Creating Neighbours

This vignette formed pp. 239–251 of the first edition of Bivand, R. S., Pebesma, E. and Gómez-Rubio V. (2008) Applied Spatial Data Analysis with R, Springer-Verlag, New York. It was retired from the second edition (2013) to accommodate material on other topics, and is made available in this form with the understanding of the publishers.

Introduction

Creating spatial weights is a necessary step in using areal data, perhaps just to check that there is no remaining spatial patterning in residuals. The first step is to define which relationships between observations are to be given a non-zero weight, that is to choose the neighbour criterion to be used; the second is to assign weights to the identified neighbour links.

The 281 census tract data set for eight central New York State counties featured prominently in (Waller and Gotway 2004) will be used in many of the examples, supplemented with tract boundaries derived from TIGER 1992 and distributed by SEDAC/CIESIN. The boundaries have been projected from geographical coordinates to UTM zone~18. This file is not identical with the boundaries used in the original source, but is very close and may be re-distributed, unlike the version used in the book. Starting from the census tract queen contiguities, where all touching polygons are neighbours, used in (Waller and Gotway 2004) and provided as a DBF file on their website, a GAL format file has been created and read into R.

library(spdep)
if (packageVersion("spData") >= "2.3.2") {
    NY8a <- sf::st_read(system.file("shapes/NY8_utm18.gpkg", package="spData"))
} else {
    NY8a <- sf::st_read(system.file("shapes/NY8_bna_utm18.gpkg", package="spData"))
    sf::st_crs(NY8a) <- "EPSG:32618"
    NY8a$Cases <- NY8a$TRACTCAS
}
## Reading layer `NY8_utm18' from data source 
##   `/home/rsb/lib/r_libs/spData/shapes/NY8_utm18.gpkg' using driver `GPKG'
## Simple feature collection with 281 features and 17 fields
## Geometry type: POLYGON
## Dimension:     XY
## Bounding box:  xmin: 358241.9 ymin: 4649755 xmax: 480393.1 ymax: 4808545
## Projected CRS: WGS 84 / UTM zone 18N
NY8 <- as(NY8a, "Spatial")
NY_nb <- read.gal(system.file("weights/NY_nb.gal", package="spData"), region.id=as.character(as.integer(row.names(NY8))-1L))

For the sake of simplicity in showing how to create neighbour objects, we work on a subset of the map consisting of the census tracts within Syracuse, although the same principles apply to the full data set. We retrieve the part of the neighbour list in Syracuse using the subset method.

Syracuse <- NY8[!is.na(NY8$AREANAME) & NY8$AREANAME == "Syracuse city",]
Sy0_nb <- subset(NY_nb, !is.na(NY8$AREANAME) & NY8$AREANAME == "Syracuse city")
summary(Sy0_nb)
## Neighbour list object:
## Number of regions: 63 
## Number of nonzero links: 346 
## Percentage nonzero weights: 8.717561 
## Average number of links: 5.492063 
## Link number distribution:
## 
##  1  2  3  4  5  6  7  8  9 
##  1  1  5  9 14 17  9  6  1 
## 1 least connected region:
## 164 with 1 link
## 1 most connected region:
## 136 with 9 links

Creating Contiguity Neighbours

We can create a copy of the same neighbours object for polygon contiguities using the poly2nb() function in spdep. It takes an object extending the SpatialPolygons class as its first argument, and using heuristics identifies polygons sharing boundary points as neighbours. It also has a snap= argument, to allow the shared boundary points to be a short distance from one another.

class(Syracuse)
## [1] "SpatialPolygonsDataFrame"
## attr(,"package")
## [1] "sp"
Sy1_nb <- poly2nb(Syracuse)
isTRUE(all.equal(Sy0_nb, Sy1_nb, check.attributes=FALSE))
## [1] TRUE

As we can see, creating the contiguity neighbours from the Syracuse object reproduces the neighbours from (Waller and Gotway 2004). Careful examination of the next figure shows, however, that the graph of neighbours is not planar, since some neighbour links cross each other. By default, the contiguity condition is met when at least one point on the boundary of one polygon is within the snap distance of at least one point of its neighbour. This relationship is given by the argument queen=TRUE by analogy with movements on a chessboard. So when three or more polygons meet at a single point, they all meet the contiguity condition, giving rise to crossed links. If queen=FALSE, at least two boundary points must be within the snap distance of each other, with the conventional name of a `rook’ relationship. The figure shows the crossed line differences that arise when polygons touch only at a single point, compared to the stricter rook criterion.

