Bayesian Local Indicators of Spatial Association (Bayesian LISA)
Also known as: Bayesian LISA, Bayesian local spatial autocorrelation, Bayesian local Moran, B-LISA
Bayesian Local Indicators of Spatial Association extend the classical LISA framework by embedding local spatial association statistics within a Bayesian hierarchical model. Rather than relying on asymptotic permutation-based significance tests, this approach places prior distributions on spatial parameters and derives posterior probabilities that a location is part of a genuine spatial cluster, accounting for uncertainty and borrowing strength across nearby units.
Key highlights
- Controls multiple-testing inflation: posterior exceedance probabilities provide a coherent alternative to unadjusted permutation p-values across hundreds of locations.
- Quantifies uncertainty at each location via credible intervals rather than a single point estimate.
- Handles sparse count data naturally through the Poisson or negative-binomial likelihood in a hierarchical model.
- Allows prior information (e.g., population size, known risk factors) to be formally incorporated.
- Produces spatially smooth cluster maps that are less sensitive to single-unit outliers than classical LISA.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Bayesian LISA when you need locally identified spatial clusters and also want rigorous uncertainty quantification — for example in disease mapping with sparse counts, crime analysis across many small areas, or environmental monitoring where the classical permutation approach suffers from severe multiple-testing problems. It is particularly valuable when sample sizes per unit are small (count data with low event rates) and when you can incorporate informative priors from domain knowledge. Avoid it when computation time is a hard constraint (MCMC is expensive for large lattices), when a simple exploratory scan is sufficient, or when the audience expects a classical frequentist p-value map; in those cases classical LISA or the Getis-Ord Gi* hot-spot test are simpler and faster.
Strengths & limitations
- Controls multiple-testing inflation: posterior exceedance probabilities provide a coherent alternative to unadjusted permutation p-values across hundreds of locations.
- Quantifies uncertainty at each location via credible intervals rather than a single point estimate.
- Handles sparse count data naturally through the Poisson or negative-binomial likelihood in a hierarchical model.
- Allows prior information (e.g., population size, known risk factors) to be formally incorporated.
- Produces spatially smooth cluster maps that are less sensitive to single-unit outliers than classical LISA.
- Computationally intensive: MCMC sampling does not scale well to lattices with tens of thousands of units without approximations such as INLA.
- Sensitivity to prior choice: the spatial autocorrelation parameter's prior can influence posterior cluster maps, especially in small datasets.
- Harder to communicate than classical p-value maps; practitioners must explain posterior probabilities to non-statistical audiences.
- Requires specifying a full generative model (likelihood + prior), introducing additional modelling decisions compared to classical LISA.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does Bayesian LISA differ from classical LISA?
Classical LISA computes a local Moran statistic and tests it with a permutation distribution, yielding a p-value for each location. Bayesian LISA embeds the local statistic in a hierarchical model and derives posterior probabilities of cluster membership, providing uncertainty quantification and a natural correction for multiple comparisons without requiring an ad-hoc Bonferroni adjustment.
What software can I use to run Bayesian LISA?
The R packages R-INLA (via the INLA project) and CARBayes support Bayesian spatial models from which local cluster probabilities can be extracted. Stan and JAGS can be used for custom MCMC formulations. Classical LISA for comparison is available in the spdep and GeoDa packages.
How do I choose the threshold for declaring a location a cluster?
A common practice is to classify a location as a cluster if the posterior exceedance probability P(I_i > 0 | data) exceeds 0.80 or 0.95. The threshold is a substantive decision balancing sensitivity against false discovery; report the full probability map rather than only the binary classification.
Can Bayesian LISA handle continuous outcomes, or only counts?
Both. For count outcomes a Poisson or negative-binomial likelihood is used with a CAR prior on random effects. For continuous outcomes a Gaussian likelihood with a spatial CAR or SAR component is standard. The choice of likelihood is part of the model specification.
Is Bayesian LISA the same as Bayesian spatial regression?
No. Bayesian spatial regression models the mean of an outcome as a function of covariates plus a spatial random effect and focuses on coefficient estimation. Bayesian LISA focuses on identifying local clusters of spatial association through location-specific statistics derived from the posterior, though the two approaches share the same hierarchical modelling machinery.
Sources
- 1.Anselin, L. (1995). Local indicators of spatial association—LISA. Geographical Analysis, 27(2), 93–115.
- 2.Banerjee, S., Carlin, B. P., & Gelfand, A. E. (2004). Hierarchical Modeling and Analysis for Spatial Data. Chapman and Hall/CRC.ISBN 978-1584884101
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Local Indicators of Spatial Association. ScholarGate. https://scholargate.app/spatial-analysis/bayesian-local-indicators-of-spatial-association