spacom (version 1.0-5)

spacom-package: Spatially weighted context data for multilevel modelling

Description

The package provides tools to construct and exploit spatially weighted context data. Spatial weights are derived by a Kernel function from a user-defined matrix of distances between contextual units. Spatial weights can then be applied either to precise contextual measures or to aggregate estimates based on micro-level survey data, to compute spatially weighted context data. Available aggregation functions include indicators of central tendency, dispersion, or inter-group variability, and take into account survey design weights. The package further allows combining the resulting spatially weighted context data with individual-level predictor and outcome variables, for the purposes of multilevel modelling. An ad hoc stratified bootstrap resampling procedure generates robust point estimates for multilevel regression coefficients and model fit indicators, and computes confidence intervals adjusted for measurement dependency and measurement error of aggregate estimates. As an additional feature, residual and explained spatial dependency can be estimated for the tested models.

Details

Package:

spacom

Type:

Package

Version:

1.0-0

Date:

2012-12-21

Authors:

Till Junge

Sandra Penic

Mathieu Cossuta

Guy Elcheroth

Contributors:

Davide Morselli

Stephanie Glaeser

Maintainer:

Till Junge

License:

GPL (>= 2)

LazyLoad:

yes

Depends:

methods, R (>= 2.14-0), spdep, foreach, lme4, boot, classInt, iterators, rgeos

The package is structured by seven core functions, organised along four logical steps. In each of the first three steps three types of input data required for multilevel modelling with spatially weighted context data are successively imported and transformed, to generate output objects required in subsequent steps:

Step 1: Creation of spatial weighting matrices. User-specified Kernel functions and bandwidth values (function WeightMatrix) are applied to one or several user-defined distance matrices.

Step 2: Construction of spatially weighted context data. The spatial weights generated in step 1 are combined with (a) imported precise contextual measures (function SpawExact) and/or (b) imported micro-level survey data to compute contextual estimates through a user-specified aggregation function (function SpawAggregate).

Step 3: Multilevel modelling with spatially weighted context data. Imported individual-level predictor and outcome variables are combined with the precise (functions MLSpawExact and ResampleMLSpawExact) or aggregate (function ResampleMLSpawAggregate) context data generated in step 2.

Finally, contextual residuals from multilevel models can be re-used in an additional, fourth step to compute estimates of explained spatial dependency (function MLSpawResidMoran).

The package furthermore provides a fast-track procedure, which condenses steps 1 to 3 and directly generates point estimates for regression coefficients across a range of different bandwidth values, in order to explore the scale-dependency of contextual effects prior to more intensive analyses (function ExploreMLSpawExact)

References

Elcheroth, G., Penic, S., Fasel, R., Giudici, F., Glaeser, S., Joye, D., Le Goff, J.-M., Morselli, D., & Spini, D. (2012). Spatially weighted context data: a new approach for modelling the impact of collective experiences. LIVES Working Papers, 19.

Examples

## Testing the model described in Elcheroth et al (2012), page 25, Equation 9.4
## Collective guilt assignment (outcome) is predicted by a series of individual level predictors,
## and two contextual predictors: risk of war victimization weighted by geographical proximity and
## common territorial identification

## DATA PREPARATION
## Import data
## load individual level data
data(traces_ind)

## exclude variable "cg_acc" and missing values
traces_ind <- na.exclude(traces_ind[,-6])

## load contextual indicator for aggregation
data(traces_event)


## load distance matrices
data(d_geo)
data(d_ident)

## EXPLORATORY ANALYSIS

## Examining the impact of the risk of war victimization weighted by geographical
## proximity weights with bandwidths of 25, 50 and 100 km

explo.wv <- ExploreMLSpawAggregate(individual.level.data=traces_ind,
                                   contextual.name="w_all",
                                   contextual.data=traces_event,
                                   context.id="area.name",
                                   formula=cg_ass~victim_d+comb_d+
                                   male+age_1990+high_school+higher_edu+
                                   (1|area.name),
                                   distance.matrix=d_geo,
                                   multilevel.bandwidths=c(25, 50, 100),
                                   design.weight.names = "weight",
                                   aggregation.function = "weighted.mean",
                                   kernel = NULL)

## STEP 1: Create spatial weights

## geographical proximity, bandwidth=50
geow.50 <- WeightMatrix(d_geo, bandwidth=50)

## common territorial identification, bandwidth=2
w.id <- WeightMatrix(d_ident,bandwidth=2)

## STEP 2: Create spatially weighted aggregated indicator of risk of war victimization

wv.agg <- SpawAggregate(contextual.data=traces_event,
                        context.id="area.name",
                        contextual.names=c("w_all", "w_all"),
                        contextual.weight.matrices=list(geow.50, w.id),
                        aggregation.functions="weighted.mean",
                        design.weight.names="weight",
                        nb.resamples=2)
## If wanted, the names of aggregated indicators in wv.agg can be changed
names(wv.agg) <- c("wv.50", "wv.id")

## STEP 3: Perform multilevel analysis (Eq 9.4, Elcheroth et al (2012), p. 25)

mla.results <- ResampleMLSpawAggregate(individual.level.data=traces_ind,
									   context.id="area.name",
			formula= cg_ass ~ victim_d+ comb_d+ male+ age_1990+
                                     high_school+ higher_edu + (1|area.name)+ wv.50 + wv.id,
                        aggregates = wv.agg, precise.data=NULL)

## STEP 4: Test the residual spatial dependency of the previous model
mla.results.moran <- MLSpawResidMoran(ml.spaw.obj=mla.results,
							          distance.matrix=d_geo,
							          bandwidths=c(25,50,100,200))

## Plotting spatial variogram with the values in mla.results.moran
## See plot() and par() for details
## Not run: 
# plot(mla.results.moran[,1], xaxt="n", xlab="Bandwidth values", ylab="Moran's I", type="b")
# axis(side=1, at=1:nrow(mla.results.moran), labels=rownames(mla.results.moran)) ## End(Not run)