{"id":"7accf24a-54c4-49ba-9baa-f707bb8c2bfd","arxiv_id":"2411.09025","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A negative binomial soft BART model summarized with ALE is applied to Atlanta asthma ED data, yielding a harmful ozone association, a negative NO2 association, and temperature modification of both.","lead":"Using soft Bayesian additive regression trees and accumulated local effects, this paper estimates how mixtures of air pollutants and temperature relate to asthma emergency department visits in Atlanta. It finds that ozone is associated with higher risk, nitrogen dioxide with lower risk, and both associations are stronger on cooler days.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The applied headline (negative NO2, harmful O3, temperature modification) rests on the unverifiable assumption that Equation (3.1)'s confounders absorb all relevant spatial/temporal confounding; the simulation's independence setup cannot validate this, so a negative-control reanalysis is needed.","rationale":"The reader's weakest assumption — residual spatial/temporal confounding — is the correct load-bearing concern. The simulation is well executed and independently supports the methodological claim, so I do not see a basis for rejecting the paper. However, the application's headline associations are not protected by the simulation's clean DGP, where confounders and spatial effects are independent of exposures. A negative-control outcome analysis is a concrete, feasible check that would either reproduce the pattern under confounding or fail to do so, thereby testing whether the concern actually lands. Since the reader already conditioned the verdict on application-level robustness, my assessment does not change the verdict.","tokens_in":17379,"tokens_out":8821,"duration_ms":108688,"concrete_test":"Refit the Section 5 mixture model (same soft-BART specification, T=25, same priors, same confounders as in Equation (3.1), same ZIP-day aggregation) with the outcome replaced by unintentional-injury ED visits from the same Georgia Hospital Association data. Unintentional injury should share healthcare-seeking, neighborhood, and seasonal structure with asthma but should not share the hypothesized air-pollution mixture effects. If the injury ALEs reproduce the negative NO2 pattern or the O3-by-temperature pattern from Figures 4-5, the application headline is likely a confounding artifact; if the injury ALEs are null while the asthma ALEs retain the reported pattern, the concern is substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The methodological core — soft BART with a negative-binomial likelihood plus ALE summaries recovering nonlinear/interactive mixture surfaces for count outcomes — is supported by the simulation (Table 1, Figures 1, B2-B3), and I find no internal flaw in the MCMC updates in Appendix A. The load-bearing risk is in the application (Section 5). Equations (3.1) and the confounder list assume that the day-of-year spline, holiday indicator, ZIP-level poverty, and pCAR random intercept absorb all relevant spatial/temporal confounding, so that f(Z) captures exposure effects rather than place-based or seasonal structure. The simulation cannot test this: its confounders X are drawn independently of Z and of the spatial effects ν, so the DGP contains no unmeasured confounding. In the Atlanta data, plausible omitted factors — daily relative humidity, pollen, healthcare access, year-specific coding or access trends — are correlated with both exposure mixtures and asthma ED visits. The negative NO2 ALE and its temperature interaction could therefore reflect residual confounding or mutual adjustment between anti-correlated pollutants rather than a real conditional association. The paper's 'not causal' caveat (Section 6) does not remove this bias from the descriptive estimate, so the applied conclusion is conditional on a confounding assumption the simulation does not cover.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a negative binomial regression model with a soft BART component for the exposure-response surface, combined with accumulated local effects (ALE) summaries for interpreting marginal and joint exposure effects in the presence of correlated mixtures. The model is applied to warm-season asthma emergency department visits in Atlanta (2011-2018) with five daily exposures (PM2.5, NO2, O3, CO, maximum temperature), day-of-year splines, a holiday indicator, ZIP-level poverty, and a proper CAR random intercept. A simulation study using the Friedman surface among five active and five noise exposures shows that soft BART with a sparse branching process and T=25 trees recovers the surface with low bias, low RMSE, and