Bayesian Spatial Autocorrelation
Bayesian Spatial Autocorrelation Analysis · Also known as: Bayesian spatial dependence, Bayesian LISA, Bayesian spatial clustering, BSA
Bayesian Spatial Autocorrelation embeds spatial dependence directly into a Bayesian hierarchical model. A Conditional Autoregressive (CAR) prior encodes the expectation that neighboring areas are more similar than distant ones, and posterior inference is obtained via MCMC. This approach is especially valuable in disease mapping, ecology, and regional science, where small-area estimates need borrowing strength across neighbors.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian Spatial Autocorrelation when you have areal (polygon) data with spatially referenced counts or rates and you need both smoothed small-area estimates and uncertainty quantification. It is especially appropriate in disease mapping, crime analysis, ecological risk assessment, and regional economics when sample sizes per area are small and classical tests would be unreliable. It is also preferred when you want to decompose spatial variation into structured (neighboring) and unstructured (overdispersion) components. Avoid it when data lack a meaningful adjacency structure, when you have no prior knowledge to inform hyperparameter choices, or when computational resources are very limited — classical Moran's I is sufficient for a simple global clustering test.
Strengths & limitations
- Fully probabilistic: produces credible intervals and exceedance probabilities for each area, not just a point estimate.
- Borrowing strength: small-area estimates are stabilised by pooling information from neighbors via the CAR prior.
- Decomposes spatial variation into structured clustering and unstructured overdispersion (BYM model).
- Naturally handles missing data and irregular geographies through the prior structure.
- INLA implementations (R-INLA) make computation feasible even for large lattices.
- Computationally demanding with MCMC for very large grids; INLA mitigates this but adds approximation error.
- Results can be sensitive to the choice of CAR prior hyperparameters and adjacency definition.
- Requires correctly specified adjacency matrix W; errors in neighborhood structure propagate into posterior inference.
- More complex to interpret and communicate than frequentist Moran's I, especially to non-statistical audiences.
Frequently asked
How does Bayesian Spatial Autocorrelation differ from Moran's I?
Moran's I is a single descriptive statistic that tests for global spatial clustering with a p-value. Bayesian Spatial Autocorrelation is a full hierarchical model that estimates smoothed area-level effects, decomposes spatial and non-spatial variation, and yields posterior distributions rather than a single number. They answer different questions: 'Is there clustering?' versus 'What is the locally varying risk with uncertainty?'
What is the BYM model?
BYM stands for Besag-York-Mollie (1991). It specifies the log-risk in each area as the sum of a spatially structured CAR component (φ) and an independent unstructured heterogeneity component (θ). This decomposition lets the model capture both global clustering and local overdispersion that cannot be explained by spatial proximity alone.
Can I use INLA instead of MCMC?
Yes. Integrated Nested Laplace Approximation (INLA) provides fast, accurate posterior approximations for latent Gaussian models, including CAR-based spatial models. The R-INLA package implements this and is much faster than full MCMC for most disease-mapping and areal-data applications, making it the preferred tool in practice.
How do I define the neighborhood matrix W?
The most common approach is queen or rook contiguity: areas sharing a boundary (or a corner, for queen) are coded as neighbors (W_ij = 1). For point data snapped to a lattice, k-nearest-neighbor or distance-band adjacency may be more appropriate. The choice of W directly shapes the estimated spatial autocorrelation structure and should reflect the substantive process being modeled.
What software implements Bayesian Spatial Autocorrelation?
R-INLA (via the INLA package) is the most widely used tool for areal Bayesian spatial models. Alternatives include Stan (via brms or custom Stan code) for MCMC-based inference, and the CARBayes package in R for BYM and related models. Python users can use PyMC with custom CAR log-probability functions.
Sources
- Besag, J., York, J., & Mollie, A. (1991). Bayesian image restoration, with two applications in spatial statistics. Annals of the Institute of Statistical Mathematics, 43(1), 1–20. DOI: 10.1007/BF00116466 ↗
- Gelfand, A. E., Diggle, P., Guttorp, P., & Fuentes, M. (Eds.). (2010). Handbook of Spatial Statistics. CRC Press. ISBN: 978-1420072877
How to cite this page
ScholarGate. (2026, June 3). Bayesian Spatial Autocorrelation Analysis. ScholarGate. https://scholargate.app/en/spatial-analysis/bayesian-spatial-autocorrelation
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian KrigingSpatial analysis↔ compare
- Bayesian Spatial RegressionSpatial analysis↔ compare
- Local Indicators of Spatial AssociationSpatial analysis↔ compare
- Local Spatial AutocorrelationSpatial analysis↔ compare
- Moran's ISpatial analysis↔ compare
- Spatial AutocorrelationSpatial analysis↔ compare