{"id":"e66ab382-8f85-4144-922d-9c5d096ecbda","arxiv_id":"2412.02970","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Bayesian spatial functional concurrent regression model estimates a 5-11 day lead time from wastewater SARS-CoV-2 levels to clinical positivity rates, with a time-varying association that weakens during outbreaks.","lead":"This paper builds a Bayesian statistical model that links weekly SARS-CoV-2 levels in Houston wastewater to daily COVID-19 positivity rates across 28 sewer service areas. The model estimates that wastewater spikes lead reported infections by roughly 5 to 11 days, and that the strength of the link changes between outbreak and non-outbreak periods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-case-count missingness rule in §1.1 censors the response exactly in the non-outbreak regime where γ(τ) is claimed to be strongest; because the model treats missingness as ignorable and never tests it, the dynamic-association result may be an artifact.","rationale":"The reader's weakest assumption identifies the same load-bearing concern. I agree with that diagnosis. The most load-bearing claim is the dynamic association between wastewater levels and positivity rates, and the inverse γ–outbreak pattern could be a selection artifact because the missingness rule removes exactly the low-prevalence periods where γ is claimed to be largest. The lag estimate may be less directly affected because it is identified primarily during outbreak periods, but γ and Δ are estimated jointly, so a biased γ could also distort the lag posterior. The Discussion explicitly acknowledges that reported positivity rates become unreliable as testing declines and suggests incorporating test counts, which supports the concern that the missingness mechanism matters. Additional weaknesses—no simulation study, no baseline comparison—are secondary. The paper's methodology is credible and the lag estimate is consistent with prior work, so I do not call for rejection; the missingness issue is addressable and should be resolved before the dynamic-association claim is accepted. Hence I keep the conditional verdict unchanged.","tokens_in":921,"tokens_out":798,"duration_ms":63290,"concrete_test":"Generate synthetic data from the fitted model (or a simplified version with known Δ and γ), apply the actual missingness rule: simulate p_i(τ) ~ Binomial(n_i(τ), y_i(τ)) using observed test counts, drop observations with p_i(τ) < 5, and refit the model. If the posterior of γ(τ) in low-count periods and the 5–11 day HPD for Δ deviate substantially from the truth, the missingness mechanism is informative. A complementary check is to re-estimate with a joint model for positive counts that includes the threshold mechanism, e.g., p_i(τ) ~ Binomial(n_i(τ), q_i(τ)) with q linked to X_i(τ−Δ), and compare the posterior of γ and Δ; the central claims should be robust to both analyses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 1.1, days with fewer than five positive cases in a given WWTP area are considered missing. Since the response is positivity rate y_i(τ) = p_i(τ) / n_i(τ), this is a deterministic, outcome-dependent censoring rule: low-prevalence periods are systematically removed, and the missingness also depends on test volume n_i(τ), which is not in the model. The observation equation (7) assumes independent errors and ignores this mechanism, so the missingness is non-ignorable. The central dynamic finding (Figure 2) is that γ(τ) is large when wastewater levels are low—precisely the non-outbreak periods in which y is mostly unobserved; at those times γ(τ) is identified only through temporal smoothness and spatial borrowing, not through direct response data. The 2023 tail, which is almost entirely imputed, is similarly unsupported. No simulation, sensitivity analysis, or posterior predictive check is provided to show that the <5-case rule does not bias γ(τ) or Δ. I am not claiming the result is wrong; I am claiming that the paper has not established that the flagship inverse γ–outbreak association survives the data-processing rule that defines the outcome.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian spatial functional concurrent regression model to characterize the time-varying association between SARS-CoV-2 wastewater concentrations and reported clinical positivity rates across 28 Houston wastewater treatment plant service areas. The model includes a shared time-lag parameter, a measurement-error model for the wastewater concentrations, functional factor models with multiplicative gamma process priors, and a spatially correlated CAR prior for the functional random effects. The application yields a posterior lag distribution concentrated between 5 and 11 days and an estimated regression coefficient function that is inversely related to the level of wastewater concentration, with stronger associations during non-outbreak periods. The authors argue this provides an early-warning tool with a quantified lead time and shows that the wastewater-positivity relationship is dynamic over the pandemic.","tokens_in":19165,"tokens_out":3451,"duration_ms":37382,"significance":"If the central estimates are reliable, the paper makes a useful contribution to wastewater-based epidemiology by providing a unified Bayesian framework that propagates uncertainty from sparse functional data, measurement error, and an unknown lag. The novelty is moderate: the key modeling components (Bayesian functional factor models, CAR priors, and ultra-efficient MCMC) are adapted from prior work by the same group, but the combination with an unknown lag in a spatial functional concurrent regression is a sensible extension. The paper ships a full