Sy2_nb <- poly2nb(Syracuse, queen=FALSE)
isTRUE(all.equal(Sy0_nb, Sy2_nb, check.attributes=FALSE))
## [1] FALSE
oopar <- par(mfrow=c(1,2), mar=c(3,3,1,1)+0.1)
plot(Syracuse, border="grey60")
plot(Sy0_nb, coordinates(Syracuse), add=TRUE, pch=19, cex=0.6)
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="a)", cex=0.8)
plot(Syracuse, border="grey60")
plot(Sy0_nb, coordinates(Syracuse), add=TRUE, pch=19, cex=0.6)
plot(diffnb(Sy0_nb, Sy2_nb, verbose=FALSE), coordinates(Syracuse),
  add=TRUE, pch=".", cex=0.6, lwd=2, col="orange")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="b)", cex=0.8)
par(oopar)

a: Queen-style census tract contiguities, Syracuse; b: Rook-style contiguity differences shown as thicker orange lines

If we have access to a GIS such as GRASS or ArcGIS, we can export the SpatialPolygonsDataFrame object and use the topology engine in the GIS to find contiguities in the graph of polygon edges - a shared edge will yield the same output as the rook relationship.

This procedure does, however, depend on the topology of the set of polygons being clean, which holds for this subset, but not for the full eight-county data set. Not infrequently, there are small artefacts, such as slivers where boundary lines intersect or diverge by distances that cannot be seen on plots, but which require intervention to keep the geometries and data correctly associated. When these geometrical artefacts are present, the topology is not clean, because unambiguous shared polygon boundaries cannot be found in all cases; artefacts typically arise when data collected for one purpose are combined with other data or used for another purpose. Topologies are usually cleaned in a GIS by snapping' vertices closer than a threshold distance together, removing artefacts -- for example, snapping across a river channel where the correct boundary is the median line but the input polygons stop at the channel banks on each side. Thepoly2nb()function does have asnap` argument, which may also be used when input data possess geometrical artefacts.

library(rgrass)
v <- terra::vect(sf::st_as_sf(Syracuse))
SG <- terra::rast(terra::ext(v), crs=terra::crs(v))
pr <- initGRASS("/home/rsb/topics/grass/g840/grass84", tempdir(), SG=SG, override=TRUE)
write_VECT(v, "SY0", flags=c("o", "overwrite"))
contig <- vect2neigh("SY0")
Sy3_nb <- sn2listw(contig, style="B")$neighbours
isTRUE(all.equal(Sy3_nb, Sy2_nb, check.attributes=FALSE))
## [1] TRUE

Similar approaches may also be used to read ArcGIS coverage data by tallying the left neighbour and right neighbour arc indices with the polygons in the data set, using either RArcInfo or rgdal.

In our Syracuse case, there are no exclaves or islands' belonging to the data set, but not sharing boundary points within the snap distance. If the number of polygons is moderate, the missing neighbour links may be added interactively using theedit()method fornb` objects, and displaying the polygon background. The same method may be used for removing links which, although contiguity exists, may be considered void, such as across a mountain range.

Creating Graph-Based Neighbours

Continuing with irregularly located areal entities, it is possible to choose a point to represent the polygon-support entities. This is often the polygon centroid, which is not the average of the coordinates in each dimension, but takes proper care to weight the component triangles of the polygon by area. It is also possible to use other points, or if data are available, construct, for example population-weighted centroids. Once representative points are available, the criteria for neighbourhood can be extended from just contiguity to include graph measures, distance thresholds, and \(k\)-nearest neighbours.

