When air pollutants rise and fall together, a survival model can report that one of them, say sulfur dioxide, triples the risk of disease, while its correlated neighbours look harmless. That attribution may be an artifact of which variable happened to absorb the shared signal. A Cox proportional hazards model, the standard regression for time-to-event data with patients who drop out or are still event-free at study end, cannot tell these apart. It still reports one confident hazard ratio per pollutant. We propose to change what the model reports. It should give a credible set: the smallest group of exposures that, with 95% posterior probability, contains the one carrying a given effect. It should also flag which pollutant contrasts the data cannot identify at all. The tool we build is a single-effect Bayesian Cox model adapted to environmental mixtures. It handles exposures that change over follow-up, effects that live at the level of emission sources rather than individual molecules, and diffuse small effects that break the sparsity assumption borrowed from genetics.
The stakes are concrete. In the CHARLS cohort, Individual and joint exposure to air pollutants and patterns of multiple chronic conditions reports a hazard ratio of 3.175 (95% CI 2.291–4.401) per 10 μg/m³ of SO₂ for incident multimorbidity. The authors themselves concede that correlation among their 11 pollutants "complicates the assessment of their individual effects." In East China, Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations finds hazard ratios near 1.0 for six pollutants, with PM₂.₅ at HR 1.01 (p = 0.599). Regulation depends on which of these statements survives an honest treatment of collinearity.
Core Claim
The method comes from genetics. CoxPH-SuSiE extends the "sum of single effects" prior to Cox models. That prior writes the risk as a sum of a few components, each of which picks exactly one variable. The paper shows that this yields near-nominal 95% credible sets and calibrated inclusion probabilities when genetic variants are almost perfectly correlated. It was built and tested only on genotype correlation matrices, with fixed covariates and a truly sparse signal.
Environmental mixtures violate all three conditions. Exposures vary over follow-up. Pollutants share emission sources, so the natural causal unit may be "coal combustion" rather than SO₂. Many pollutants may contribute small effects at once.
Our claim is that a version of the single-effect Cox model with three changes will produce credible sets with close to nominal coverage on real pollutant and exposome correlation structures. The three changes are a counting-process likelihood for time-varying exposure, single effects that may select either a variable or a source group, and a dense ridge-like background term. By contrast, Cox-LASSO, elastic net, and stability selection will give confident attributions to the wrong member of a correlated cluster at a measurable rate. The novelty is not "Bayesian selection in Cox models," which already exists. It is making that machinery honest under the specific correlation and non-sparsity patterns of environmental data, and showing empirically when it is and is not.
Research Program
The model is
$$h(t \mid x(t)) = h_0(t)\exp\Big(\sum_{l=1}^{L} b_l^\top x(t) + u^\top x(t)\Big).$$
Here $x(t)$ is the vector of exposures at time $t$. Each $b_l$ has exactly one nonzero entry, drawn from a menu that includes every individual exposure plus source groups; a group's entry acts on the group's standardized mean. The background term $u \sim N(0,\tau^2 I)$ absorbs diffuse effects. Source groups are defined in two ways, which we compare. The first uses chemistry, for example combustion (SO₂, CO, PM) versus secondary inorganic aerosol (NO₃⁻, SO₄²⁻, NH₄⁺). The second uses hierarchical clustering of the exposure correlation matrix.
We fit the model by the iterative stepwise procedure of CoxPH-SuSiE, using its Laplace-approximated Bayes factors. The one change is that the partial likelihood is evaluated over at-risk sets built from start–stop intervals, so exposures can change within a person's follow-up.
Alongside the credible sets we report an identifiability diagnostic. We take the eigendecomposition of the observed information of the partial likelihood. Directions with near-zero eigenvalue are exposure contrasts, such as "SO₂ minus CO," that the cohort cannot estimate. We report which credible sets are wide because of such directions and which are wide because the signal is weak.
Evaluation has two parts. The first is semi-synthetic, which is the only way to know the truth. We take real exposure matrices and real censoring and follow-up patterns, then simulate event times from known causal exposures. The exposure sources are the CHARLS-linked 11-pollutant estimates and the 1,308-variable OneFlorida+ exposome from External Exposome Factors and Adverse Heart Failure Outcomes. Scenarios cover one causal pollutant, a causal source group, several small effects, and pure diffuse signal.
The second part is real-data reanalysis. We re-examine the CHARLS SO₂ finding and refit OneFlorida+ as continuous-time readmission-or-death rather than the 1-year binary endpoint the original study used.
Following Stop Chasing the C-index, predictive evaluation reports three things. These are Integrated Brier Score and log-likelihood for accuracy, D-calibration for calibration, and C-index only as a secondary discrimination measure, all computed with the SurvivalEVAL library that paper identifies. The primary metrics for attribution are credible-set coverage, set size, and calibration of inclusion probabilities.
