Bayesian Multiscale Geographically Weighted Regression
Also known as: Bayesian MGWR, B-MGWR, Bayesian multiscale GWR, Bayesian spatially varying coefficient model
Bayesian Multiscale Geographically Weighted Regression (Bayesian MGWR) extends the MGWR framework by placing Bayesian priors on each spatially varying coefficient. Each predictor is allowed its own bandwidth — its own geographic scale of influence — while Bayesian inference replaces classical back-fitting with posterior sampling, yielding full uncertainty quantification for every local coefficient surface.
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 MGWR when you have georeferenced data and suspect that the relationship between predictors and outcome varies across space at different geographic scales — for example, income's effect on health may operate at a neighbourhood scale while climate's effect operates at a regional scale. It is the right choice when uncertainty quantification for local coefficients matters (e.g., policy decisions, hypothesis testing about spatial non-stationarity). Do not use it when spatial non-stationarity is unlikely (global OLS or spatial lag models are simpler and more interpretable), when samples are very small (fewer than ~100 locations), when computation resources are limited (MCMC is expensive), or when a global average effect is the only quantity of interest.
Strengths & limitations
- Each predictor gets its own geographic bandwidth, revealing multi-scale spatial heterogeneity that single-bandwidth GWR cannot detect.
- Full posterior distributions provide credible intervals for every local coefficient, enabling principled uncertainty quantification.
- Bayesian priors regularise local estimates and prevent over-fitting in areas with sparse data.
- Posterior predictive checks and model comparison via WAIC or LOO-CV are straightforward within the Bayesian framework.
- Handles spatially structured residuals and non-stationary relationships simultaneously.
- Computationally intensive: MCMC sampling over high-dimensional coefficient surfaces is slow and may require parallel hardware.
- Prior specification for bandwidths and coefficient variances requires domain knowledge; misspecified priors can distort results.
- Interpretation of hundreds of local posterior summaries requires careful cartographic and statistical communication.
- Requires a sufficient density of georeferenced observations — sparse spatial coverage leads to poorly identified local estimates.
- Software support is less mature than for classical GWR; custom implementation is often needed.
Frequently asked
How is Bayesian MGWR different from standard GWR?
Standard GWR uses a single bandwidth for all predictors and produces point estimates. Bayesian MGWR allows each predictor its own bandwidth (revealing multi-scale heterogeneity) and returns full posterior distributions, so you get credible intervals rather than a single best-guess surface.
What sample size do I need for Bayesian MGWR?
As a practical minimum, at least 100-200 georeferenced observations are advisable. Fewer locations make the local coefficient surfaces poorly identified, and wide credible intervals will dominate results. With very sparse data consider a Bayesian spatial lag or error model instead.
Do I need to specify priors for the bandwidths?
Yes. A common choice is a weakly informative half-normal or uniform prior over the bandwidth parameter for each predictor, letting the data dominate. If you have strong prior knowledge about the likely geographic scale of a relationship, informative priors can regularise estimation in sparse regions.
How do I know if spatial non-stationarity is present at all?
Run a global test such as Leung et al.'s F-test for GWR or compare WAIC between a Bayesian MGWR and a global Bayesian regression. If the posterior bandwidths collapse toward infinity for all predictors, non-stationarity is minimal and a simpler model suffices.
Which software supports Bayesian MGWR?
Classical MGWR is available in the MGWR 2.2 Python package (slu-opendata.org). Bayesian extensions typically require custom MCMC implementations in Stan, PyMC, or R (e.g., via INLA with spatially structured effects); no off-the-shelf Bayesian MGWR package has achieved universal adoption as of 2026.
Sources
- Fotheringham, A. S., Yang, W., & Kang, W. (2017). Multiscale Geographically Weighted Regression (MGWR). Annals of the American Association of Geographers, 107(6), 1247-1265. DOI: 10.1080/24694452.2017.1352480 ↗
- Li, Z., Fotheringham, A. S., Li, W., & Oshan, T. (2020). Fast Geographically Weighted Regression (FastGWR): a scalable algorithm to investigate spatial process heterogeneity in millions of observations. International Journal of Geographical Information Science, 33(1), 155-175. DOI: 10.1080/13658816.2018.1521523 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Multiscale Geographically Weighted Regression. ScholarGate. https://scholargate.app/en/spatial-analysis/bayesian-multiscale-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 Geographically Weighted RegressionSpatial analysis↔ compare
- 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