Adding 95% contours around scatterplot points with ggplot2

Hi all,

I have been looking for means of add a contour around some points in a
scatterplot as a means of representing the center of density for of the
data. I'm imagining something like a 95% confidence estimate drawn around
the data.

So far I have found some code for drawing polygons around the data. These
look nice, but in some cases the polygons are strongly influenced by
outlying points. Does anyone have a thought on how to draw a contour which
is more along the lines of a 95% confidence space?

I have provided a working example below to illustrate the drawing of the
polygons. As I said I would rather have three "ovals"/95% contours drawn
around the points by "level" to capture the different density distributions
without the visualization being heavily influenced by outliers.

I have looked into the code provided here from Hadley
https://groups.google.com/forum/?fromgroups=#!topic/ggplot2/85q4SQ9q3V8
using the mvtnorm package and the dmvnorm function, but haven't been able
to get it work for my data example. The calculated densities are always
zero (at this step of Hadley's code: dgrid$dens <-
dmvnorm(as.matrix(dgrid), ex_mu, ex_sigma)   )

I appreciate any assistance.

Thanks,
Nate

x<-c(seq(0.15,0.4,length.out=30),seq(0.2,0.6,length.out=30),
seq(0.4,0.6,length.out=30))
y<-c(0.55,x[1:29]+0.2*rnorm(29,0.4,0.3),x[31:60]*rnorm(30,0.3,0.1),x[61:90]*rnorm(30,0.4,0.25))
data<-data.frame(level=c(rep(1, 30),rep(2,30), rep(3,30)), x=x,y=y)


find_hull <- function(data) data[chull(data$x, data$y), ]
hulls <- ddply(data, .(level), find_hull)

fig1 <- ggplot(data=data, aes(x, y, colour=(factor(level)),
fill=level))+geom_point()
fig1 <- fig1 + geom_polygon(data=hulls, alpha=.2)
fig1

        [[alternative HTML version deleted]]

Thanks Ista,

I have played a bit with stat_density2d as well. It doesn't completely
capture what I am looking for and ends up being quite busy at the same time.
I'm looking for a way of helping those looking that the figure to see the
broad patterns of where in the x/y space the data from different groups are
distributed. Using the 95% CI type idea is so that I don't end up
arbitrarily drawing circles around each set of points. I appreciate your
direction though.

Nate


On Mon, Jan 28, 2013 at 10:50 AM, Ista Zahn <istazahn at gmail.com> wrote:

Hi Nathan,

This only fits some of your criteria, but have you looked at
?stat_density2d?

Best,
Ista

On Mon, Jan 28, 2013 at 12:53 PM, Nathan Miller <natemiller77 at gmail.com>
wrote:

Hi all,

I have been looking for means of add a contour around some points in a
scatterplot as a means of representing the center of density for of the
data. I'm imagining something like a 95% confidence estimate drawn
around
the data.

So far I have found some code for drawing polygons around the data.
These
look nice, but in some cases the polygons are strongly influenced by
outlying points. Does anyone have a thought on how to draw a contour
which
is more along the lines of a 95% confidence space?

I have provided a working example below to illustrate the drawing of the
polygons. As I said I would rather have three "ovals"/95% contours drawn
around the points by "level" to capture the different density
distributions
without the visualization being heavily influenced by outliers.

I have looked into the code provided here from Hadley
https://groups.google.com/forum/?fromgroups=#!topic/ggplot2/85q4SQ9q3V8
using the mvtnorm package and the dmvnorm function, but haven't been
able
to get it work for my data example. The calculated densities are always
zero (at this step of Hadley's code: dgrid$dens <-
dmvnorm(as.matrix(dgrid), ex_mu, ex_sigma)   )

I appreciate any assistance.

