Bayesian Spatio-temporal Modeling of Multi-Site Wastewater Surveillance Data

arXiv:2611.07602 · stat.AP · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Bayesian Spatio-temporal Modeling of Multi-Site Wastewater Surveillance Data".

Jane: Here's the gist of the paper. Bayesian spatio-temporal modeling of multi-site wastewater surveillance data develops a hierarchical framework to decompose wastewater measurements into shared epidemic curves, regional deviations, persistent site effects, and within-site temporal dependence The resulting Bayesian framework has three main contributions First, a hierarchical decomposition places regional trajectories on a common statistical basis despite unequal site histories Second, a direct marginal spatial covariance distinguishes the magnitude of regional fluctuations from neighboring association Third, region-specific time shifts separate regional leads and lags from differences in signal level and local fluctuations.

Tom: Model Framework Description The baseline model is defined by the equation Yij (t) = g(t) + vi(t) + uij + eij (t), where g(t) captures the shared temporal component, v(t) represents time-varying deviations from the shared curve, uij are site-level random intercepts capturing persistent differences, and eij (t) is a mean-zero Gaussian residual process The shared temporal component g(t) is estimated using cubic B-spline basis functions to capture broad epidemic waves The site-level random intercepts uij capture persistent differences that are not explained by the shared curve or regional process The residual process eij (t) captures within-site temporal dependence and measurement variation after accounting for the other components.

Jane: Spatial Covariance Specification The spatial formulation for regional deviations specifies the covariance of v(t) directly using an areal covariance based on regional adjacency, where W = (wii′) is a symmetric adjacency matrix The first-order marginal autoregressive (MAR) model is specified as Σv = σ squared v (I + ρW∗), −1 < ρ < 1, allowing the magnitude of regional variation to be distinguished from the strength of local association This structure allows multiple sites within one region to share a regional deviation when observed at the same time and neighboring regions to share local spatial variation.

Tom: Regional Timing Extension Model 2 extends the baseline model by allowing that trajectory to occur earlier or later in each region within a selected epidemic period, yielding Yij (t) = g(t − δi) + vi(t) + uij + eij (t) The timing layer is introduced using a standard orthonormal sum-to-zero construction to treat all regions symmetrically and gives a zero-sum Gaussian distribution with Cov(δ σδ) = σ squared δHH⊤ This shift model draws on Bayesian curve registration to allow regional leads and lags to be estimated in weeks without attributing them to site levels or residual fluctuations.

Paper summary: Tom: So, looking at the full paper, the title "Bayesian Spatio-temporal Modeling of Multi-Site Wastewater Surveillance Data" points directly to this comprehensive approach where they break down measurements into shared epidemic curves, spatial regional deviations, persistent site effects, and within-site temporal dependence.

Jane: What I see is that the authors developed a hierarchical framework specifically designed to handle the reality of multi-site data, accommodating unequal observation windows while still making sense of the underlying epidemic signal.

Lu: The main contribution they make is having this hierarchical decomposition where regional trajectories share a common statistical basis even if those sites have different histories. They're showing you how to separate what’s shared from what’s unique to each location within the system.

Meng: From an operational standpoint, it’s about moving beyond simple curve fitting and instead understanding the distinct sources of variation that are driving the data, which is crucial for building reliable models at scale.

Lalam: The implication is that we can gain a clearer picture of disease spread by accounting for these multiple layers of complexity rather than just looking at a single aggregated signal.

Tom: The paper shows how to distinguish between regional leads and lags from differences in signal level and local fluctuations, which is really important when trying to understand the timing of an outbreak across different parts of a community.

Jane: It’s about using this framework to interpret regional surveillance through those distinct sources of variation instead of just relying on curve fitting alone.

Lu: The work supports that by showing how to separate those components, particularly the timing layers, and how omitting true shifts can lead to inflated temporal residual scales and ranges.

Meng: It seems like the paper provides a strong methodology for handling the complexity inherent in this type of surveillance data, even if deploying it at scale will require careful engineering decisions about priors and computational resources.