credible interval coverage near 95%, and that ALE main and second-order effect estimates match the truth. In the application, the mixture model improves WAIC over single-exposure models, and the ALE summaries show a negative NO2 association, a harmful O3 association, and temperature modification of pollutant effects.","tokens_in":17632,"tokens_out":4093,"duration_ms":44326,"significance":"If the claims hold, the paper offers a practical alternative to BKMR for count outcomes with large sample sizes: soft BART can capture nonlinear and interacting exposure-response surfaces, and ALE provides interpretable summaries without extrapolating to unrealistic mixture profiles. The simulation is well designed, with a known generating function, a realistic correlation structure among active exposures, and evaluation of bias, coverage, and RMSE across ensemble sizes and prior choices. The code is publicly available. The methodological core is credible and useful. However, the headline applied findings about NO2 and O3 in Atlanta rest on strong confounding assumptions that the simulation does not test, and the paper's own 'not causal' caveat does not resolve that limitation for descriptive associations.","major_comments":[{"comment":"The applied conclusions—negative association between NO2 and asthma ED visits and stronger temperature modification of NO2 and O3 effects—are identified only under the assumption that the day-of-year spline, holiday indicator, ZIP-level poverty, and pCAR random intercept absorb all relevant temporal and spatial confounding. The simulation in Section 4 cannot support this assumption because its confounders X are drawn independently of exposures Z and spatial effects ν, so the data-generating process contains no unmeasured confounding. In the Atlanta data, plausible omitted factors such as humidity, pollen, healthcare access, and year-specific coding or access trends may be correlated with both the exposure mixture and asthma ED visits, which would bias the ALE estimates, including the negative NO2 association. The Section 6 statement that the analysis is 'not causal in nature' does not remove this bias from the descriptive estimates. Please add a negative-control reanalysis, adjust for additional confounders, or provide a sensitivity analysis bounding the impact of unmeasured confounding; at minimum, the abstract and discussion should temper the applied claims until such evidence is provided.","section":"Section 5, Eq. (3.1)"},{"comment":"No MCMC convergence diagnostics are reported for either the simulation or the application. The application uses 5,000 burn-in iterations and 1,000 posterior samples after thinning (Section 5), and Table 1 reports credible interval coverage, but without trace plots, R-hat statistics, or effective sample sizes the reader cannot assess whether the reported 95% credible intervals are reliable. This is load-bearing for the coverage claims in the simulation and for the uncertainty bands in Figures 4 and 5. Please report convergence diagnostics for the MCMC chains, or provide a justification for why the chosen run length and thinning are sufficient for the soft BART/NB sampler in Algorithm 1.","section":"Section 3.3 and Section 5"},{"comment":"The single-exposure and mixture ALE estimates differ substantially: O3 shifts from essentially null to harmful, and NO2 shows a strong negative association only in the mixture model. This pattern suggests sensitivity to mutual adjustment among correlated pollutants and to the model specification, rather than a robust conditional association. The paper's interpretation in Section 6 offers a plausible mechanism (NO2 as an ozone precursor with an inverse relationship to ozone on warmer days), but no analysis is provided to distinguish real mutual adjustment from artifacts of multicollinearity or from omitted confounding. A correlation matrix or variance-inflation summary for the four pollutants and temperature, and a check of the ALE results under different numbers of quantile intervals K, would substantially strengthen the applied claim. Without such evidence, the headline negative NO2 result should be presented as preliminary and model-dependent.","section":"Section 5, Figure 4 and Section 6"}],"minor_comments":[{"comment":"The caption contains a typo: 'Widley applicable information criterion' should be 'Widely applicable information criterion'.","section":"Figure 2 caption"},{"comment":"The notation θij ∼ Gamma(ξ, exp(ηij)) should explicitly state the shape/scale parameterization used, since Gamma distributions are sometimes parameterized by shape and rate; this would avoid ambiguity in the mean/variance formulas in Appendix A.1.","section":"Section 