The most direct graph representation of neighbours is to make a Delaunay triangulation of the points, shown in the first panel in the figure. The neighbour relationships are defined by the triangulation, which extends outwards to the convex hull of the points and which is planar. Note that graph-based representations construct the interpoint relationships based on Euclidean distance, with no option to use Great Circle distances for geographical coordinates. Because it joins distant points around the convex hull, it may be worthwhile to thin the triangulation as a Sphere of Influence (SOI) graph, removing links that are relatively long. Points are SOI neighbours if circles centred on the points, of radius equal to the points’ nearest neighbour distances, intersect in two places (Avis and Horton 1985). Functions for graph-based neighbours were kindly contributed by Nicholas Lewin-Koh.

coords <- coordinates(Syracuse)
IDs <- row.names(as(Syracuse, "data.frame"))
#FIXME library(tripack)
Sy4_nb <- tri2nb(coords, row.names=IDs)
if (require(dbscan, quietly=TRUE)) {
  Sy5_nb <- graph2nb(soi.graph(Sy4_nb, coords), row.names=IDs)
} else Sy5_nb <- NULL
## 
## Attaching package: 'dbscan'
## The following object is masked from 'package:stats':
## 
##     as.dendrogram
Sy6_nb <- graph2nb(gabrielneigh(coords), row.names=IDs)
Sy7_nb <- graph2nb(relativeneigh(coords), row.names=IDs)
oopar <- par(mfrow=c(2,2), mar=c(1,1,1,1)+0.1)
plot(Syracuse, border="grey60")
plot(Sy4_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="a)", cex=0.8)
plot(Syracuse, border="grey60")
if (!is.null(Sy5_nb)) {
  plot(Sy5_nb, coords, add=TRUE, pch=".")
  text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="b)", cex=0.8)
}
plot(Syracuse, border="grey60")
plot(Sy6_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="c)", cex=0.8)
plot(Syracuse, border="grey60")
plot(Sy7_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="d)", cex=0.8)
par(oopar)

a: Delauney triangulation neighbours; b: Sphere of influence neighbours (if available); c: Gabriel graph neighbours; d: Relative graph neighbours

Delaunay triangulation neighbours and SOI neighbours are symmetric by design – if \(i\) is a neighbour of \(j\), then \(j\) is a neighbour of \(i\). The Gabriel graph is also a subgraph of the Delaunay triangulation, retaining a different set of neighbours (Matula and Sokal 1980). It does not, however, guarantee symmetry; the same applies to Relative graph neighbours (Toussaint 1980). The graph2nb() function takes a sym argument to insert links to restore symmetry, but the graphs then no longer exactly fulfil their neighbour criteria. All the graph-based neighbour schemes always ensure that all the points will have at least one neighbour. Subgraphs of the full triangulation may also have more than one graph after trimming. The functions is.symmetric.nb() can be used to check for symmetry, with argument force=TRUE if the symmetry attribute is to be overridden, and n.comp.nb() reports the number of graph components and~the components to which points belong (after enforcing symmetry, because the algorithm assumes that the graph is not directed). When there are more than one graph component, the matrix representation of the spatial weights can become block-diagonal if observations are appropriately sorted.

nb_l <- list(Triangulation=Sy4_nb, Gabriel=Sy6_nb,
  Relative=Sy7_nb)
if (!is.null(Sy5_nb)) nb_l <- c(nb_l, list(SOI=Sy5_nb))
sapply(nb_l, function(x) is.symmetric.nb(x, verbose=FALSE, force=TRUE))
## Triangulation       Gabriel      Relative           SOI 
##          TRUE         FALSE         FALSE          TRUE
sapply(nb_l, function(x) n.comp.nb(x)$nc)
## Triangulation       Gabriel      Relative           SOI 
##             1             1             1             1

Distance-Based Neighbours

An alternative method is to choose the \(k\) nearest points as neighbours – this adapts across the study area, taking account of differences in the densities of areal entities. Naturally, in the overwhelming majority of cases, it leads to asymmetric neighbours, but will ensure that all areas have \(k\) neighbours. The knearneigh() function returns an intermediate form converted to an nb object by knn2nb(); knearneigh() can also take a longlat= argument to handle geographical coordinates.

