Spatial Bayesian Model Averaging
Also known as: spatial BMA, BMA for spatial data, Bayesian model averaging with spatial effects, spatial model uncertainty averaging
Spatial Bayesian model averaging (spatial BMA) extends classical BMA to settings where observations are georeferenced and spatial dependence must be modelled. Rather than selecting a single spatial regression model — which spatial weight matrix to use, which regressors to include, which spatial lag or error structure to adopt — it averages the predictions and parameter estimates across all candidate models, weighting each by its posterior probability given the data.
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 spatial BMA when (1) you face genuine uncertainty about the correct spatial model specification — lag vs error vs Durbin structures — or about which regressors to include; (2) the data are georeferenced and spatial autocorrelation is plausible; (3) you need robust coefficient or spillover estimates that do not over-commit to a single model. Do not use spatial BMA when you have a strong theoretical justification for a specific spatial model with a fixed variable set — in that case, fit the specified model directly and save computational cost. Also avoid when the candidate model space is poorly defined or when spatial weights are entirely unknown and no sensitivity analysis is feasible.
Strengths & limitations
- Propagates model uncertainty into inference, producing parameter estimates and predictions that are averaged over competing spatial structures and variable sets.
- Posterior inclusion probabilities provide a principled variable-selection diagnostic free of arbitrary p-value thresholds.
- Accounts for direct and indirect (spillover) effects uncertainty simultaneously across models.
- Flexible: can incorporate multiple spatial weight matrices as model dimensions, explicitly testing spatial connectivity assumptions.
- Reduces risk of over-fitting a single, possibly misspecified spatial model to the data.
- Computationally demanding: the model space grows exponentially with the number of candidate regressors and spatial structures; MCMC model composition is required for large spaces.
- Requires a spatial weights matrix W even as a candidate; misspecification of all candidate W matrices can bias all averaged estimates.
- Marginal likelihoods for spatial models are not always in closed form, requiring numerical approximations whose accuracy affects posterior weights.
- Interpretation of averaged coefficients in spatial models is non-trivial because direct and indirect effects must be decomposed for each model before averaging.
Frequently asked
What is a posterior inclusion probability (PIP) and how should I interpret it?
A PIP for a variable is the sum of posterior weights of all models that include that variable. A PIP above 0.5 means the data favours including the variable more than excluding it. Values near 1 indicate a robust predictor across spatial model structures; values near 0.5 signal genuine model uncertainty about whether the variable matters.
How does spatial BMA differ from just fitting a spatial lag model with LASSO or stepwise selection?
LASSO and stepwise selection pick one model and discard the rest, so all downstream inference conditions on that single model as if it were known to be correct. Spatial BMA retains all models and weights them, so uncertainty about the spatial structure and variable set is carried through to every estimate and prediction.
Which spatial weight matrices should I include as candidates?
Include matrices that reflect plausible connectivity theories for your substantive domain: contiguity (queen/rook), inverse distance with different bandwidth parameters, and k-nearest-neighbour matrices. Including too many weakly justified matrices inflates the model space without adding information; four to six candidates per study is a common pragmatic choice.
Do I still need to decompose direct and indirect effects when using spatial BMA?
Yes. In spatial lag or Durbin models, the slope coefficient is not the marginal effect because of feedback through the spatial multiplier. You must decompose impacts into average direct and average indirect (spillover) effects for each sampled model and then average those decompositions using the posterior model weights.
How do I assess convergence of the MC3 sampler over the model space?
Run multiple independent MC3 chains starting from different random models and compare the posterior model probabilities and PIPs across chains. Stability of the top-ranked models and their weights across chains is the primary convergence check; some implementations also track the proportion of the total model space visited.
Sources
- LeSage, J. P. & Pace, R. K. (2009). Introduction to Spatial Econometrics. CRC Press / Taylor & Francis. ISBN: 978-1420064247
- Fernandez, C., Ley, E. & Steel, M. F. J. (2001). Benchmark priors for Bayesian model averaging. Journal of Econometrics, 100(2), 381-427. DOI: 10.1016/S0304-4076(00)00076-2 ↗
How to cite this page
ScholarGate. (2026, June 3). Spatial Bayesian Model Averaging. ScholarGate. https://scholargate.app/en/bayesian/spatial-bayesian-model-averaging
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 Model AveragingBayesian↔ compare
- Bayesian RegressionBayesian↔ compare
- Hierarchical Bayesian InferenceBayesian↔ compare
- Spatial Bayesian InferenceBayesian↔ compare
- Spatial Variational InferenceBayesian↔ compare