Design-weighted graphics

There are (at least) three useful strategies for graphing survey data.
  1. Draw a conventional graph and annotate it to indicate sampling weights
  2. Graph some feature of the estimated population distribution
  3. Simulate a simple random sample from the estimated population distribution and graph that.
The survey package has examples of all three of these in the svyplot and svyhist functions.

The first approach produces the "bubble plot". Here we show the relationship between the 2000 and 1999 API scores for a two-stage cluster sample of California schools. The radius of the bubble for each school is proportional to the sampling weight.

data(api)
dclus2<-svydesign(id=~dnum+snum, fpc=~fpc1+fpc2, data=apiclus2)
svyplot(api00~api99, dclus2)

The second strategy is used by the hexagonal binning plots and survey-weighted histograms. Hexagonal binning plots divide the screen on a hexagonal grid and show the number of points falling in each hexagon. The survey-weighted hexbin plot shows estimated population numbers in each cell. It requires the "hexbin" package from Bioconductor.

This example uses data from NHANES 2, and shows the estimated relationship between serum iron and serum transferrin in the US population

svyplot(iron~trnsfern,style="hex",dhanes,xlab="Transferrin",ylab="Iron",legend=2)
svyplot(iron~trnsfern,style="grayhex",dhanes,xlab="Transferrin",ylab="Iron",legend=2)

Another example uses the API data and compares the estimated cumulative distribution to the known population cumulative distribution

data(api)
dstrat <- svydesign(id = ~1, strata = ~stype, weights = ~pw, data = apistrat,      fpc = ~fpc)
 cdf.est<-svycdf(~enroll+api00+api99, dstrat)
cdf.pop<-ecdf(apipop$enroll)

plot(cdf.pop, main="Population vs estimate",xlab="Enrollment")
lines(cdf.est[[1]],col.points="red")
legend("right",col=c("red","black"),pch=1, legend=c("Estimate","Population"),bty="n",lty=1)
 

Smoothing

This example uses data from PEAS Examplar 2, which examines internet use in Scotland from the Scottish Household Survey.

load("ex2.RData")
shs.des <- svydesign(id=~PSU, weights=~GROSSWT,strata=~STRATUM,data=shs)

plot(c(15,83),c(0,65),type='n',xlab='age',ylab='% internet users')
legend("topright",lty=c(1,1),col=c("blue","red"),legend=c("Male","Female"),bty="n")
lf<-svysmooth(I(100*INTUSE)~AGE,subset(shs.des,SEX=="female" & !is.na(AGE)),bandwidth=10)
lines(lf,col="red")
lm<-svysmooth(I(100*INTUSE)~AGE,subset(shs.des,SEX=="male" & !is.na(AGE)),bandwidth=10)
lines(lm, col="blue")

shs.des <- update(shs.des, agegp=cut(AGE, c(0,25,35,45,55,66,75,Inf)))
means<-svyby(~INTUSE,~agegp+SEX,svymean,design=shs.des, na.rm=TRUE)
points((2:8)*10, means[1:7,3]*100,col="blue")
points((2:8)*10, means[7+1:7,3]*100,col="red")
The graph of internet use by age and sex looks like the one in the PEAS examplar, but using svysmooth makes it much easier to produce


Thomas Lumley
Last modified: Thu Jun 29 09:43:10 PDT 2006