Baselines
The naive reference is the analysis in the literature: a multi-pollutant time-varying Cox model with Wald intervals, paired with quantile g-computation weights, as in the CHARLS study. Its reported benchmark is the SO₂ HR of 3.175. We rerun it on identical data. It fails if its 95% intervals cover the true simulated coefficient far below 95% or if it assigns large weight to non-causal correlated pollutants.
The strongest penalized competitor is Cox elastic net, together with SCAD-penalized Cox with iterative screening as in High-dimensional variable selection for Cox's proportional hazards model. That method recruits "marginally weak but jointly vital" features and reports roughly 6 minutes against over 10 hours for LASSO on 10,707 probes. We wrap both in stability selection, which refits on 100 subsamples, so they output selection frequencies directly comparable to inclusion probabilities. All methods get the same exposure matrix, the same cross-validation folds for penalty tuning, and matched wall-clock budgets. Our method beats them if its sets reach within a few points of 95% coverage while the stability-selection "top variable" misattributes a clearly larger fraction of causal signals at comparable set size.
Unmodified CoxPH-SuSiE is the closest existing method. It is also the core ablation: it has fixed baseline exposures, variable-only effects, and no background term. The paper reports coverage "close to the target 95%" on genomic simulations. We found no reported result on Lacuna for any environmental or time-varying exposure setting. We add each ingredient one at a time. The proposal is justified only if at least one ingredient measurably restores coverage in the source-group or diffuse-effect scenarios. If plain CoxPH-SuSiE already achieves nominal coverage there, the contribution shrinks to an application study, and we will say so.
DeepSurv (DeepSurv; C-index 0.867 on METABRIC vs. 0.816 for linear Cox) serves only as a predictive ceiling on Integrated Brier Score. It does not produce attributions.
Why not just…?
Why not just use Cox elastic net or ridge, which are stable under collinearity? They are stable because they spread weight across a correlated cluster. That gives a good predictor but no statement about which member matters, and no calibrated uncertainty. This is a real threat for prediction, and we expect elastic net to match our Integrated Brier Score. Stability selection is the fair way to extract attribution from it. The semi-synthetic coverage comparison settles whether its selection frequencies are as calibrated as posterior inclusion probabilities. If they are, the simpler tool wins and we report that.
Why not just estimate the mixture's total effect with quantile g-computation? If the policy question is "does this mixture harm health," that answer suffices and we do not compete with it. But the CHARLS paper uses qgcomp weights to name SO₂ as the top contributor. That weighting still rests on an unstable multi-pollutant fit and assumes largely additive, same-direction contributions. Our simulations measure how often those weights name the wrong pollutant.
Why not just collapse exposures into principal components or a deprivation index? Collapsing is legitimate when the data truly cannot separate the pieces. The identifiability diagnostic tells us exactly when that holds. Source-group single effects are essentially a data-driven version of this aggregation that is applied only where needed. Methodological challenges in spatial and contextual exposome-health studies shows that blanket aggregation changes significance conclusions and removes the pollutant-level answer that regulation requires.
Why not just use BKMR or BART? The same review reports that they do not scale to large cohorts. Standard implementations also do not target censored, time-varying proportional hazards. We include them only on the CHARLS subsample, where they are feasible.
Decisive Checks
The first test is the semi-synthetic coverage study on real CHARLS and OneFlorida+ correlation structures. It measures credible-set coverage, set size, and misattribution rate across the four signal scenarios, comparing the full model, its ablations, and all baselines.
The second is the CHARLS reanalysis. The key question is whether the SO₂ effect resolves to a singleton set or to a combustion-source set containing CO and PM. The diagnostic also tells us whether SO₂-versus-CO is identifiable in 1,938 people at all. A wide, identifiability-flagged set would be a substantive finding: the 3.175 attribution is not supported by the data.
The third is OneFlorida+ as continuous-time survival. We check whether the May maximum-temperature signal (aOR 1.04, 95% CI 1.02–1.06) persists as a credible set or dissolves into a broader climate cluster. We also check whether calibration metrics distinguish the methods where C-index does not.
References
- Individual and joint exposure to air pollutants and patterns of multiple chronic conditions
- Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations from 2015 to 2018 in East China
- BAYESIAN VARIABLE SELECTION IN A COX PROPORTIONAL HAZARDS MODEL WITH THE "SUM OF SINGLE EFFECTS" PRIOR.
- External Exposome Factors and Adverse Heart Failure Outcomes in the OneFlorida+ Network: Retrospective Cohort Study
- Stop Chasing the C-index: This Is How We Should Evaluate Our Survival Models
- High-dimensional variable selection for Cox's proportional hazards model
- DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network
- Methodological challenges in spatial and contextual exposome-health studies