Gibbs sampler and demonstrates its application on a real public health dataset. However, the flagship claims—the 5–11 day lead time and the dynamic inverse association—depend on a data-processing rule that censors the response, and the manuscript provides no simulation or sensitivity analysis to establish that the results are not an artifact of that rule.","major_comments":[{"comment":"The missingness rule 'Days with less than five positive cases in a given WWTP area are considered as missing' is a deterministic, outcome-dependent censoring mechanism on the response variable y_i(τ) = p_i(τ)/n_i(τ). Because the model in Eq. (7) assumes independent observation errors and does not model the testing volume or the censoring threshold, the missingness is non-ignorable. Low-count periods, which are precisely the non-outbreak regime where the paper claims the strongest association, are systematically unobserved; the posterior of γ(τ) and Δ is then identified only through temporal smoothing and spatial borrowing rather than through direct response data. The manuscript does not provide any simulation, sensitivity analysis, or posterior predictive check demonstrating that this rule does not bias γ(τ) or the lag estimate. I am not claiming the result is false; I am claiming the central inverse-association finding has not been shown to survive the missingness mechanism that defines the outcome.","section":"Section 1.1, Eq. (7)"},{"comment":"The paper's flagship dynamic-relationship claim is that γ(τ) is large precisely when wastewater levels are low, i.e., during non-outbreak periods. This is also the regime in which positivity rates are most likely to be censored by the <5-case rule. The fitted positivity curves during those periods in Figure 4 are largely imputed from the smoothness prior and the spatial CAR component, so the inverse relationship in Figure 2 may be a consequence of where data are observed rather than a feature of the underlying association. The authors should perform a sensitivity analysis with alternative thresholds (e.g., censoring at <1 or <10 cases), or model the test-count process, to show that the inverse pattern is stable to the censoring rule.","section":"Section 3, Figure 2"},{"comment":"The model assumes a single shared lag parameter Δ across all locations and all time periods, yet the abstract's lead-time claim is stated in terms of 'spikes' in wastewater levels versus positivity rates. If the temporal association γ(τ) is in fact dynamic, and if reporting delays or testing practices vary across WWTPs and over the pandemic, a constant lag may be an average that does not characterize any particular spike. At minimum, the paper should test sensitivity to allowing site-specific lags or a time-varying lag, or should justify the common-lag assumption with a model comparison. Without such a check, the 5–11 day HPD interval cannot be interpreted as the lead time for outbreak spikes.","section":"Section 2.1.1, Figure 6"},{"comment":"The paper presents a new inferential method but includes no simulation study, no out-of-sample validation, and no posterior predictive checks. The text claims (after Figure 5) that the model 'is able to provide estimates to the sparsely observed regions through the wastewater measurements by borrowing information from nearby service areas,' yet for the 2023 tail, where positivity data 'halting completely,' the imputed positivity curves are entirely unverifiable. A simulation study under realistic missingness patterns, or a holdout prediction exercise (e.g., hiding the last observed positivity values and comparing predictions), is necessary to support both the methodological claims and the applied conclusions about the dynamic association.","section":"Section 3, Section 4"}],"minor_comments":[{"comment":"If Δ takes values 0,...,21, that is 22 possible values, so the uniform prior should be Categorical(1/22), not Categorical(1/21) as written in the supplement; the text also says 'equal probability for each possible lag,' which conflicts with the stated probability.","section":"Section 2.1.1 and Supplement Eq. (9)"},{"comment":"There are typographical errors: 'postivity' should be 'positivity,' 'drat' should be 'draft,' and 'devleopment' should be 'development.'","section":"Section 4 and Author Attribution"},{"comment":"Equation (7) contains a double comma after the covariance matrix: 'N_{m_i^y}(0, σ^2_{\\epsilon_y} I_{m_i^y}),,' which should be a single comma.","section":"Eq. (7)"},{"comment":"The caption says '95% pointwise credible intervals (dashed lines),' but the text in Section 3 refers to 'shaded regions' for Figure 4 and 5; the presentation is inconsistent and should be clarified.","section":"Figure 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is probably better framed for an applied statistics or environmental epidemiology audience than as a methods-first stat.ME paper, since the core building blocks (functional factor models, CAR priors, lag sampling) are adapted from prior work, and the novelty is in the combination and the application. The non-ignorable missingness issue is the main technical risk; the revision should include a sensitivity analysis or a simulation study, not just discussion. The authors should also make code and data available, as reproducibility is a strength that would increase confidence in the results. The paper's lead-time estimate is broadly consistent with the existing wastewater literature, which is reassuring, but the dynamic inverse-association claim is not yet supported by the evidence presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stress-test concern lands. The paper is a legitimate methodological extension: it jointly estimates an unknown discrete lag, spatial CAR random effects, and measurement error in a Bayesian functional concurrent regression. That combination is new in the wastewater literature, and the MCMC appears carefully built. The posterior for the lead time (5–11 days HPD, mode 8–9) is plausible and consistent with earlier cross-correlation work. The authors also do a good job illustrating the sparsity and heterogeneity of the data.