r - ggplot2 odd results from geom_density

Question

I'm trying to replicate a conditional density plot as shown in the docs for stat_density, but I need to do it with ggplot and not qplot because I need to facet the plot. However, my data has unequal numbers of rows in each facet, and when I plot it, I get weird results. Here's an MRE:

test.df = data.frame(model=c(rep("A",100),rep("B",50),rep("C",20)),
as.factor(cnt=c(rbinom(100,1,0.5),rbinom(50,1,0.5),rbinom(20,1,0.5))),
xvar=c(seq(1,100),seq(1,50),seq(1,20)))

# This works fine, ignoring the model facets.
ggplot(test.df,aes(x=xvar,y=..count..,fill=cnt)) + geom_density(position="fill")

# Something's wrong with this one.
ggplot(test.df,aes(x=xvar,y=..count..,fill=cnt)) + geom_density(position="fill") + 
facet_wrap(~model)

Here's an example of what it looks like ignoring the model (and it looks the same if you subset and plot each group individually):

vs with faceting:

Any thoughts on what I might be doing wrong here?

edit: this is with R 2.15.1 / Rstudio 0.97.314 under OSX 10.8.2, ggplot2 0.9.3.1

Answer 1

The problems arise because it doesn't really make sense to estimate the density when you are extrapolating the x variable. (the limits for model = B are [0,50], and for model=C are [0,20]

If you set scales='free_x' then you won't be forcing ggplot to do something silly.

ggplot(test.df,aes(x=xvar,y=..count..,fill=cnt)) + 
 geom_density(position = "fill") + 
 facet_wrap(~model,scales = 'free_x')

You could alternatively pass trim = TRUE to geom_density (and thus stat_density), which would trim to the range of the observed data when calculating (and plotting) the densities

ggplot(test.df,aes(x=xvar,y=..count..,fill=cnt)) + 
 geom_density(position="fill",trim = TRUE) + 
 facet_wrap(~model)

Perhaps there should be some kind of warning regarding the nonsense results in the original attempt, but I think this may be covered by fortune(15)

library(fortunes)
fortune(15)
## It really is hard to anticipate just how silly users can be.
##   -- Brian D. Ripley
##   R-devel (October 2003)