Modeling Spatial Heterogeneity in Exposure Buffers and Risk: A Hierarchical Bayesian Approach

1. Problem Statement

In place-based epidemiology, researchers often define "exposure" to spatial risk factors (e.g., healthcare facilities, pollution sources) using circular buffers with a fixed radius. This approach assumes that:

1. The exposure effect is constant beyond a specific distance threshold.
2. This threshold (buffer radius) is uniform across the entire spatial domain.

Limitations of Current Methods:

• Arbitrary Selection: Buffer radii are often chosen arbitrarily or based on precedent rather than data, leading to the "uncertain geographic context problem."
• Spatial Homogeneity Assumption: Traditional models assume the effective range of exposure is identical everywhere, ignoring local variations in geography, infrastructure, or behavior.
• Suboptimal Inference: Incorrect modeling choices (fixed vs. varying radii) can lead to biased estimates of exposure effects and poor statistical inference.
• Limitations of Existing Alternatives: Previous methods like Distributed Lag Models (DLMs) or two-stage functional regression often require pre-defined discrete distance intervals, rely on indirect inference for radii, or fail to account for spatial correlation and heterogeneity simultaneously.

2. Methodology: The SVBR Model

The authors propose SVBR (Spatially-Varying Buffer Radii), a flexible hierarchical Bayesian spatial change-point model.

Core Framework:
The model jointly estimates:

1. Location-specific buffer radii ($\delta(s_j)$): The distance at which exposure influence drops to zero for a specific location $s_j$.
2. Location-specific exposure effects ($\theta(s_j)$): The magnitude of the impact of exposure within that buffer.

Mathematical Formulation:

• Outcome Model: The outcome $Y_i(s_j)$ for individual $i$ at location $s_j$ follows a generalized linear model (Gaussian, Binomial, or Negative Binomial):
  $$g(\mu_i(s_j)) = O_i(s_j) + x_i(s_j)^T \beta + z(s_j; \delta(s_j)) \theta(s_j)$$
  Where $z(s_j; \delta(s_j))$ is the exposure metric (e.g., count of facilities within $\delta$) and $\theta(s_j)$ is the effect parameter.
• Exposure Definitions: The model supports two definitions:
  1. Counts: Binary presence of sources within the radius.
  2. Spherical: Continuous weighting of sources based on distance, decaying to zero at the radius $\delta$.
• Spatial Heterogeneity Modeling:
  • Radii ($\delta$): Modeled via a transformed regression with spatial random effects:
    $$\Phi^{-1}\left(\frac{\delta(s_j) - a}{b - a}\right) = w(s_j)^T \gamma + \phi(s_j)$$
    Where $\phi(s_j)$ follows a Gaussian Process (GP) with exponential covariance to capture spatial correlation.
  • Effects ($\theta$): Modeled as a linear function of covariates:
    $$\theta(s_j) = q(s_j)^T \eta$$
• Nested Models: The framework includes simpler models as limiting cases:
  • SVBR[$\delta, \theta$]: Single global radius and effect (traditional approach, but radius is estimated).
  • SVBR[$\delta(s_j), \theta$]: Spatially varying radii, constant effect.
  • SVBR[$\delta(s_j), \theta(s_j)$]: Fully spatially varying radii and effects.

Inference & Computation:

• MCMC: Parameters are estimated using Markov Chain Monte Carlo (MCMC).
• Priors: Weakly informative priors are used. Specific priors on $\gamma$ and spatial variance $\tau^2_\phi$ are designed to ensure identifiability and prevent overfitting by partitioning variability between fixed and random effects.
• Scalability: A predictive process approximation (using "knots") is implemented for large datasets ($m \gg K$) to reduce computational burden.
• Software: Implemented in the R package EpiBuffer.

3. Key Contributions

1. Data-Driven Buffer Estimation: Moves away from arbitrary fixed radii to a method that estimates optimal, location-specific radii directly from the data.
2. Spatial Heterogeneity: Explicitly models spatial variation in both the range of exposure (radii) and the magnitude of the effect, without requiring spatial replication.
3. Continuous Parameterization: Treats buffer radii as continuous variables rather than discrete intervals, eliminating arbitrary cut-points.
4. Unified Framework: Provides a single modeling framework that nests traditional fixed-buffer models and more complex spatially varying models, allowing for rigorous model comparison.
5. Open Source Tool: Development of the EpiBuffer R package to facilitate adoption.

4. Results

A. Simulation Study:

• Setup: Simulated data based on the Madagascar case study with five settings: No Effect, Single Radius/Effect, Varying Radii, Varying Radii & Effect, and Misspecified Data.
• Performance:
  • Correct Specification: The model matching the true data-generating process performed best.
  • Flexibility: Over-parameterized models (e.g., SVBR[$\delta(s_j), \theta(s_j)$]) adapted well to simpler structures without significant loss in performance.
  • Under-parameterization: Models assuming a single radius (SVBR[$\delta, \theta$]) failed significantly when true spatial variation existed (high bias, poor coverage).
  • Model Selection: Metrics like WAIC, PSIS-LOO, and LPML reliably identified the correct model structure across all settings.
• Robustness: The framework remained robust even under misspecified data generation (e.g., unmeasured confounding or non-Gaussian processes).

B. Application: Antenatal Care (ANC) in Madagascar:

• Context: Analyzed 1,818 women across 109 clusters in Toliara Province to determine the relationship between walking distance to healthcare facilities and completing $\ge$ 4 ANC visits.
• Model Comparison:
  • The SVBR[$\delta(s_j), \theta$] model (spatially varying radii, constant effect) provided the best fit (lowest WAIC/PSIS-LOO).
  • Fixed buffer models (5 km or 12.33 km) performed significantly worse.
• Key Findings:
  • Spatial Variation in Access: The effective buffer radius varied substantially across the region, ranging from 2.47 km to 17.08 km.
  • Urban vs. Rural: Areas with higher Global Human Footprint Index (more urban) had significantly smaller effective buffer radii, suggesting that in urban areas, proximity matters less (or facilities are denser), whereas in rural areas, women travel much further.
  • Effect Size: Each additional facility within an individual's specific buffer increased the odds of adequate ANC by 86% (OR = 1.86).
  • Policy Implication: A uniform 5 km access standard misclassifies accessibility in heterogeneous landscapes. Rural areas may require service planning for distances $>15$ km, while urban areas may be served by $<10$ km.

5. Significance and Implications

• Methodological Advancement: SVBR addresses a critical gap in spatial epidemiology by providing a principled, data-driven way to define exposure zones, moving beyond the "one-size-fits-all" buffer approach.
• Public Health Policy: The study demonstrates that uniform distance-based policies (e.g., "5 km access standard") are insufficient for heterogeneous environments. Resource allocation and facility siting should be context-specific, accounting for local spatial heterogeneity in access and utilization.
• Generalizability: While applied to healthcare, the framework is applicable to any spatially distributed risk factor (pollution, green space, crime) where the effective range of influence is unknown and likely varies by location.
• Reproducibility: The release of the EpiBuffer package ensures that these advanced spatial methods are accessible to the broader research community.