3.2"},{"comment":"The ALE estimation description could clarify that the lower and upper evaluation points (zlower_k, zupper_k) refer to the boundaries of each quantile interval, and should mention how the choice of K affects the finite-difference approximation and the smoothness of the resulting ALE curves.","section":"Section 3.4"},{"comment":"The paper positions soft BART as a more feasible alternative to BKMR for large datasets, but no runtime or performance comparison with BKMR is provided. A brief benchmark on a subsample, or even a discussion of the computational scaling, would make the feasibility claim more concrete.","section":"Section 4 and Section 6"}],"recommendation":"major_revision","confidential_remarks":"The methodological contribution is sound and the simulation is convincing, but the applied findings—especially the negative NO2 association—are the main headline of the abstract and are sensitive to confounding assumptions that the simulation does not cover. A negative-control or sensitivity analysis would substantially increase my confidence. The paper would also benefit from convergence diagnostics and a more cautious framing of the applied claims. This seems within scope for Biostatistics-style applied statistical journals, and the paper is likely acceptable after such revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a well-written applied methods paper that combines soft BART, negative binomial regression via Pólya-gamma augmentation, and ALE for summarizing mixture effects. The simulation study is the strong part: it shows that soft sparse BART recovers a Friedman-type surface and null noise exposures, with good coverage and low RMSE, across tree counts. The ALE summaries recover the truth. That's credible evidence the estimator works for count outcomes with correlated exposures. The code is on GitHub, and the MCMC details in the appendix are standard and appear correct.\n\nWhere it gets softer is the application. The headline claims — negative NO2 association, harmful O3 association, temperature modification — are conditional on a structural assumption: the day-of-year spline, holiday indicator, poverty, and pCAR intercept absorb all relevant spatial/temporal confounding. The simulation cannot validate this, because its confounders are drawn independently of Z and the spatial effects. The paper honestly says the analysis is 'not causal' (Section 6), but the descriptive estimates still inherit the confounding risk. Plausible omitted time-varying factors, like humidity, pollen, or year-specific health system changes, could shift those ALE curves. Also, the negative NO2 finding stands in contrast to prior literature; the mutual adjustment between correlated pollutants is a plausible explanation, but the paper does not test it.\n\nOther soft spots: no MCMC convergence diagnostics are reported, no sensitivity analyses (e.g., spline df, ALE K) for the application, and the data are restricted so the application results cannot be independently reproduced. The claim that soft BART is more feasible than BKMR at this sample size is plausible but not benchmarked.\n\nOverall: the methodological core is sound, the simulation is careful, and the ALE workflow is a useful contribution to environmental mixture analysis. The application findings are interesting but should be treated as hypothesis-generating. The paper deserves a serious referee, but the authors should be pushed to add convergence diagnostics, sensitivity analyses, and ideally a negative control or a more explicit treatment of unmeasured confounding. If I were handling it, I'd send it out.","headline":"Solid applied demonstration of soft BART + ALE for count mixtures, but the headline findings rest on untested confounding assumptions; the simulation alone cannot carry them.","tokens_in":18173,"tokens_out":1791,"would_cite":true,"duration_ms":18589,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62G08","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Soft BART inside a negative binomial model can recover nonlinear, interacting exposure-mixture surfaces for count outcomes, and ALE summaries of the Atlanta asthma data show ozone associated with more emergency-department visits and NO2…","keywords":["Bayesian additive regression trees","soft BART","negative binomial regression","environmental exposure mixtures","accumulated local effects","asthma emergency department visits","air pollution","Atlanta"],"falsifier":"Re-run the same 25-tree mixture model on a dataset where the exposure time series are shifted by several weeks relative to the health outcomes; under the null the ALE curves for all five exposures should be flat near zero, and a strongly non-flat