Thanks,
Nate

x<-c(seq(0.15,0.4,length.out=30),seq(0.2,0.6,length.out=30),
seq(0.4,0.6,length.out=30))

y<-c(0.55,x[1:29]+0.2*rnorm(29,0.4,0.3),x[31:60]*rnorm(30,0.3,0.1),x[61:90]*rnorm(30,0.4,0.25))
data<-data.frame(level=c(rep(1, 30),rep(2,30), rep(3,30)), x=x,y=y)


find_hull <- function(data) data[chull(data$x, data$y), ]
hulls <- ddply(data, .(level), find_hull)

fig1 <- ggplot(data=data, aes(x, y, colour=(factor(level)),
fill=level))+geom_point()
fig1 <- fig1 + geom_polygon(data=hulls, alpha=.2)
fig1

        [[alternative HTML version deleted]]

Hi Ista,

Thanks. That does look pretty nice and I hadn't realized that was possible.
Do you know how to extract information regarding those curves? I'd like to
be able to report something about what portion of the data they encompass or
really any other feature about them in a figure legend. I'll look into
stat_density2d and see if I can determine how they are set.

Thanks for your help,

Nate


On Mon, Jan 28, 2013 at 12:37 PM, Ista Zahn <istazahn at gmail.com> wrote:

Hi Nate,

You can make it less busy using the bins argument. This is not
documented, except in the examples to stat_contour, but try

ggplot(data=data, aes(x, y, colour=(factor(level)), fill=level))+
        geom_point()+
        stat_density2d(bins=2)

HTH,
Ista

On Mon, Jan 28, 2013 at 2:43 PM, Nathan Miller <natemiller77 at gmail.com>
wrote:

Thanks Ista,

I have played a bit with stat_density2d as well. It doesn't completely
capture what I am looking for and ends up being quite busy at the same
time.
I'm looking for a way of helping those looking that the figure to see
the
broad patterns of where in the x/y space the data from different groups
are
distributed. Using the 95% CI type idea is so that I don't end up
arbitrarily drawing circles around each set of points. I appreciate your
direction though.

Nate


On Mon, Jan 28, 2013 at 10:50 AM, Ista Zahn <istazahn at gmail.com> wrote:

Hi Nathan,

This only fits some of your criteria, but have you looked at
?stat_density2d?

Best,
Ista

On Mon, Jan 28, 2013 at 12:53 PM, Nathan Miller
<natemiller77 at gmail.com>
wrote:

Hi all,

I have been looking for means of add a contour around some points in
a
scatterplot as a means of representing the center of density for of
the
data. I'm imagining something like a 95% confidence estimate drawn
around
the data.

So far I have found some code for drawing polygons around the data.
These
look nice, but in some cases the polygons are strongly influenced by
outlying points. Does anyone have a thought on how to draw a contour
which
is more along the lines of a 95% confidence space?

I have provided a working example below to illustrate the drawing of
the
polygons. As I said I would rather have three "ovals"/95% contours
drawn
around the points by "level" to capture the different density
distributions
without the visualization being heavily influenced by outliers.

I have looked into the code provided here from Hadley

https://groups.google.com/forum/?fromgroups=#!topic/ggplot2/85q4SQ9q3V8
using the mvtnorm package and the dmvnorm function, but haven't been
able
to get it work for my data example. The calculated densities are
always
zero (at this step of Hadley's code: dgrid$dens <-
dmvnorm(as.matrix(dgrid), ex_mu, ex_sigma)   )

I appreciate any assistance.

Thanks,
Nate

x<-c(seq(0.15,0.4,length.out=30),seq(0.2,0.6,length.out=30),
seq(0.4,0.6,length.out=30))


y<-c(0.55,x[1:29]+0.2*rnorm(29,0.4,0.3),x[31:60]*rnorm(30,0.3,0.1),x[61:90]*rnorm(30,0.4,0.25))
data<-data.frame(level=c(rep(1, 30),rep(2,30), rep(3,30)), x=x,y=y)


find_hull <- function(data) data[chull(data$x, data$y), ]
hulls <- ddply(data, .(level), find_hull)

fig1 <- ggplot(data=data, aes(x, y, colour=(factor(level)),
fill=level))+geom_point()
fig1 <- fig1 + geom_polygon(data=hulls, alpha=.2)
fig1

        [[alternative HTML version deleted]]