gwr.sel {spgwr}

Crossvalidation of bandwidth for geographically weighted regression

Description

The function finds a bandwidth for a given geographically weighted regression by optimzing a selected function. For cross-validation, this scores the root mean square prediction error for the geographically weighted regressions, choosing the bandwidth minimizing this quantity.

Usage

gwr.sel(formula, data=list(), coords, adapt=FALSE, gweight=gwr.Gauss,
 method = "cv", verbose = TRUE, longlat=NULL, RMSE=FALSE, weights,
 tol=.Machine$double.eps^0.25, show.error.messages = FALSE)

Arguments

formula

regression model formula as in lm

data

model data frame as in lm, or may be a SpatialPointsDataFrame or SpatialPolygonsDataFrame object as defined in package sp

coords

matrix of coordinates of points representing the spatial positions of the observations

adapt

either TRUE: find the proportion between 0 and 1 of observations to include in weighting scheme (k-nearest neighbours), or FALSE — find global bandwidth

gweight

geographical weighting function, at present gwr.Gauss() default, or gwr.gauss(), the previous default or gwr.bisquare()

method

default "cv" for drop-1 cross-validation, or "aic" for AIC optimisation (depends on assumptions about AIC degrees of freedom)

verbose

if TRUE (default), reports the progress of search for bandwidth

longlat

TRUE if point coordinates are longitude-latitude decimal degrees, in which case distances are measured in kilometers; if x is a SpatialPoints object, the value is taken from the object itself

RMSE

default FALSE to correspond with CV scores in newer references (sum of squared CV errors), if TRUE the previous behaviour of scoring by LOO CV RMSE

weights

case weights used as in weighted least squares, beware of scaling issues — only used with the cross-validation method, probably unsafe

tol

the desired accuracy to be passed to optimize

show.error.messages

default FALSE; may be set to TRUE to see error messages if gwr.sel returns without a value

Details

If the regression contains little pattern, the bandwidth will converge to the upper bound of the line search, which is the diagonal of the bounding box of the data point coordinates for “adapt=FALSE”, and 1 for “adapt=TRUE”; see the simulation block in the examples below.

Value

returns the cross-validation bandwidth.

Note

Use of method="aic" results in the creation of an n by n matrix, and should not be chosen when n is large.

Author(s)

Roger Bivand [e-mail]

References

Fotheringham, A.S., Brunsdon, C., and Charlton, M.E., 2002, Geographically Weighted Regression, Chichester: Wiley; Paez A, Farber S, Wheeler D, 2011, "A simulation-based study of geographically weighted regression as a method for investigating spatially varying relationships", Environment and Planning A 43(12) 2992-3010; http://gwr.nuim.ie/

See Also

gwr.bisquare, gwr.gauss

Examples

data(columbus, package="spData")
gwr.sel(CRIME ~ INC + HOVAL, data=columbus,
  coords=cbind(columbus$X, columbus$Y))
gwr.sel(CRIME ~ INC + HOVAL, data=columbus,
  coords=cbind(columbus$X, columbus$Y), gweight=gwr.bisquare)
## Not run: 
data(georgia)
set.seed(1)
X0 <- runif(nrow(gSRDF)*3)
X1 <- matrix(sample(X0), ncol=3)
X1 <- prcomp(X1, center=FALSE, scale.=FALSE)$x
gSRDF$X1 <- X1[,1]
gSRDF$X2 <- X1[,2]
gSRDF$X3 <- X1[,3]
yrn <- rnorm(nrow(gSRDF))
gSRDF$yrn <- sample(yrn)
bw <- gwr.sel(yrn ~ X1 + X2 + X3, data=gSRDF, method="cv", adapt=FALSE, verbose=FALSE)
bw
bw <- gwr.sel(yrn ~ X1 + X2 + X3, data=gSRDF, method="aic", adapt=FALSE, verbose=FALSE)
bw
bw <- gwr.sel(yrn ~ X1 + X2 + X3, data=gSRDF, method="cv", adapt=TRUE, verbose=FALSE)
bw
bw <- gwr.sel(yrn ~ X1 + X2 + X3, data=gSRDF, method="aic", adapt=TRUE, verbose=FALSE)
bw

## End(Not run)