NO2 curve would indicate residual confounding or a model artifact.","tokens_in":17164,"feed_emoji":"🌬️","tokens_out":5955,"duration_ms":52342,"temperature":0.7,"pith_summary":"The paper argues that soft Bayesian additive regression trees can be embedded in a negative binomial count model to estimate smooth, nonlinear, and interacting exposure-response surfaces for mixtures of air pollutants and temperature, with accumulated local effects used to summarize the fitted surface. If the approach works, it gives environmental epidemiologists a practical alternative to Gaussian-process methods for large count datasets. In the Atlanta warm-season data, the mixture model suggests ozone is associated with higher asthma emergency-department visit rates and nitrogen dioxide with lower rates, with both associations stronger on cooler days. The authors are careful to describe these as associations, not causal effects.","feed_headline":"Ozone and cool days tied to asthma ER visits in Atlanta","feed_subtitle":"A five-pollutant model also links NO2 to fewer visits, with both effects strongest on lower-temperature days.","key_machinery":"The load-bearing machinery is soft BART as a nonparametric component inside the linear predictor of a negative binomial model, together with Polya-gamma data augmentation that makes the tree-leaf updates conditionally conjugate. Soft trees replace hard decision rules with smooth gating functions, so the ensemble approximates smooth exposure-response surfaces without requiring many trees. A sparsity-inducing Dirichlet prior helps the ensemble ignore noise exposures. Accumulated local effects then summarize the fitted surface by integrating conditional partial derivatives over the observed exposure distribution, isolating the effect of one or two exposures while averaging over plausible values of the others. The combination of the soft-tree ensemble and the ALE summary is what allows nonlinear, interacting mixture effects to be extracted without extrapolation.","core_discovery":"The paper claims that soft BART, placed inside a Poisson-gamma negative binomial regression with a proper conditional autoregressive prior on ZIP-code random intercepts, recovers nonlinear and interactive mixture exposure-response surfaces for count outcomes, and that accumulated local effects then summarize those surfaces without extrapolating to implausible exposure combinations. In 200 simulation repetitions using a benchmark surface with five signal and five noise exposures, soft BART with a sparse branching-process prior showed low bias, near-nominal credible-interval coverage for the exposure function, and low root mean squared error, outperforming rigid-tree BART. Applied to 478,311 asthma emergency-department visits across 128 Atlanta ZIP codes over eight warm seasons, the 25-tree mixture model had much lower WAIC than single-exposure models, and its ALE summary shows a harmful association for ozone and a negative association for nitrogen dioxide, with both effects more pronounced on lower-temperature days.","pith_inferences":["If the negative NO2 association is real, it may reflect atmospheric chemistry in Atlanta, such as ozone scavenging by nitrogen oxides, rather than a protective biological effect of NO2; a natural test would be to stratify the ALE analysis by ozone concentration and see whether the NO2 association reverses.","A skeptical check would be a negative-control outcome analysis, such as injury emergency-department visits, with the same exposure mixture; flat ALE curves for the control outcome would increase confidence in the confounding adjustment.","The method could be extended to handle lagged exposure windows explicitly by feeding the BART component a structured distributed-lag form, addressing the limitation the paper acknowledges.","Because ALE is computed from the fitted ensemble, the same software template should transfer to other count outcomes and other fused exposure products, such as PM2.5 species or wildfire smoke."],"forward_implications":["The proposed negative binomial soft-BART sampler scales to roughly 220,000 daily count observations, a size at which Gaussian-process BKMR becomes computationally infeasible.","Simulation results indicate that ALE main-effect and pairwise plots recover the true functional forms of important exposures and correctly flatten for noise exposures across the central 95% of each exposure range.","WAIC comparisons in the application favor the five-exposure mixture model over any single-exposure model, implying that analyzing pollutants jointly changes the apparent effect of an individual pollutant such as ozone.","The Atlanta results imply that pollutant effects vary across the warm