Sy8_nb <- knn2nb(knearneigh(coords, k=1), row.names=IDs)
## Warning in knn2nb(knearneigh(coords, k = 1), row.names = IDs): neighbour object has 15
## sub-graphs
Sy9_nb <- knn2nb(knearneigh(coords, k=2), row.names=IDs)
Sy10_nb <- knn2nb(knearneigh(coords, k=4), row.names=IDs)
nb_l <- list(k1=Sy8_nb, k2=Sy9_nb, k4=Sy10_nb)
sapply(nb_l, function(x) is.symmetric.nb(x, verbose=FALSE, force=TRUE))
##    k1    k2    k4 
## FALSE FALSE FALSE
sapply(nb_l, function(x) n.comp.nb(x)$nc)
## k1 k2 k4 
## 15  1  1
oopar <- par(mfrow=c(1,3), mar=c(1,1,1,1)+0.1)
plot(Syracuse, border="grey60")
plot(Sy8_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="a)", cex=0.8)
plot(Syracuse, border="grey60")
plot(Sy9_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="b)", cex=0.8)
plot(Syracuse, border="grey60")
plot(Sy10_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="c)", cex=0.8)
par(oopar)

a: \(k=1\) neighbours; b: \(k=2\) neighbours; c: \(k=4\) neighbours

The figure shows the neighbour relationships for \(k= 1, 2, 4\), with many components for \(k=1\). If need be, \(k\)-nearest neighbour objects can be made symmetrical using the make.sym.nb() function. The \(k=1\) object is also useful in finding the minimum distance at which all areas have a distance-based neighbour. Using the nbdists() function, we can calculate a list of vectors of distances corresponding to the neighbour object, here for first nearest neighbours. The greatest value will be the minimum distance needed to make sure that all the areas are linked to at least one neighbour. The dnearneigh() function is used to find neighbours with an interpoint distance, with arguments d1 and d2 setting the lower and upper distance bounds; it can also take a longlat argument to handle geographical coordinates.

dsts <- unlist(nbdists(Sy8_nb, coords))
summary(dsts)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   395.7   587.3   700.1   760.4   906.1  1544.6
max_1nn <- max(dsts)
max_1nn
## [1] 1544.615
Sy11_nb <- dnearneigh(coords, d1=0, d2=0.75*max_1nn, row.names=IDs)
## Warning in dnearneigh(coords, d1 = 0, d2 = 0.75 * max_1nn, row.names = IDs): neighbour
## object has 4 sub-graphs
Sy12_nb <- dnearneigh(coords, d1=0, d2=1*max_1nn, row.names=IDs)
Sy13_nb <- dnearneigh(coords, d1=0, d2=1.5*max_1nn, row.names=IDs)
nb_l <- list(d1=Sy11_nb, d2=Sy12_nb, d3=Sy13_nb)
sapply(nb_l, function(x) is.symmetric.nb(x, verbose=FALSE, force=TRUE))
##   d1   d2   d3 
## TRUE TRUE TRUE
sapply(nb_l, function(x) n.comp.nb(x)$nc)
## d1 d2 d3 
##  4  1  1
oopar <- par(mfrow=c(1,3), mar=c(1,1,1,1)+0.1)
plot(Syracuse, border="grey60")
plot(Sy11_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="a)", cex=0.8)
plot(Syracuse, border="grey60")
plot(Sy12_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="b)", cex=0.8)
plot(Syracuse, border="grey60")
plot(Sy13_nb, coords, add=TRUE, pch=".")
text(bbox(Syracuse)[1,1], bbox(Syracuse)[2,2], labels="c)", cex=0.8)
par(oopar)

a: Neighbours within 1,158m; b: neighbours within 1,545m; c: neighbours within 2,317m

