Bayesian Spatial Lag Model
Bayesian Spatial Autoregressive Lag Model · Also known as: Bayesian SAR model, Bayesian spatial autoregressive model, BSLM, Bayesian SLM
The Bayesian Spatial Lag Model (BSLM) extends the classical spatial autoregressive (SAR) regression by placing prior distributions over all parameters and recovering full posterior distributions via MCMC sampling. It explicitly accounts for spatial dependence — the outcome in one location is partly driven by outcomes in neighboring locations — and yields uncertainty-quantified estimates of both regression coefficients and the spatial autocorrelation parameter rho.
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 the Bayesian Spatial Lag Model when your outcome variable exhibits positive or negative spatial autocorrelation (Moran's I significant) and you believe this dependence arises from genuine spillovers between units — for example, crime diffusion, economic growth contagion, disease spread, or housing price cascades. It is preferable over frequentist maximum-likelihood SAR estimation when sample sizes are small to moderate (fewer than a few hundred observations), when you want full posterior uncertainty on rho rather than a point estimate and asymptotic standard error, or when you need to incorporate informative prior knowledge about the strength of spatial dependence. Do NOT use it when spatial dependence originates in correlated errors rather than outcome spillovers — in that case use the Bayesian Spatial Error Model. Avoid it when the spatial weights matrix W is poorly specified or when the data are point-pattern rather than areal/lattice.
Strengths & limitations
- Provides full posterior distributions for all parameters, giving honest uncertainty quantification even in small samples.
- Correctly decomposes marginal effects into direct and indirect (spillover) impacts, which classical OLS or naive spatial models ignore.
- Handles the simultaneous-equation nature of the spatial lag specification without relying on large-sample approximations.
- Allows incorporation of prior knowledge about the plausible range or direction of spatial autocorrelation.
- Robust to moderate spatial weights matrix misspecification because the posterior averages over parameter uncertainty.
- Computationally intensive: MCMC sampling can be slow for large spatial datasets (n > 10,000) because the Jacobian |I - rho*W| must be evaluated at each iteration.
- Results depend on the choice of spatial weights matrix W, which is rarely uniquely determined by theory.
- Interpretation of indirect spillover effects can be complex and is often overlooked in applied work.
- Convergence of the MCMC chain must be diagnosed; poor mixing leads to unreliable posterior summaries.
- Requires careful prior specification for rho; an overly tight prior can dominate the likelihood in small samples.
Frequently asked
How is the Bayesian Spatial Lag Model different from running OLS on spatially autocorrelated data?
OLS ignores the simultaneity introduced by Wy on the right-hand side, yielding biased and inconsistent coefficient estimates whenever rho is non-zero. The Bayesian SAR model treats the spatial lag as endogenous, produces unbiased posterior estimates of beta and rho, and correctly accounts for the feedback loop when computing marginal effects.
How do I choose the spatial weights matrix W?
Common choices are contiguity matrices (queen or rook adjacency) for areal units, or inverse-distance matrices for location data. The specification should reflect the actual mechanism of spatial interaction. Sensitivity analyses across two or three plausible W matrices are good practice; large changes in rho across specifications signal fragility.
What is the difference between direct and indirect impacts?
The direct impact of a covariate is the average effect on a unit's own outcome. The indirect (spillover) impact is the average effect on all other units' outcomes transmitted through the spatial multiplier (I - rho*W)^-1. Total impact is their sum. Reporting only beta_k as the marginal effect underestimates the full response when rho > 0.
How many MCMC iterations are needed?
A typical practice is 10,000 to 50,000 burn-in iterations (discarded) followed by 10,000 to 100,000 retained draws. Trace plots, effective sample size, and the Gelman-Rubin R-hat statistic should all confirm adequate convergence before reporting posterior summaries.
When should I prefer the Bayesian Spatial Error Model instead?
Prefer the Spatial Error Model when spatial autocorrelation appears in the residuals of a non-spatial regression but the outcome of one unit is not thought to directly cause outcomes in neighbors — i.e., the dependence reflects omitted spatially correlated covariates rather than genuine spillovers. Robust LM tests in the frequentist literature, or DIC/WAIC model comparisons in the Bayesian setting, can guide the choice.
Sources
- LeSage, J. P., & Pace, R. K. (2009). Introduction to Spatial Econometrics. CRC Press / Taylor & Francis. ISBN: 978-1420064247
- LeSage, J. P. (1997). Bayesian Estimation of Spatial Autoregressive Models. International Regional Science Review, 20(1-2), 113-129. DOI: 10.1177/016001769702000107 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Spatial Autoregressive Lag Model. ScholarGate. https://scholargate.app/en/spatial-analysis/bayesian-spatial-lag-model
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 Durbin ModelSpatial analysis↔ compare
- Bayesian Spatial Error ModelSpatial analysis↔ compare
- Geographically Weighted RegressionSpatial analysis↔ compare
- Spatial AutocorrelationSpatial analysis↔ compare
- Spatial Lag ModelSpatial analysis↔ compare