Abstract
Discretization of a geographical region is quite common in spatial analysis. There have been few studies into the impact of different geographical scales on the outcome of spatial models for different spatial patterns. This study aims to investigate the impact of spatial scales and spatial smoothing on the outcomes of modelling spatial point-based data. Given a spatial point-based dataset (such as occurrence of a disease), we study the geographical variation of residual disease risk using regular grid cells. The individual disease risk is modelled using a logistic model with the inclusion of spatially unstructured and/or spatially structured random effects. Three spatial smoothness priors for the spatially structured component are employed in modelling, namely an intrinsic Gaussian Markov random field, a second-order random walk on a lattice, and a Gaussian field with Matérn correlation function. We investigate how changes in grid cell size affect model outcomes under different spatial structures and different smoothness priors for the spatial component. A realistic example (the Humberside data) is analyzed and a simulation study is described. Bayesian computation is carried out using an integrated nested Laplace approximation. The results suggest that the performance and predictive capacity of the spatial models improve as the grid cell size decreases for certain spatial structures. It also appears that different spatial smoothness priors should be applied for different patterns of point data.
Introduction
In spatial epidemiology and ecology, researchers frequently work with "point-based" data, such as the exact locations of disease occurrences. To protect privacy or simplify computational complexity, this data is often aggregated into administrative districts or regular grid cells. However, this discretization introduces significant challenges, most notably the Modifiable Areal Unit Problem (MAUP), where statistical results vary drastically based on the chosen boundaries or grid resolution. Furthermore, researchers must choose appropriate mathematical "priors" to smooth results across neighboring areas. This study investigates how the choice of grid resolution (spatial scale) and smoothing techniques influence the accuracy of Bayesian spatial models, demonstrating that the optimal configuration depends heavily on the underlying spatial pattern of the data.
Methods
The study utilizes a Bayesian hierarchical framework to model disease risk. To overcome the computational costs associated with traditional Markov Chain Monte Carlo (MCMC) methods, the researchers employ Integrated Nested Laplace Approximation (INLA), which is optimized for latent Gaussian models.
Four specific models were compared:
- IID: An unstructured model assuming no spatial correlation.
- IGMRF (Intrinsic Gaussian Markov Random Field): A standard "Besag" model smoothing based on immediate neighbors.
- RW2D (Second-order Random Walk): A flexible prior considering both first and second-order neighbors on a lattice.
- MATERN2D: A model using the Matérn correlation function to fine-tune the smoothness and range of spatial influence.
These models were tested across ten grid resolutions, from a coarse 5x5 grid to a fine 50x50 grid.
Datasets
The models were evaluated using both simulated and real-world data:
- Simulated Data (X1–X6): Six synthetic datasets mimicking scenarios like varying density (inhomogeneous Poisson processes), local repulsion, and local clustering.
- Humberside Data: A real-world dataset comprising 62 cases of childhood leukemia and lymphoma (and 141 controls) in North Humberside, England, characterized by sparse points and potential clustering.
Results and Discussion
The analysis revealed several key insights regarding model performance across different scales:
Scale Sensitivity and Spatial Structure
For most spatial structures, model performance—measured by lower Deviance Information Criterion (DIC) and Logarithmic Score (LS)—improved as grid cells became smaller, though with diminishing returns. In dense datasets (X5 and X6), the choice of grid scale was less critical, as the unstructured (IID) noise component was sufficient to capture the pattern.
Performance of Smoothing Priors
The RW2D model proved particularly effective for sparse, clustered data (e.g., Dataset X3). By considering second-order neighbors, it "borrows strength" from a wider area to fill gaps in sparse data. Conversely, the MATERN2D model showed high sensitivity to scale, performing best at intermediate resolutions (e.g., 20x20) but struggling at extreme coarse or fine scales.
Real-World Application: Humberside
In the Humberside case study, the MATERN2D model at a 30x30 grid resolution provided the best fit. This suggests that for sparse epidemiological data with potential clusters, a moderately fine grid combined with a flexible correlation function is optimal. While IID and IGMRF models were more stable across scales, they rarely provided the superior fit offered by specialized models like RW2D or Matérn.
Conclusion
This research underscores that grid size selection should not be a matter of convenience. The "signal" of a spatial phenomenon, such as a disease cluster, may only become apparent at specific resolutions. Researchers are encouraged to test multiple scales and smoothing priors, using metrics like DIC and LS to identify the combination that best captures the underlying spatial truth. Future work may extend these findings to irregular administrative boundaries and a broader continuum of spatial structures.
Summary updated