Cressie and Chan develop and illustrate a Markov-random-field approach to modeling regional (areal) data, using the now-classic dataset of Sudden Infant Death Syndrome counts in the counties of North Carolina, 1974–1978. They separate large-scale variation (a spatial trend, or mean surface) from small-scale variation (the variance and spatial dependence in the residuals), and show that combining resistant (robust) trend fitting with a simple spatial auto-Gaussian (conditional autoregressive) model for the residuals is an effective way to analyze the transformed data.
"We analyze a data set of Sudden Infant Deaths, 1974–1978, in the counties of North Carolina, using a Markov-random-field approach to spatial modeling. We model the spatial trend with … large-scale-variation parameters, and the variance and spatial dependence with small-scale variation parameters."
A template-setting applied paper: it takes Besag's fairly abstract auto-model / CAR machinery and shows, on a real and messy county-count dataset, how to actually build a spatial model — decompose into trend and dependence, be robust in the trend, and let a Gaussian CAR mop up the residual autocorrelation. Its most enduring legacy is arguably sociological: the North Carolina SIDS data became the worked example for areal spatial statistics, appearing in Cressie's textbook and countless later methods papers (Neelon-Gelfand, disease-mapping work), which is how a good benchmark disciplines a field. The large-scale/small-scale split is also a clean way to think about the identification tension the CAR page emphasizes — where the trend ends and the spatially-dependent noise begins is exactly what a mis-specified mean surface can confound.