 years of county-level IHD mortality, and it is honest about its limitations. The headline epidemiological claim—deceleration after 2014 and a post-2019 uptick—is not, however, supported by a censoring-appropriate analysis, and the imputation procedure could generate exactly that pattern.\n\nWhat is actually new: the specific county-level estimates for 1999-2021, including the urban/rural and regional breakdowns, and an imputation procedure for CDC WONDER suppressed counts (per-year spatial Poisson model, round, cap at 9). The modeling machinery itself is from the same group's earlier work (Orozco-Acosta et al.) and is not new. The paper does this cleanly: full code and data on GitHub, standard DIC/WAIC model selection, and a sensible computational strategy for 3,105 counties. The discussion is candid about sensitivity to imputation and about the COVID-19 confound.\n\nThe soft spot is load-bearing. Suppressed counts are not missing at random; they are suppressed when the true count is ≤9, and the fraction of suppressed cells rises over time as mortality declines. The authors replace each suppressed count with a rounded, truncated posterior-median prediction and then treat those synthetic values as observed in the final spatio-temporal model. No imputation uncertainty reaches the credible intervals. The per-year spatial imputation borrows strength from neighboring counties, which can pull estimates up or down, and the 9-cap induces a systematic upward bias for small counties. The direction of the net bias is not clear from the paper, but the concern is that a time-varying bias—rather than a true change in risk—could produce the reported deceleration and post-2019 bump. The aggregate observed-SMR sanity check in Section 3 does not isolate the national post-2014 slope, and change points are identified by eye, not formally. The post-2019 increase rests on two years that include the pandemic; the authors admit this.\n\nIf I were editing: sent to peer review. The computational demo is valuable and the authors are transparent, but a referee should require a reanalysis that treats counts as censored (e.g., the Valeriano et al. model or a multiple-imputation / sensitivity analysis) before the county-level details are used for policy. The descriptive maps and regional trends are probably robust, but the headline change-point claim needs the censoring-appropriate redo.","headline":"Solid, reproducible bigDM application; honest about limits, but the headline change-point claim rests on single imputation that could create the pattern.","tokens_in":15849,"tokens_out":4015,"would_cite":false,"duration_ms":33984,"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":"Using 23 years of county-level data for every contiguous US county, this paper claims the long decline in ischemic heart disease mortality slowed after 2009, flattened after 2014, and turned into a slight rise after 2019.","keywords":["ischemic heart disease mortality","spatio-temporal disease mapping","divide-and-conquer Bayesian models","county-level trends","urban-rural disparities","suppressed count imputation","mortality change points","United States counties"],"falsifier":"Fit the same model to the complete, un-suppressed death records for a subset of counties; if the post-2014 flattening or the post-2019 uptick disappears, the reported trajectory is an artifact of imputation.","tokens_in":14888,"feed_emoji":"📉","tokens_out":11682,"duration_ms":102510,"temperature":0.7,"pith_summary":"Using 23 years of county-level death counts for the contiguous United States, the paper asks whether the long national decline in ischemic heart disease mortality has continued evenly across places and populations. It answers no: the decline decelerated after 2009, flattened after 2014, and turned into a slight increase after 2019, with a 2020 peak that may reflect COVID-19 misclassification. The authors build this picture from Bayesian spatio-temporal models made fast by splitting the country into state-based partitions, and they impute the roughly 12.5% of county-year counts suppressed by confidentiality rules before smoothing. A reader should care because national averages hide the county and rural-urban disparities that determine where prevention efforts are needed, and because the post-2014 stall lines up with the stagnation of US life expectancy.","feed_headline":"Heart-disease deaths stopped falling in 2014, rose after 2019","feed_subtitle":"County-level maps of 23 years of US data also show rural counties and the Northeast still carry the highest risk.","key_machinery":"The central mechanism is a hierarchical Bayesian spatio-temporal disease mapping model. Death counts follow $Y_{it} \\sim \\text{Poisson}(e_{it} r_{it})$, and the log relative risk is decomposed as $\\log r_{it} = \\alpha_0 + \\xi_i + \\gamma_t + \\delta_{it}$, with $\\xi_i$ a spatially structured county effect, $\\gamma_t$ a first-order random walk over years, and $\\delta_{it}$ a Type II space-time interaction (structured in time, unstructured in space). To handle 3105 counties over 23 years, the paper replaces one global fit with a 'divide and conquer' scheme: partition the US into state-based subregions, add first-order neighboring counties across borders, fit the model separately in each partition, and allow different smoothing parameters per partition. Before fitting, suppressed counts (values below 10) are imputed with rounded, truncated predictions from a per-year Bayesian spatial model, so the smoothing machinery operates on a complete data grid.","core_discovery":"The paper's central claim is that the risk of dying from ischemic heart disease in US counties followed a four-phase national trajectory over 1999-2021: a clear decline to about 2009, a slowing decline from 2009 to 2014, a flat trend from 2014 to 2019, and an increase after 2019 with a peak in 2020. Beneath that national curve, no single trend describes the country. The West remains below the national risk, the Northeast above it; rural counties carry higher risk than urban counties in the West, Midwest, and South; and counties such as Valley (Montana) and