Kamarianakis (n.d.) Hierarchical Bayesian Modeling for Spatial Time Series: An Alternative Approach to Spatial SUR

Summary

Adapts the three-stage hierarchical Bayesian space-time framework of Wikle, Berliner, and Cressie (1998) — originally developed for environmental temperature modeling — to panel regression settings with both spatial and temporal structure. The model extends spatial seemingly unrelated regressions (SUR) by allowing regression coefficients to vary across locations under a conditional autoregressive (CAR) prior and by adding a dynamic spatio-temporal residual process modeled as a vector autoregression (VAR). All full conditional distributions are derived in closed form (Gaussian for state vectors and coefficients, Inverse-Gamma for variances) enabling a Gibbs sampler; only the spatial autocorrelation parameters $\rho_i$ require Metropolis-Hastings updates due to a non-standard Uniform prior over the feasible eigenvalue range.

Key Claims

• Three-stage hierarchy. Stage 1: $Y(s,t) = M(X,\beta(s)) + K(s,t) + \gamma(s,t)$ where $M$ is a location-specific regression and $K(s,t)$ is a spatio-temporal residual. Stage 2: $\beta_i(s)$ follows a CAR prior conditional on neighboring $\beta_i(l)$; $K_t = HK_{t-1} + \eta_t$ (VAR with spatially structured $H$). Stage 3: Gaussian hyperpriors on regression hyperparameters $\alpha_i$, Uniform priors on $\rho_i$, Inverse-Gamma priors on all variances.
• CAR prior for coefficients. Each $\beta_i(s)|\{\beta_i(l): s \neq l\} \sim N(\alpha_i + \rho_i \sum_l w_{sl}\beta_i(l),\, \tau_i^2)$, implying the joint distribution $\beta_i \sim N((I - \rho_i W)^{-1}\alpha_i,\, \tau_i^2[(I-\rho_i W)'(I-\rho_i W)]^{-1})$. Nearby locations share information through $W$ and $\rho_i$.
• Feasible prior for $\rho_i$. $\rho_i \sim U[\lambda_{min}^{-1}, \lambda_{max}^{-1}]$ where $\lambda_{min}, \lambda_{max}$ are eigenvalues of row-standardized $W$. This restricts the CAR model to stationary (positive-definite covariance) configurations; the conditional for $\rho_i$ is truncated normal so Metropolis-Hastings is used.
• Gibbs sampler derivation. Full conditionals for $K_t$ (filtering recursion: Gaussian with precision $\sigma_\gamma^{-2}I + \sigma_\eta^{-2}H'H + \sigma_\eta^{-2}I$), $\beta_i$ (Gaussian combining CAR prior and data likelihood), $\alpha_i$ (Gaussian), $b(s)$, $c(s)$, variances $\sigma_\gamma^2, \sigma_\eta^2, \tau_i^2$ (all Inverse-Gamma). Sampling cycles through these in sequence.
• Identifiability caveat. $K(s,t)$ and $\gamma(s,t)$ appear only through their sum in stage 1. Formally unidentified, but proper priors at all stages allow valid Bayesian inference — Wikle et al. (1998) give the same caveat for environmental models.
• Connection to SUR. Setting $K(s,t) \equiv 0$ recovers a spatial SUR with CAR-prior coefficients. The full model generalizes this by allowing a dynamic spatio-temporal component to capture serial and cross-sectional dependence beyond the regression fit.
• Convergence diagnostics. Recommends Gelman-Rubin (1992) multi-chain $\hat{R}$ and batch-means Monte Carlo standard errors (MCSE; Roberts 1996) for estimating standard errors given Markov chain Monte Carlo (MCMC) autocorrelation.
• No empirical application. Purely methodological; uses Greek regional economics as motivating context.

Concepts Introduced or Extended

• Bayesian Hierarchical Model
• Conditional Autoregressive Model — CAR prior on location-varying regression coefficients; feasible $\rho_i$ range from eigenvalues of row-standardized $W$
• Gibbs Sampler

Entities Mentioned

• (none with existing wiki pages)

Quotes

"The proposed modeling strategy is useful in quantifying the temporal evolution and the interrelationships between variables that have a spatial reference, while accounting for the location of the measurements and their neighboring relations."

My Take

A clean methodological paper that bridges the spatial-statistics literature (Wikle-Berliner-Cressie) and the econometric panel/SUR literature. The derivation of all Gibbs conditionals is the main technical contribution. The model is quite general but the lack of any empirical application makes it hard to assess practical performance — in particular, the identifiability issue between $K(s,t)$ and $\gamma(s,t)$ is noted but not resolved. The CAR prior on regression coefficients is a natural spatial analogue of the hierarchical shrinkage priors common in Bayesian panel data. The paper is an undated working paper and likely predates LeSage-Pace (2009) which became the standard reference for Bayesian spatial econometrics.