Lalam: Ultimately, this Bayesian spatio-temporal modeling framework is about providing a more nuanced statistical lens to view wastewater surveillance data.

Conclusion: Tom: So, this paper is about taking wastewater data from multiple sites across a whole region and using Bayesian methods to figure out what’s going on over time and space.

Jane: It takes all these different measurements from different places—like cities or regions—and it tries to untangle the signal into several parts: shared epidemic curves, local regional differences, things that happen at one specific site, and how the measurements change over time within that same site.

Lu: The authors build this hierarchical framework so that even if some sites have been measured less often than others or have different histories, they can still be statistically linked using a common structure.

Meng: From what I've seen in production, the big win here is separating the noise. It’s not just fitting one smooth curve; it’s figuring out exactly what that smooth curve represents versus what’s just random site-specific variation.

Lalam: The main point they make is how they handle those time shifts. They introduce a layer where they allow different regions to start their epidemic wave at different times, and they use Bayesian curve registration to estimate those timing differences directly in weeks without blaming it on the local site or just random noise.

Tom: That's interesting because when you look at the results, especially for that timing scale—the shift—it seems region-specific rather than a fixed thing about the location itself.

Jane: Exactly. The simulation validation showed that if you omit those true time shifts, your estimates of how fast the epidemic is moving through each region get way too big and noisy.

Lu: They found that for a first year of Omicron data, the timing differences were much more pronounced than in later waves, suggesting this model captures important dynamics during the initial spread phase.

Meng: If we're thinking about deploying something like this at scale, it means we need to be very careful how we define our priors for those regional effects because they are sensitive to how you structure the adjacency matrix between your sites.

Lalam: It confirms that surveillance isn't just about measuring a single number; it’s about using a structured model to understand the entire system of spread across different spaces and times.

Tom: And that leads into what this means for us: if we can accurately disentangle these factors, we stop guessing whether a rise in wastewater is local noise or part of a bigger regional pattern.

Department of Civil Engineering, University of Ottawa · School of Science, Harbin Institute of Technology, Shenzhen Campus of University Town of Shenzhen · Department of Statistics and Actuarial Science, Simon Fraser University · Department of Statistics and Actuarial Science, University of Waterloo

stat.AP

Submitted: 2026-11-05

Updated: 2026-11-05

Code: https://github.com/OntarioWastewaterSurveillanceConsortium/sars-cov-2-data

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Bayesian spatio-temporal modeling of multi-site wastewater surveillance data develops a hierarchical framework to decompose wastewater measurements into shared epidemic curves…

Key concepts

Hierarchical Decomposition
This is a statistical method used to break down complex data into simpler components. In this study, it separates wastewater measurements into shared patterns (the main epidemic curve), regional variations, persistent site differences, and within-site temporal changes. This allows researchers to understand the different sources of variation influencing the data.
Spatial Covariance Specification
This describes how the spatial relationship between different monitoring sites affects each other's measurements. The model uses an areal covariance based on regional adjacency to distinguish between fluctuations caused by a shared regional trend and those caused by local site-specific differences, clarifying the nature of spatial association.
Regional Timing Extension
This extension allows the model to estimate if an epidemic wave occurs earlier or later in different geographic regions. It uses Bayesian curve registration techniques to estimate these time shifts independently of other factors like signal level, helping to identify regional leads and lags.

Terminology

Summary

Bayesian spatio-temporal modeling of multi-site wastewater surveillance data develops a hierarchical framework to decompose wastewater measurements into shared epidemic curves, regional deviations, persistent site effects, and within-site temporal dependence The resulting Bayesian framework has three main contributions First, a hierarchical decomposition places regional trajectories on a common statistical basis despite unequal site histories Second, a direct marginal spatial covariance distinguishes the magnitude of regional fluctuations from neighboring association Third, region-specific time shifts separate regional leads and lags from differences in signal level and local fluctuations. p. 2

Model Framework Description

The baseline model is defined by the equation Yij (t) = g(t) + vi(t) + uij + eij (t), where g(t) captures the shared temporal component, v(t) represents time-varying deviations from the shared curve, uij are site-level random intercepts capturing persistent differences, and eij (t) is a mean-zero Gaussian residual process The shared temporal component g(t) is estimated using cubic B-spline basis functions to capture broad epidemic waves The site-level random intercepts uij capture persistent differences that are not explained by the shared curve or regional process The residual process eij (t) captures within-site temporal dependence and measurement variation after accounting for the other components.