Menominee (Michigan) show rising trends even while the national risk falls. The authors identify 2009 and 2019 as candidate change points while cautioning that the two post-2019 years cannot separate a persistent increase from a transient pandemic effect.","pith_inferences":["A consequence the authors leave implicit is that a real post-2014 flattening strengthens the case that cardiovascular mortality is the main brake on US life-expectancy gains, since ischemic heart disease is the leading cause of death.","The imputation step could be audited by fitting the same pipeline to complete, un-suppressed death records for a subset of states; if the flattening or the post-2019 uptick disappears, the finding would be an artifact of filling in small counts.","Extending the series through 2023 and later would settle whether the post-2019 increase is persistent or a one-off COVID-19 spike, something the paper's data ending in 2021 cannot resolve.","The same partitioned workflow transfers to other suppressed-count chronic diseases such as stroke or heart failure, where county-level screening could reuse the imputation and smoothing machinery."],"forward_implications":["If the trajectory is right, the decades-long US decline in heart disease mortality has levelled off, so public health planning should target a stalled epidemic rather than assume continuation of historical improvements.","The 2009 and 2019 change points give surveillance systems concrete moments to monitor, especially whether the post-2019 increase persists outside the pandemic years.","Rural counties in the West, Midwest, and South emerge as the highest-risk areas, reinforcing that national campaigns need county-level and rural-specific strategies.","The partitioned fitting strategy makes it practical to screen spatio-temporal disease patterns for all US counties in minutes on a desktop, not just for heart disease but for any condition with similar data."],"supporting_citations":[{"why":"Documents urban-rural differences in coronary heart disease mortality from 1999 to 2009; supplies the county population classification and the disparity baseline the paper extends.","marker":"Kulshreshtha et al. (2014)"},{"why":"Attributes the post-2010 stall in US life expectancy to cardiovascular disease; supports interpreting the flattening as a national health problem.","marker":"Mehta et al. (2020)"},{"why":"Earlier multivariate space-time Bayesian model of county heart disease mortality; the approach this paper accelerates and simplifies.","marker":"Quick et al. (2018)"},{"why":"County-level coronary heart disease trends by race, sex, and age group; the reference for county-specific increases that diverge from the national decline.","marker":"Vaughan et al. (2020)"},{"why":"Introduces the partitioned modeling approach and software used to fit the spatio-temporal models.","marker":"Orozco-Acosta et al. (2023)"},{"why":"Develops the scalable partition-based Bayesian smoothing and border-effect handling that the method relies on.","marker":"Orozco-Acosta et al. (2021)"},{"why":"Recommends Bayesian spatial models for highly censored public mortality data; the rationale for the imputation of suppressed counts.","marker":"Quick (2019)"},{"why":"Defines the four space-time interaction types, including the Type II interaction selected as the best model.","marker":"Knorr-Held (2000)"},{"why":"Provides the scaled spatial prior used in imputation and in comparing smoothing parameters across partitions.","marker":"Riebler et al. (2016)"}],"fun_headline_variants":["US heart disease death rate flattens after 2014, rises post-2019","Rural counties and Northeast still face highest heart disease risk","New fast method maps 23 years of US heart disease mortality trends","Heart disease death rate nationally flat since 2014, up after 2019","Bayesian divide-and-conquer reveals county-level heart disease shifts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the roughly 12.5% of county-year death counts suppressed because they are below 10 can be replaced by spatially predicted, rounded, and truncated values without systematically distorting the smoothed trends; the paper itself concedes the results may be sensitive to this imputation.","fun_headline_variants_meta":{"raw":{"variants":["US heart disease death rate flattens after 2014, rises post-2019","Rural counties and Northeast still face highest heart disease risk","New fast method maps 23 years of US heart disease mortality trends","Heart disease death rate nationally flat since 2014, up after 2019","Bayesian divide-and-conquer reveals county-level heart disease shifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000827,"raw_usage":{"total_tokens":3594,"prompt_tokens":905,"completion_tokens":2689,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":2603}},"tokens_in":521,"tokens_out":2689,"duration_ms":19764,"temperature":1.0,"reasoning_tokens":2603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:47:48.517537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the same model to the complete, un-suppressed death records for a subset of counties; if the post-2014 flattening or the post-2019 uptick disappears, the reported trajectory is an artifact of imputation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents urban-rural differences in coronary heart disease mortality from 1999 to 2009; supplies the county population classification and the disparity baseline the paper extends."},{"cited_title":"K., Abrams, L","cited_arxiv_id":null,"evidence_quote":"Attributes the post-2010 stall in US life expectancy to cardiovascular disease; supports interpreting the flattening as a national health problem."},{"cited_title":"S., Schieb, L., & Casper, M","cited_arxiv_id":null,"evidence_quote":"County-level coronary heart disease trends by race, sex, and age group; the reference for county-specific increases that diverge from the national decline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recommends Bayesian spatial models for highly censored public mortality data; the rationale for the imputation of suppressed counts."}],"review_version":1}