The figure shows how the numbers of distance-based neighbours increase with moderate increases in distance. Moving from \(0.75\) times the minimum all-included distance, to the all-included distance, and \(1.5\) times the minimum all-included distance, the numbers of links grow rapidly. This is a major problem when some of the first nearest neighbour distances in a study area are much larger than others, since to avoid no-neighbour areal entities, the distance criterion will need to be set such that many areas have many neighbours.

dS <- c(0.75, 1, 1.5)*max_1nn
res <- sapply(nb_l, function(x) table(card(x)))
mx <- max(card(Sy13_nb))
res1 <- matrix(0, ncol=(mx+1), nrow=3)
rownames(res1) <- names(res)
colnames(res1) <- as.character(0:mx)
res1[1, names(res$d1)] <- res$d1
res1[2, names(res$d2)] <- res$d2
res1[3, names(res$d3)] <- res$d3
library(RColorBrewer)
pal <- grey.colors(3, 0.95, 0.55, 2.2)
# RSB quietening greys
barplot(res1, col=pal, beside=TRUE, legend.text=FALSE, xlab="numbers of neighbours", ylab="tracts")
legend("topright", legend=format(dS, digits=1), fill=pal, bty="n", cex=0.8, title="max. distance")

Distance-based neighbours: frequencies of numbers of neighbours by census tract

The figure shows the counts of sizes of sets of neighbours for the three different distance limits. In Syracuse, the census tracts are of similar areas, but were we to try to use the distance-based neighbour criterion on the eight-county study area, the smallest distance securing at least one neighbour for every areal entity is over 38km.

dsts0 <- unlist(nbdists(NY_nb, coordinates(NY8)))
summary(dsts0)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    82.7  1505.0  3378.7  5865.8  8954.3 38438.1

If the areal entities are approximately regularly spaced, using distance-based neighbours is not necessarily a problem. Provided that care is taken to handle the side effects of “weighting” areas out of the analysis, using lists of neighbours with no-neighbour areas is not necessarily a problem either, but certainly ought to raise questions. Different disciplines handle the definition of neighbours in their own ways by convention; in particular, it seems that ecologists frequently use distance bands. If many distance bands are used, they approach the variogram, although the underlying understanding of spatial autocorrelation seems to be by contagion rather than continuous.

Higher-Order Neighbours

Distance bands can be generated by using a sequence of d1 and d2 argument values for the dnearneigh() function if needed to construct a spatial autocorrelogram as understood in ecology. In other conventions, correlograms are constructed by taking an input list of neighbours as the first-order sets, and stepping out across the graph to second-, third-, and higher-order neighbours based on the number of links traversed, but not permitting cycles, which could risk making \(i\) a neighbour of \(i\) itself (O’Sullivan and Unwin 2003). The nblag() function takes an existing neighbour list and returns a list of lists, from first to maxlag= order neighbours.

Sy0_nb_lags <- nblag(Sy0_nb, maxlag=9)
## Warning in nblag(Sy0_nb, maxlag = 9): lag 6 neighbour object has 7 sub-graphs
## Warning in nblag(Sy0_nb, maxlag = 9): lag 7 neighbour object has 22 sub-graphs
## Warning in nblag(Sy0_nb, maxlag = 9): lag 8 neighbour object has 52 sub-graphs
## Warning in nblag(Sy0_nb, maxlag = 9): lag 9 neighbour object has 63 sub-graphs
names(Sy0_nb_lags) <- c("first", "second", "third", "fourth", "fifth", "sixth", "seventh", "eighth", "ninth")
res <- sapply(Sy0_nb_lags, function(x) table(card(x)))
mx <- max(unlist(sapply(Sy0_nb_lags, function(x) card(x))))
nn <- length(Sy0_nb_lags)
res1 <- matrix(0, ncol=(mx+1), nrow=nn)
rownames(res1) <- names(res)
colnames(res1) <- as.character(0:mx)
for (i in 1:nn) res1[i, names(res[[i]])] <- res[[i]]
res1
##          0 1 2 3 4  5  6 7 8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
## first    0 1 1 5 9 14 17 9 6  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
## second   0 0 0 0 2  2  0 6 6 11 11  4  3  7  4  6  1  0  0  0  0  0  0  0  0
## third    0 0 0 0 0  0  0 1 3  5  5  7 14  6  8  3  3  0  1  1  1  3  1  0  1
## fourth   0 0 0 0 0  0  0 0 1  3  5  7 16 16  5  3  3  2  0  1  1  0  0  0  0
## fifth    0 0 0 1 1  3  1 1 3  7 13 12  8  9  3  1  0  0  0  0  0  0  0  0  0
## sixth    6 3 0 2 8  2  5 5 4  8  9  5  5  1  0  0  0  0  0  0  0  0  0  0  0
## seventh 21 7 4 5 9  7  3 5 1  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0
## eighth  49 6 5 2 1  0  0 0 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
## ninth   63 0 0 0 0  0  0 0 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0