\n\nBut the central empirical claim about a time-varying association is not established. The rule in Section 1.1—days with fewer than five positive cases are missing—is outcome-dependent censoring. Positivity rate is positives over tests, so the rule removes exactly the low-prevalence regime where the model estimates the largest gamma. The observation equation treats the missingness as ignorable, with no sensitivity analysis, simulation, or posterior predictive check. The inverse gamma–outbreak pattern in Figure 2 may therefore be an artifact of imputation rather than a real feature. The 2023 tail is almost entirely imputed, and the paper offers no validation for those curves. The discussion does concede that testing declined, but that admission does not address the censoring mechanism.\n\nThe novelty claim also self-contradicts: the paper says no FDA methods have been applied in COVID-19 WBE literature, then immediately cites Dai et al. 2022 and 2024 as exactly that. Those papers pre-estimate the lag, so the current combination is still new, but the claim as written is overstated. A minor point: the linear concurrent model with a single shared lag may force the dynamic coefficient to absorb misspecification; that is worth checking but is not itself a flaw.\n\nWho gets value from this? Statisticians interested in functional data with structured missingness, and wastewater epidemiologists who want a principled posterior for lead time. The paper deserves a serious referee: the method is coherent, the data are real, and the missingness issue is addressable. I would send it to peer review with a request for a missing-not-at-random sensitivity analysis, a simulation study, and some out-of-sample forecasting check. I would not cite it for the dynamic finding as it stands, but I might cite the model.","headline":"Useful model with plausible lag inference, but the flagship dynamic-association result is not yet supported because the missingness rule censors the low-count regime where that association is claimed.","tokens_in":19737,"tokens_out":3015,"would_cite":false,"duration_ms":31040,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62M30","62F15","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Wastewater SARS-CoV-2 levels in Houston lead reported COVID-19 positivity rates by about a week, and the wastewater–positivity link is stronger between outbreaks than during them.","keywords":["wastewater-based epidemiology","functional concurrent regression","SARS-CoV-2","sparse functional data","Bayesian factor models","time lag inference","spatial dependence","COVID-19 positivity rate"],"falsifier":"Re-run the analysis on a data window in which low-case-count days were still reported, or on fully observed simulated data with a known lag and the same missingness mask. If the posterior for the lag Δ and the coefficient γ(τ) moves materially, the missingness rule rather than the underlying biology is producing the headline lead-time and state-dependent findings.","tokens_in":18692,"feed_emoji":"🦠","tokens_out":5103,"duration_ms":45081,"temperature":0.7,"pith_summary":"The paper tries to establish that the relationship between SARS-CoV-2 wastewater concentrations and reported COVID-19 positivity rates is a dynamic, time-varying link rather than a fixed correlation, and that the wastewater signal leads the clinical signal by a quantifiable interval. Using data from 28 wastewater treatment plant service areas in a major city, the authors estimate that a spike in wastewater viral load precedes a spike in reported positivity by roughly 5 to 11 days, with the most probable lag around 8 or 9 days. They also find that the strength of the association fluctuates: it is larger during non-outbreak periods and smaller, though more precisely estimated, during outbreak waves. If correct, these results give public health officials a probabilistic early-warning lead time and show that wastewater surveillance remains informative even when clinical testing data become sparse.","feed_headline":"Wastewater signals COVID-19 waves 5 to 11 days early","feed_subtitle":"Bayesian model of Houston sewersheds also shows the wastewater–positivity link is strongest between outbreaks.","key_machinery":"The central object is a Bayesian functional concurrent regression model, yi(τ)=Xi(τ−Δ)γ(τ)+θi(τ)+ε, in which the predictor curve is the true wastewater curve Xi, the response is the positivity rate curve yi, Δ is a categorical lag parameter over 0–21 days, γ(τ) is a smooth time-varying coefficient, and θi are spatially correlated functional random effects. The true wastewater curve itself is unobserved and reconstructed through functional factor models with a multiplicative gamma process prior, so uncertainty in the weekly measurements is propagated. The machinery that carries the argument is the joint posterior over Δ, γ(·), and the latent curves: it converts the raw, sparse series into smooth curves, shares information across the 28 sewersheds via a conditionally autoregressive prior on extended Hausdorff-distance neighbors, and lets the lag be inferred rather than fixed.","core_discovery":"The paper claims that the relationship between SARS-CoV-2 wastewater concentration and reported COVID-19 positivity rate is a