Spatial Covariance Specification

The spatial formulation for regional deviations specifies the covariance of v(t) directly using an areal covariance based on regional adjacency, where W = (wii′) is a symmetric adjacency matrix The first-order marginal autoregressive (MAR) model is specified as Σv = σ squared v (I + ρW∗), −1 < ρ < 1, allowing the magnitude of regional variation to be distinguished from the strength of local association This structure allows multiple sites within one region to share a regional deviation when observed at the same time and neighboring regions to share local spatial variation.

Regional Timing Extension

Model 2 extends the baseline model by allowing that trajectory to occur earlier or later in each region within a selected epidemic period, yielding Yij (t) = g(t − δi) + vi(t) + uij + eij (t) The timing layer is introduced using a standard orthonormal sum-to-zero construction to treat all regions symmetrically and gives a zero-sum Gaussian distribution with Cov(δ σδ) = σ squared δHH⊤ This shift model draws on Bayesian curve registration to allow regional leads and lags to be estimated in weeks without attributing them to site levels or residual fluctuations.

Application and Validation

The application to Ontario SARS-CoV-2 wastewater surveillance data reveals persistent site heterogeneity, within-site temporal dependence, and spatial association among neighboring regions The analysis of two Omicron periods shows that regional timing differences are more pronounced during the first year of Omicron than in the XBB-and-later period Simulation studies demonstrate the recovery of these components and show that omitting true time shifts can lead to increased estimates of residual temporal variation The results support interpreting regional surveillance through distinct sources of variation rather than through curve fit alone.

Key Findings from Ontario Data

The first-year Omicron period showed a larger timing scale and several region-specific contrasts, indicating meaningful differences in when the shared trajectory occurred across PHUs The second period showed that the timing scale is smaller and individual contrasts are less pronounced Regional timing was more heterogeneous in the first period than in the second. The posterior mean timing scale σδ was 1. p. 9188 weeks in the first period, compared with 0. p. 9347 weeks and (0. p. 9028, 0.799) in the second The analysis shows that regional timing summaries are specific to the epidemic period rather than fixed characteristics of a PHU. p. 2

Simulation Validation Results

The simulations support estimation of the baseline decomposition under the Ontario sampling hierarchy and show how unmodeled timing can be absorbed by the residual process The simulation study further evaluate recovery of regional time shifts under the Ontario sampling structure and assess whether differences in wave height can be mistaken for timing differences Model 2 recovers the true timing scale of 2 weeks with an average estimate of 2.005, individual shift coverage of 90. p. 145%, and direction accuracy of 95.15%. Omitting true shifts inflates the temporal residual scale and range, while Model 2 recovers both parameters. p. 9 The models give similar estimates when shifts are absent

Future Directions

Further methodological work can clarify how curve specification and boundary handling affect regional timing summaries The framework could also support surveillance where local sampling is limited, as the shared curve pools information across the network while neighboring observations can inform contemporaneous regional deviations Incorporating covariates within the hierarchy could explain persistent site differences and allow linking wastewater trajectories with regional hospital admissions through a lagged association. The validation on subsequent waves could assess whether the resulting model anticipates local increases in hospital admissions. p. 2

Data Availability

The Ontario wastewater surveillance program and measurement methods are described by D’Aoust et al. (2024) Publicly released data are available in CSV format from Zenodo (https://doi. p. 3org/10.5281/zenodo.10810195) and the Ontario Wastewater Surveillance Consortium GitHub repository (https://github. p. 0com/OntarioWastewaterSurveillanceConsortium/sars-cov-2-data). p. 0

