Bayesian Spatial Durbin Model
Also known as: Bayesian SDM, Bayesian spatial lag-X model, Bayesian SDM with spatially lagged covariates, BSDM
The Bayesian Spatial Durbin Model (BSDM) estimates a spatial regression that simultaneously includes a spatially lagged outcome variable and spatially lagged covariates, using Bayesian inference with Markov Chain Monte Carlo sampling. It captures both endogenous and exogenous spatial spillovers while providing full posterior distributions for all parameters, quantifying uncertainty beyond what classical maximum-likelihood estimation offers.
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 Durbin Model when you have georeferenced continuous outcome data, you expect both endogenous (peer outcome) and exogenous (neighbor covariate) spatial spillovers, and you need honest uncertainty quantification for the spillover effects. It is especially valuable with small or moderate samples where maximum-likelihood inference is fragile, when incorporating prior expert knowledge (e.g., expected sign of a spillover), or when comparing models via Bayes factors. Prefer it over the classical SDM when reporting credible intervals is important for policy audiences or when convergence diagnostics reveal instability in ML estimates. Do not use it when no credible spatial weights matrix can be constructed, when computational resources are very limited (MCMC is slow for large n), or when the theoretical model strongly rules out exogenous spillovers (in that case use the Bayesian Spatial Lag Model instead).
Strengths & limitations
- Captures both endogenous outcome spillovers (rho) and exogenous covariate spillovers (theta) in a single coherent model.
- Full posterior distributions replace point estimates, providing credible intervals that are valid even in small samples.
- Bayesian model comparison (marginal likelihoods, Bayes factors) gives a principled way to test SDM against nested alternatives (SAR, SEM, SLX).
- Prior information — such as the expected sign of a spatial multiplier from theory — can be formally incorporated.
- Partial-derivative approach yields posterior-distributed direct, indirect, and total effects with honest uncertainty quantification for spillover magnitudes.
- MCMC sampling is computationally expensive for large datasets (n > several thousand) due to repeated evaluation of the log-determinant of (I - rho*W).
- Results depend on the choice of spatial weights matrix W; sensitivity analysis across multiple W specifications is necessary but adds effort.
- Interpretation of two separate spatial channels (rho and theta) is more complex than simpler SAR or SEM models, requiring careful economic theory to justify the SDM specification.
- Posterior inference is sensitive to prior choices for rho when data are sparse or spatially uninformative.
Frequently asked
How does the Bayesian SDM differ from the classical (ML) SDM?
Both estimate the same model structure. The Bayesian version uses MCMC to sample the full posterior, delivering credible intervals valid in small samples and enabling formal model comparison via Bayes factors. The classical ML version produces point estimates faster but relies on asymptotic standard errors that can be misleading when n is small or data are highly collinear.
Why are direct and indirect effects computed separately rather than reading off the beta and theta coefficients?
Because spatial feedback loops mean that a shock to one unit propagates through the entire spatial system. LeSage and Pace (2009) show that the total impact of changing predictor k is captured by the diagonal (direct) and off-diagonal (indirect) elements of the matrix (I - rho*W)^{-1}*(I*beta_k + W*theta_k), averaged across units. Reading beta_k or theta_k alone omits these feedback multipliers and gives incorrect marginal effects.
How do I choose the spatial weights matrix W?
W should reflect the theoretical mechanism through which spatial dependence operates. Common choices are contiguity matrices (queen or rook adjacency), inverse-distance matrices with a distance threshold, or k-nearest-neighbor matrices. The choice must be motivated by theory; sensitivity analysis across two or three plausible W matrices is strongly recommended.
Can the Bayesian SDM be used with panel data?
Yes. The model extends to a spatial panel (time dimension added) with fixed or random effects. LeSage (2014) discusses Bayesian estimation of spatial panel SDMs, including heterogeneous panel versions. MCMC implementation becomes more complex and slower, but the conceptual structure is the same.
When should I prefer the Bayesian Spatial Lag Model over the Bayesian SDM?
Prefer the simpler SAR (Spatial Lag) model when theory gives no reason to expect that neighbors' covariates (not just neighbors' outcomes) affect your unit's outcome. If the exogenous spillover channel (theta) is theoretically motivated, SDM is more appropriate. Bayes factors comparing the two specifications can inform this choice empirically.
Sources
- LeSage, J. P., & Pace, R. K. (2009). Introduction to Spatial Econometrics. CRC Press / Taylor & Francis. ISBN: 978-1420064247
- LeSage, J. P. (2014). Spatial Econometric Panel Data Model Comparison Using Heterogeneous Panels with Local Spatial Spillovers. Empirical Economics, 46(1), 193–211. link ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Spatial Durbin Model. ScholarGate. https://scholargate.app/en/spatial-analysis/bayesian-spatial-durbin-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 Error ModelSpatial analysis↔ compare
- Bayesian Spatial Lag ModelSpatial analysis↔ compare
- Geographically Weighted RegressionSpatial analysis↔ compare
- Spatial Durbin ModelSpatial analysis↔ compare
- Spatial Error ModelSpatial analysis↔ compare
- Spatial Lag ModelSpatial analysis↔ compare