time-varying functional coefficient, not a constant link, and that the wastewater signal leads the clinical signal by a lag that can be learned from data. The estimated lag distribution places 95% posterior mass between 5 and 11 days, with mode near 8–9 days. The estimated coefficient function γ(τ) moves inversely with the population wastewater level: in quiet periods an incremental rise in wastewater has a larger association with positivity, while during outbreak waves the coefficient is smaller yet estimated more precisely. The authors argue that this dynamic structure, with full uncertainty propagation through measurement error, missingness, and spatial dependence, is needed to avoid the biases of pre-smoothed or lag-fixed correlation analyses.","pith_inferences":["The same modeling template should transfer to other pathogens and outcomes such as influenza or ICU admissions, where the lead-time question is equally urgent but testing data are even sparser.","The dynamic coefficient γ(τ) implies that alert thresholds for wastewater should be state-dependent: a small rise during a quiet period is more informative than the same rise during an active wave.","A direct check of the missingness assumption would be to re-estimate on a period with near-complete testing, or to simulate under an informative missingness mechanism; until then the non-outbreak association estimates carry an unquantified risk."],"forward_implications":["A spike in wastewater concentration can be treated as an early warning 5 to 11 days before reported positivity rises, with the most probable lead around 8 to 9 days.","The association between wastewater and positivity is not a fixed multiplier; it is stronger during non-outbreak periods and weaker, but more precisely estimated, during waves.","The method yields a full posterior distribution for the lag rather than a single cross-correlation number, so decision makers can use probabilistic statements such as the 95% HPD interval.","Spatial borrowing of information across sewersheds lets the model estimate positivity curves even where public testing faded to zero."],"supporting_citations":[{"why":"Establishes the precursor link between wastewater SARS-CoV-2 RNA and community infection dynamics that this model formalizes.","marker":"Peccia et al. 2020a"},{"why":"Provides city-wide Houston wastewater versus surveillance-indicator correlations that the model extends with a full inferential framework.","marker":"Hopkins et al. 2023"},{"why":"Systematic review and meta-analysis of wastewater–case correlations used to benchmark the lead-time estimates.","marker":"Li et al. 2023b"},{"why":"Documents the wide 0 to 14 day range of previous lead-time estimates that motivates explicit lag inference.","marker":"Olesen et al. 2021"},{"why":"Supplies the sparse-functional-data framework that motivates the Bayesian factor approach to curve estimation.","marker":"Yao et al. 2005"},{"why":"Previous functional concurrent regression with variable selection that the proposed model builds on.","marker":"Goldsmith and Schwartz 2017"},{"why":"Introduces the extended Hausdorff distance used to define spatial neighbors between irregular sewershed areas.","marker":"Min et al. 2007"},{"why":"Review noting that few wastewater models account for measurement error, which the proposed model addresses.","marker":"Torabi et al. 2023"}],"fun_headline_variants":["Sewage predicts COVID waves up to 11 days ahead","COVID's sewage signal shifts strength between waves","Bayesian sewage model pinpoints COVID's lag window","Wastewater foretells COVID outbreaks with dynamic lag","5–11 day head start from wastewater to COVID cases"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The data rule that treats any day with fewer than five positive cases as missing must be innocuous; those days are concentrated exactly in the non-outbreak periods where the model finds the strongest wastewater–positivity association, so if missingness tracks the true case count the estimated lag and the dynamic coefficient could be distorted.","fun_headline_variants_meta":{"raw":{"variants":["Sewage predicts COVID waves up to 11 days ahead","COVID's sewage signal shifts strength between waves","Bayesian sewage model pinpoints COVID's lag window","Wastewater foretells COVID outbreaks with dynamic lag","5–11 day head start from wastewater to COVID cases"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1308,"prompt_tokens":892,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":338}},"tokens_in":508,"tokens_out":416,"duration_ms":4235,"temperature":1.0,"reasoning_tokens":338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:54:21.847541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the analysis on a data window in which low-case-count days were still reported, or on fully observed simulated data with a known lag and the same missingness mask. If the posterior for the lag Δ and the coefficient γ(τ) moves materially, the missingness rule rather than the underlying biology is producing the headline lead-time and state-dependent findings.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the wide 0 to 14 day range of previous lead-time estimates that motivates explicit lag inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Previous functional concurrent regression with variable selection that the proposed model builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the extended Hausdorff distance used to define spatial neighbors between irregular sewershed areas."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Review noting that few wastewater models account for measurement error, which the proposed model addresses."}],"review_version":1}