References

Rezgar Arabzadeh, Daniel Martin Gr¨unbacher, Heribert Insam, Norbert Kreuzinger, Rudolf Markt, and Wolfgang Rauch. Data filtering methods for SARS-CoV-2 wastewater surveillance H. p. 0 Asghar, O. M. Diop, G. Weldegebriel, F. Malik, S. Shetty, L. El Bassioni, A. O Akande, E Al Maamoun, S Zaidi, A J Adeniji, C C Burns Xiaotian Dai et al. p. 1, A Bayesian framework for modeling COVID-19 case numbers through longitudinal monitoring of SARS-CoV-2 RNA in wastewater Patrick M D’Aoust et al. p. 1, SARS-CoV-2 viral titer measurements in Ontario, Canada wastewaters throughout the COVID-19 pandemic Paul H. p. 1 C. Eilers and Brian D Marx, Flexible smoothing with B-splines and penalties Noel Cressie and Christopher K Wikle, Statistics for spatio-temporal data Xiaotian Dai et al. p. 1, A spatio-temporal framework for modelling wastewater concentration during the COVID-19 pandemic Kung-Yee Liang and Scott L. p. 1 Zeger, Longitudinal data analysis using generalized linear models Shaopei Ma et al. p. 6, Spatially correlated analysis of infectious disease outcomes based on Bayesian functional hierarchical models Gertjan Medema et al. p. 0, Presence of SARS-Coronavirus-2 RNA in sewage and correlation with reported COVID-19 prevalence in the early stage of the epidemic in the Netherlands Aiswarya Rani Pappu et al. p. 1, Tracking COVID-19 trends in communities with low population by wastewater-based surveillance Michael D. p. 1 Parkins et al. p. 1, Wastewater-based surveillance as a tool for public health action: SARS-CoV-2 and beyond Jordan Peccia et al. p. 1, Measurement of SARS-CoV-2 RNA in wastewater tracks community infection dynamics K. p. 1 Ken Peng et al. p. 0 p. 0, An exploration of the relationship between wastewater viral signals and COVID-19 hospitalizations in Ottawa, Canada K. p. 0 Ken Peng et al., Joint modeling of two stochastic processes, with application to learning hospitalization dynamics from wastewater viral concentrations Walter Randazzo et al., SARS-CoV-2 RNA in wastewater anticipated COVID-19 occurrence in a low prevalence area Mary E. p. 1 Schoen et al. p. 1, SARS-CoV-2 RNA wastewater settled solids surveillance frequency and impact on predicted COVID-19 incidence using a distributed lag model Emily Somerset and Patrick E. p. 1 Brown, Wastewater surveillance using differentiable Gaussian processes Stan Development Team, Stan Reference Manual Donatello Telesca and Lurdes Y. p. 1 T Inoue, Bayesian hierarchical curve registration Melanie M Wall, A close look at the spatial structure implied by the CAR and SAR models Ettore Zuccato et al. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Bayesian spatio-temporal modeling of multi-site wastewater surveillance data can be used to decompose wastewater measurements into a shared epidemic curve, spatially correlated regional deviations, persistent site effects, and within-site temporal dependence. This decomposition accommodates unequal observation windows and sampling schedules when comparing regional trajectories while preserving site-level heterogeneity.

  2. The model allows for the estimation of region-specific time shifts to distinguish regional leads and lags in epidemic timing from differences in signal levels and local fluctuations. This enables the AI system to compare epidemic timing across different regions and periods, rather than just signal levels.

  3. A direct marginal spatial covariance specification separates the magnitude of regional variation from the strength of association between neighboring regions, providing distinct interpretations of these two sources of spatial variation. This allows the system to quantify both how much regions depart from the shared trajectory and how those departures vary together.

  4. The framework enables regional comparisons to distinguish shared temporal change, regional fluctuations, and within-site temporal dependence by retaining site differences while separating them from the shared curve and regional process. This allows for a more nuanced interpretation of surveillance patterns by identifying distinct sources of variation across sites and regions.

  5. The AI system can incorporate covariates at the corresponding sampling level to explain heterogeneity through variables such as catchment population or sewer infrastructure, allowing persistent site differences to be explained directly rather than absorbed by site-level random intercepts.

Sources

Related papers