Bayesian Geographically Weighted Regression (BGWR)
Bayesian Geographically Weighted Regression · Also known as: BGWR, Bayesian GWR, Bayesian spatially varying coefficient model, Bayesian local regression
Bayesian Geographically Weighted Regression combines the spatially varying coefficient framework of GWR with Bayesian inference, placing Gaussian process priors on the locally varying regression coefficients. This yields full posterior distributions over each coefficient at every location, providing principled uncertainty quantification rather than only point estimates.
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 BGWR when relationships between predictors and an outcome are expected to vary spatially and you need credible interval estimates around each local coefficient — for example in epidemiology (disease risk factors varying by region), ecology (species-environment relationships across a landscape), or real-estate economics (price determinants varying across a city). It is especially appropriate when sample sizes differ across sub-regions and uncertainty estimates should reflect local data sparsity. Do not use it when the relationship is genuinely stationary across space (classical OLS or spatial lag model will suffice), when the dataset is very large (n > 50,000) and computational cost is prohibitive without approximations, or when analysts lack the Bayesian modelling background to specify and diagnose priors and MCMC convergence.
Strengths & limitations
- Produces full posterior distributions for spatially varying coefficients, enabling principled credible intervals at every location.
- Naturally propagates uncertainty from sparse data regions into wider credible intervals rather than artificially narrow point estimates.
- Kernel hyperparameters (range, sill, nugget) are inferred from data rather than chosen ad hoc, reducing subjective bandwidth selection bias compared to classical GWR.
- Hierarchical GP priors provide spatial regularisation that prevents overfitting in small local samples.
- Can be extended to non-Gaussian likelihoods (logistic, Poisson) for binary or count spatial outcomes.
- Computationally demanding: full GP posterior scales as O(n^3) in memory and time, making it impractical for large datasets without approximations (e.g., sparse GP, NNGP).
- Requires specification of the GP covariance kernel family and prior distributions on hyperparameters; results can be sensitive to these choices.
- MCMC-based fitting requires convergence diagnostics and can take hours or days on moderately sized spatial datasets.
- Interpretation of spatially varying posteriors is more complex than classical regression output, requiring spatial visualisation and GIS expertise.
Frequently asked
How does BGWR differ from classical GWR?
Classical GWR estimates spatially varying coefficients by kernel-weighted local least squares, selecting bandwidth via cross-validation, and produces only point estimates. BGWR places GP priors on the coefficients and produces full posterior distributions, enabling credible intervals and formal inference about spatial variation. BGWR also jointly estimates the smoothness of coefficient surfaces from data rather than relying on a fixed bandwidth.
What is the biggest practical challenge of BGWR?
Computational cost. Full GP inference scales as O(n^3), so datasets with more than a few thousand observations require approximations such as the nearest-neighbour Gaussian process (NNGP) implemented in the R package spNNGP, or sparse GP methods in Stan or GPyTorch.
How do I choose the GP covariance kernel?
Inspect the empirical semivariogram of the residuals from a non-spatial model to gauge the plausible range and nugget. The Matern kernel with smoothness nu = 1.5 or 2.5 is a robust default. Conduct sensitivity analysis by comparing DIC or WAIC under different kernel choices.
Can BGWR handle binary or count outcomes?
Yes. By replacing the Gaussian likelihood with a logistic or Poisson likelihood and retaining the GP prior on spatially varying coefficients, you obtain Bayesian spatially varying coefficient logistic or Poisson regression. Fitting requires MCMC or Laplace approximation because the posterior is no longer Gaussian.
Which software implements BGWR?
The R package spBayes (Finley et al.) implements Bayesian spatially varying coefficient models efficiently. Stan and brms can fit custom BGWR models with GP terms. For large n, spNNGP (nearest-neighbour GP) provides scalable Bayesian spatial regression.
Sources
- Finley, A. O. (2011). Comparing spatially-varying coefficients models for analysis of ecological data with non-stationary and anisotropic residual dependence. Methods in Ecology and Evolution, 2(2), 143-154. DOI: 10.1111/j.2041-210X.2010.00060.x ↗
- Wheeler, D., & Calder, C. (2007). An assessment of coefficient accuracy in linear regression models with spatially varying coefficients. Journal of Geographical Systems, 9(2), 145-166. DOI: 10.1007/s10109-006-0040-y ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Geographically Weighted Regression. ScholarGate. https://scholargate.app/en/spatial-analysis/bayesian-geographically-weighted-regression
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 Spatial RegressionSpatial analysis↔ compare
- Geographically Weighted RegressionSpatial analysis↔ compare
- Local Spatial RegressionSpatial analysis↔ compare
- Multiscale Geographically Weighted RegressionSpatial analysis↔ compare
- Spatial Lag ModelSpatial analysis↔ compare