season, with the harmful ozone association and the negative NO2 association both stronger when maximum daily temperature is low.","A practitioner wanting to estimate mixture effects for count health data can use the provided software instead of BKMR, at the cost of roughly 20 hours of computing for a dataset of this size."],"supporting_citations":[{"why":"Supplies the soft BART ensemble and default priors that replace hard tree decisions with smooth gating functions, the core nonparametric component being proposed.","marker":"(Linero and Yang, 2018)"},{"why":"Defines accumulated local effects, the summary strategy used to interpret the fitted exposure-response surface without extrapolation.","marker":"(Apley and Zhu, 2020)"},{"why":"Introduces BKMR, the Gaussian-process mixture method that the paper positions soft BART as an alternative to.","marker":"(Bobb and others, 2015)"},{"why":"Provides the count-data BKMR framework whose modeling structure the paper adapts by substituting soft BART for the Gaussian process.","marker":"(Mutiso and others, 2024)"},{"why":"Defines the benchmark exposure-response surface used in the simulation study to assess recovery of nonlinearities and interactions.","marker":"(Friedman, 1991)"},{"why":"Introduces BART and the branching-process tree prior that soft BART extends.","marker":"(Chipman and others, 2010)"},{"why":"Supplies Polya-gamma data augmentation, which makes the tree and coefficient updates in the negative binomial model conjugate.","marker":"(Polson and others, 2013)"},{"why":"Provides the conjugate sampling routine used to update the overdispersion parameter in the negative binomial model.","marker":"(Zhou and others, 2012)"}],"fun_headline_variants":["Ozone and cool temps amplify asthma ER visits in Atlanta","NO2 tied to fewer asthma ER visits, especially on cool days","Cool days boost ozone's asthma risk in Atlanta","Asthma ER visits spike with ozone on cool days","Cool temps strengthen ozone asthma effect in Atlanta"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Atlanta findings assume that the seasonal and neighborhood adjustments in the model capture every outside factor that could drive both air pollution levels and asthma visits; if any such factor is missed, the estimated ozone and NO2 associations would be biased.","fun_headline_variants_meta":{"raw":{"variants":["Ozone and cool temps amplify asthma ER visits in Atlanta","NO2 tied to fewer asthma ER visits, especially on cool days","Cool days boost ozone's asthma risk in Atlanta","Asthma ER visits spike with ozone on cool days","Cool temps strengthen ozone asthma effect in Atlanta"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0004,"raw_usage":{"total_tokens":2087,"prompt_tokens":937,"completion_tokens":1150,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1083}},"tokens_in":553,"tokens_out":1150,"duration_ms":9489,"temperature":1.0,"reasoning_tokens":1083,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:09:19.971291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same 25-tree mixture model on a dataset where the exposure time series are shifted by several weeks relative to the health outcomes; under the null the ALE curves for all five exposures should be flat near zero, and a strongly non-flat NO2 curve would indicate residual confounding or a model artifact.","supporting_citations":[{"cited_title":"and Zhu, Jingyu","cited_arxiv_id":null,"evidence_quote":"Defines accumulated local effects, the summary strategy used to interpret the fitted exposure-response surface without extrapolation."},{"cited_title":"and Coull, Brent A","cited_arxiv_id":null,"evidence_quote":"Introduces BKMR, the Gaussian-process mixture method that the paper positions soft BART as an alternative to."},{"cited_title":"(2024, January)","cited_arxiv_id":null,"evidence_quote":"Provides the count-data BKMR framework whose modeling structure the paper adapts by substituting soft BART for the Gaussian process."},{"cited_title":"(1991, March)","cited_arxiv_id":null,"evidence_quote":"Defines the benchmark exposure-response surface used in the simulation study to assess recovery of nonlinearities and interactions."},{"cited_title":"and McCulloch, Robert E","cited_arxiv_id":null,"evidence_quote":"Introduces BART and the branching-process tree prior that soft BART extends."},{"cited_title":"and Windle, Jesse","cited_arxiv_id":null,"evidence_quote":"Supplies Polya-gamma data augmentation, which makes the tree and coefficient updates in the negative binomial model conjugate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the conjugate sampling routine used to update the overdispersion parameter in the negative binomial model."}],"review_version":1}