Skip to contents

Residential burglary rates in 49 counties of Columbus, Ohio, come with the spData package. Plot the response on the map first:

library(sf); library(spdep)
map  <- st_read(system.file("shapes/columbus.gpkg", package = "spData"), quiet = TRUE)
areal.plot(map, map$CRIME, title = "crime")

Choropleth map of residential burglary rates across the 49 Columbus counties.

An intrinsic conditional autoregression (the besag model) assumes neighbouring counties are similar, with Gaussian differences across the edges of the adjacency graph, xi−xj∼N(0,1/τ)for each edge (i,j).x_i - x_j \sim N(0,\,1/\tau)\quad\text{for each edge }(i,j). The non-Gaussian version allows a few sharp boundaries between otherwise smooth regions. Build the adjacency and fit the Gaussian model:

d    <- st_drop_geometry(map[, c("CRIME", "HOVAL", "INC")])
d$s  <- seq_len(nrow(d))
nb   <- poly2nb(map)
Wadj <- Matrix::sparseMatrix(i = rep(seq_along(nb), lengths(nb)), j = unlist(nb), x = 1)

LGM <- inla(CRIME ~ 1 + HOVAL + INC +
              f(s, model = "besag", graph = Wadj, scale.model = TRUE,
                hyper = list(prec = list(prior = "loggamma", param = c(0.5, 0.01)))),
            data = d, control.compute = list(config = TRUE))
LGM$summary.hyperpar
#>                                                 mean           sd 0.025quant
#> Precision for the Gaussian observations  0.008022714 1.647426e-03 0.00519725
#> Precision for s                         57.746536399 1.157551e+02 0.28549464
#>                                            0.5quant   0.975quant        mode
#> Precision for the Gaussian observations  0.00788201   0.01165272 0.007641854
#> Precision for s                         19.78290796 336.00659816 0.125463561
ng.check(LGM, compute.fixed = FALSE)   # areal model auto-detected

Diagnostic for the intrinsic CAR: similar small sensitivities across all graph edges, with the observed total well inside its Gaussian reference.

The 118 edges of this graph all show a similarly small discrepancy from the Gaussian reference. This field is essentially Gaussian, and both the previous checking procedure (ng.check) and the latent non-Gaussian fitting (ngvb) show that.

LnGM <- ngvb(LGM)
#> Components: s [car]
#> ngvb: reached the iteration limit after 30 iteration(s);  E[eta] = 0.073
plot(LnGM)

ngvb fit summary for the besag component: edge mixing weights near one and the non-Gaussianity parameter shrinking toward zero.

bayes.factor(LnGM, n.samples = 30, seed = 1)
#> Bayes factor (non-Gaussian vs Gaussian): 1.09
#>   log10 BF = 0.04  (weak/none)
#>   weight ESS = 25.8 of 30 draws

The intrinsic besag effect puts its mixing variables on the edges of the graph rather than on the counties, so they cannot be mapped directly. The simultaneous autoregression, built by hand in Custom models, places them on the counties instead and can be mapped.