Bayesian Methods for Modeling Cumulative Exposure to Extensive Environmental Health Hazards

Rob Trangucci, Jesse D. Contreras, Jon Zelner, Joseph N. S. Eisenberg, Yang ChenarXiv 2024 · DOI · arXiv

TL;DR

Researchers have developed a new Bayesian generative model to measure disease risk from "extensive" environmental hazards like rivers or canal networks. By combining log-Gaussian Cox processes with distance-based kernels, the model captures how hazard intensity varies along a spatial source, offering a more accurate alternative to traditional "shortest distance" metrics.

What Problem Does This Paper Address?

In environmental epidemiology, we are quite good at modeling "point sources"—think of a single factory chimney or a localized chemical spill. In these cases, exposure is usually simplified to a function of the distance from that one point.

However, many real-world hazards are extensive: they are lines or shapes in space, such as wastewater canals, rivers, or busy highways. Traditional methods often fall back on using the "shortest distance" to the source as a proxy for exposure. This approach has two major flaws:

  1. It assumes the hazard intensity is identical at every point along the source.
  2. It ignores the cumulative effect of being near multiple parts of a hazardous network.

If a child lives near a bend in a wastewater canal, they are exposed to pathogens from the entire visible stretch, not just the single closest point. Ignoring this leads to biased risk estimates and a failure to account for the true uncertainty of exposure.

Methods

The authors propose a Bayesian generative semiparametric model. Instead of a simple distance calculation, they model the hazard as a "doubly-stochastic" process:

  • Log-Gaussian Cox Process (LGCP): They use an LGCP to model the latent (unobserved) intensity of the hazard along the spatial extent of the source. This allows the "danger level" to vary smoothly along a canal or river.
  • Distance Kernels: This intensity is integrated against a kernel (like a Gaussian or exponential decay function) that represents how pathogens or pollutants travel through space to reach a household.
  • Generative Dose-Response: The model is built from the ground up, starting with a Poisson process for pathogen emission, a binomial process for how many pathogens reach an individual, and finally a Bernoulli "one-hit" model for whether an infection occurs.
  • Computational Strategy: Because the true model involves complex integrals, the authors use a discretization scheme. They prove mathematically that this approximation is "well-defined" (the posterior exists) and that the error is controllable. They implemented the model in CmdStan, leveraging parallel computing to handle the heavy lifting of Gaussian Process (GP) calculations.

Datasets

The paper evaluates the model using two primary sources:

  1. Simulation Study: A synthetic 10km x 4km region featuring a three-segment canal system. They tested the model against different household distributions ("uniform" vs. "clustered") and varying sample sizes (500 to 2,000 households).
  2. Mezquital Valley Survey: Real-world longitudinal data (2017–2019) from Mexico. This dataset tracks diarrheal illness in children living near a 4,400km network of wastewater canals used for irrigation. It includes GPS coordinates, household wealth, and parental education.

Key Insights and Ideas

  • Geometry and Distribution Matter: One of the most interesting findings was that the way people are settled affects what the model learns. If households are clustered near the hazard, the model is great at mapping the hazard's intensity. If they are spread out uniformly, the model is better at estimating the "background" risk of disease unrelated to the hazard.
  • The "Shortest Distance" Bias: The authors demonstrate that traditional models assuming a power-law decay (based on the single closest point) can significantly misrepresent how quickly risk drops off as you move away from a source.
  • Mathematical Rigor for Inverse Problems: The paper isn't just an application; it provides a formal proof that the discretized Bayesian inverse model converges to the true continuous model, giving researchers confidence in the numerical approximations.

Key Results

  • Superior Risk Characterization: In the Mezquital Valley application, the new model revealed a much sharper decline in risk than the traditional "shortest distance" model. It estimated that exposure effectively hits zero beyond 200 meters, whereas the simpler model suggested a much slower, less realistic decline.
  • Discretization Sensitivity: The simulation showed that grid resolution is critical. Using a coarse grid (e.g., 40 points along a segment) led to poor "coverage"—meaning the model's confidence intervals didn't actually contain the true values. Moving to a finer grid (160 points) corrected this, highlighting a clear trade-off between computational speed and inferential accuracy.
  • Parameter Recovery: The model successfully recovered the "length-scale" ($\rho$) of the hazard—essentially how far the "stink" or danger travels—even when the intensity of the source was unknown and varying.

Limitations and Future Work

  • Computational Scaling: Gaussian Processes are notoriously "heavy." As the complexity of the canal network or the number of households grows, the current implementation may struggle to scale without more advanced GP approximations (like Vecchia or inducing points).
  • Identifiability: The authors note it is difficult to perfectly separate the intensity of the source from the scale of the distance decay ($\rho$). In some scenarios, a very intense source with a short reach looks like a weak source with a long reach.
  • Static Households: The current model treats households as fixed points. The authors suggest that future versions could incorporate human mobility data—acknowledging that people move around the hazard throughout the day, accumulating exposure in a more dynamic way.

Summary updated

Abstract

Measuring the impact of an environmental point source exposure on the risk of disease, like cancer or childhood asthma, is well-developed. Modeling how an environmental health hazard that is extensive in space, like a wastewater canal, impacts disease risk is not. We propose a novel Bayesian generative semiparametric model for characterizing the cumulative spatial exposure to an environmental health hazard that is not well-represented by a single point in space. The model couples a dose-response model with a log-Gaussian Cox process integrated against a distance kernel with an unknown length-scale. We show that this model is a well-defined Bayesian inverse model, namely that the posterior exists under a Gaussian process prior for the log-intensity of exposure, and that a simple integral approximation adequately controls the computational error. We quantify the finite-sample properties and the computational tractability of the discretization scheme in a simulation study. Finally, we apply the model to survey data on household risk of childhood diarrheal illness from exposure to a system of wastewater canals in Mezquital Valley, Mexico.
BibTeX
@misc{trangucci2024bayesian,
  author = {Rob Trangucci and Jesse D. Contreras and Jon Zelner and Joseph N. S. Eisenberg and Yang Chen},
  title = {{Bayesian Methods for Modeling Cumulative Exposure to Extensive Environmental Health Hazards}},
  year = {2024},
  eprint = {2404.04398},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.04398},
}
Figure