This shows how the wave of connectedness in the graph spreads to the third order, receding to the eighth order, and dying away at the ninth order - there are no tracts nine steps from each other in this graph. Both the distance bands and the graph step order approaches to spreading neighbourhoods can be used to examine the shape of relationship intensities in space, like the variogram, and can be used in attempting to look at the effects of scale.

Grid Neighbours

When the data are known to be arranged in a regular, rectangular grid, the cell2nb() function can be used to construct neighbour lists, including those on a torus. These are useful for simulations, because, since all areal entities have equal numbers of neighbours, and there are no edges, the structure of the graph is as neutral as can be achieved. Neighbours can either be of type rook or queen.

cell2nb(7, 7, type="rook", torus=TRUE)
## Neighbour list object:
## Number of regions: 49 
## Number of nonzero links: 196 
## Percentage nonzero weights: 8.163265 
## Average number of links: 4
cell2nb(7, 7, type="rook", torus=FALSE)
## Neighbour list object:
## Number of regions: 49 
## Number of nonzero links: 168 
## Percentage nonzero weights: 6.997085 
## Average number of links: 3.428571

When a regular, rectangular grid is not complete, then we can use knowledge of the cell size stored in the grid topology to create an appropriate list of neighbours, using a tightly bounded distance criterion. Neighbour lists of this kind are commonly found in ecological assays, such as studies of species richness at a national or continental scale. It is also in these settings, with moderately large \(n\), here \(n=3\),103, that the use of a sparse, list based representation shows its strength. Handling a \(281 \times 281\) matrix for the eight-county census tracts is feasible, easy for a \(63 \times 63\) matrix for Syracuse census tracts, but demanding for a 3,\(103 \times 3\),103 matrix.

data(meuse.grid)
coordinates(meuse.grid) <- c("x", "y")
gridded(meuse.grid) <- TRUE
dst <- max(slot(slot(meuse.grid, "grid"), "cellsize"))
mg_nb <- dnearneigh(coordinates(meuse.grid), 0, dst)
mg_nb
## Neighbour list object:
## Number of regions: 3103 
## Number of nonzero links: 12022 
## Percentage nonzero weights: 0.1248571 
## Average number of links: 3.874315
table(card(mg_nb))
## 
##    1    2    3    4 
##    1  133  121 2848

References

Avis, D., and J. Horton. 1985. “Remarks on the Sphere of Influence Graph.” In Discrete Geometry and Convexity, edited by J. E. Goodman, 323–27. New York: New York Academy of Sciences, New York.
Matula, D. W., and R. R. Sokal. 1980. “Properties of Gabriel Graphs Relevant to Geographic Variation Research and the Clustering of Points in the Plane.” Geographical Analysis 12: 205–22.
O’Sullivan, D., and D. J. Unwin. 2003. Geographical Information Analysis. Hoboken, NJ: Wiley.
Toussaint, G. T. 1980. “The Relative Neighborhood Graph of a Finite Planar Set.” Pattern Recognition 12: 261–68.
Waller, L. A., and C. A. Gotway. 2004. Applied Spatial Statistics for Public Health Data. Hoboken, NJ: John Wiley & Sons.