denstrip: Density strips

Description

The density strip illustrates a univariate distribution as a shaded rectangular strip, whose darkness at a point is proportional to the probability density. The strip is darkest at the maximum density and fades into the background at the minimum density. It may be used to generalise the common point-and-line drawing of a point and interval estimate, by representing the entire posterior or predictive distribution of the estimate. This function adds a density strip to an existing plot.

Usage

denstrip(x, dens, at, width, horiz=TRUE, colmax, colmin="white", scale=1, gamma=1, ticks=NULL, tlen=1.5, twd, tcol, mticks=NULL, mlen=1.5, mwd, mcol, lattice=FALSE, ...)
panel.denstrip(...)

Arguments

Either the vector of points at which the density is evaluated (if dens supplied), or a sample from the distribution (if dens not supplied).

dens

Density at x. If dens is not supplied, the density of the sample x is estimated by kernel density estimation, using density(x,...).

Position of the centre of the strip on the y-axis (if horiz=TRUE) or the x-axis (if horiz=FALSE).

width

Thickness of the strip, that is, the length of its shorter dimension. Defaults to 1/30 of the axis range.

horiz

Draw the strip horizontally (TRUE) or vertically (FALSE).

colmax

Colour at the maximum density, either as a built-in R colour name (one of colors()) or an RGB hex value. Defaults to par("fg") which is normally "black", or "#000000". Or in lattice, defaults to trellis.par.get("add.line")$col.

colmin

Colour to shade the minimum density, likewise. Defaults to "white". If this is set to "transparent", and the current graphics device supports transparency (see rgb), then overlapping strips will merge smoothly.

scale

Proportion of colmax to shade the maximum density, for example scale=0.5 with colmax="black" for a mid-grey colour.

gamma

Gamma correction to apply to the colour palette. The default of 1 should give an approximate perception of darkness proportional to density, but this may need to be adjusted for different displays. Values of gamma greater than 1 produce colours weighted towards the lighter end, and values of between 0 and 1 produce darker colours.

ticks

Vector of x-positions on the strip to draw tick marks, or NULL for no ticks.

tlen

Length of these tick marks relative to the strip width.

twd

Line thickness of these marks (defaults to par("lwd"), or in lattice, to

trellis.par.get("add.line")$lwd*2.).

tcol

Colour of the tick marks. Defaults to colmax.

mticks

x-position to draw a thicker tick mark or tick marks (for example, at the mean or median).

mlen

Length of this mark relative to the strip width.

mwd

Line thickness of this mark (defaults to par("lwd")*2, or in lattice, to trellis.par.get("add.line")$lwd*2.).

mcol

Colour of this mark. Defaults to colmax.

lattice

Set this to TRUE to make denstrip a lattice panel function instead of a base graphics function. panel.denstrip(x,...) is equivalent to denstrip(x, lattice=TRUE, ...).

...

Additional arguments supplied to density(x,...), if the density is being estimated. For example, bw to change the bandwidth of the kernel.

In other software

An add-on which enables the WinBUGS 1.4 software for Bayesian analysis to draw density strips is available from http://www.mrc-bsu.cam.ac.uk/personal/chris/papers/denstrip_wbpatch.txt. Open this file in WinBUGS and select Tools->Decode->Decode All. They will then be available via the Inference/Compare menu. In OpenBUGS (http://www.openbugs.info) density strips are available via the Inference/Compare menu. See this blog post: http://blogs.sas.com/content/graphicallyspeaking/2012/11/03/density-strip-plot/, for density strips in SAS.

References

Jackson, C. H. (2008) Displaying uncertainty with shading. The American Statistician, 62(4):340-347. http://www.mrc-bsu.cam.ac.uk/personal/chris/papers/denstrip.pdf

Examples

Run this code

## Illustrate a known standard normal distribution
## Various settings to change the look of the plot

x <- seq(-4, 4, length=10000)
dens <- dnorm(x)
plot(x, xlim=c(-5, 5), ylim=c(-5, 5), xlab="x", ylab="x", type="n")
denstrip(x, dens, at=0) # default width 
denstrip(x, dens, width=0.5, at=0) 
denstrip(x, dens, at=-4, ticks=c(-2, 0, 2)) 
denstrip(x, dens, at=-3, ticks=c(-2, 2), mticks=0) 
denstrip(x, dens, at=-2, ticks=c(-2, 2), mticks=0, mlen=3, 
         mwd=4, colmax="#55AABB") 
denstrip(x, dens, at=1, ticks=c(-2, 2), tlen=3, twd=3) 
denstrip(x, dens, at=-4, ticks=c(-2, 2), mticks=0, colmax="darkgreen", 
         horiz=FALSE)
x <- rnorm(1000) # Estimate the density
denstrip(x, width=0.2, at=-3, ticks=c(-2, 2), mticks=0, colmax="darkgreen",
         horiz=FALSE)
denstrip(x, at=2, width=0.5, gamma=2.2) 
denstrip(x, at=3, width=0.5, gamma=1/2.2) 

### Specifying colour of minimum density 
par(bg="lightyellow")
plot(x, xlim=c(-5, 5), ylim=c(-5, 5), xlab="x", ylab="x", type="n")
x <- seq(-4, 4, length=10000)
dens <- dnorm(x)
## Equivalent ways of drawing same distribution 
denstrip(x, dens, at=-1, ticks=c(-2, 2), mticks=0, colmax="darkmagenta")
denstrip(x, dens, at=-2, ticks=c(-2, 2), mticks=0, colmax="darkmagenta",
         colmin="lightyellow")
## ...though the next only works if graphics device supports transparency 
denstrip(x, dens, at=-3, ticks=c(-2, 2), mticks=0, colmax="darkmagenta",
         colmin="transparent")
denstrip(x, dens, at=-4, ticks=c(-2, 2), mticks=0, colmax="#8B008B", colmin="white")

## Alternative to density regions (\link{densregion.survfit}) for
## survival curves - a series of vertical density strips with no
## interpolation

library(survival)
fit <- survfit(Surv(time, status) ~ 1, data=aml, conf.type="log-log")
plot(fit, col=0)
lse <- (log(-log(fit$surv)) - log(-log(fit$upper)))/qnorm(0.975)
n <- length(fit$time)
lstrip <- fit$time - (fit$time-c(0,fit$time[1:(n-1)])) / 2
rstrip <- fit$time + (c(fit$time[2:n], fit$time[n])-fit$time) / 2
for (i in 1:n) { 
    y <- exp(-exp(qnorm(seq(0,1,length=1000)[-c(1,1000)], 
                        log(-log(fit$surv))[i], lse[i])))
    z <- dnorm(log(-log(y)), log(-log(fit$surv))[i], lse[i])
    denstrip(y, z, at=(lstrip[i]+rstrip[i])/2,
                 width=rstrip[i]-lstrip[i],
                 horiz=FALSE, colmax="darkred")
}
par(new=TRUE)
plot(fit, lwd=2)

## Use for lattice graphics (first example from help(xyplot))

library(lattice)
Depth <- equal.count(quakes$depth, number=8, overlap=.1)
xyplot(lat ~ long | Depth, data = quakes,
       panel = function(x, y) { 
           panel.xyplot(x, y)
           panel.denstrip(x, horiz=TRUE, at=-10, ticks=mean(x))
           panel.denstrip(y, horiz=FALSE, at=165, ticks=mean(y))
       } 
       )

## Lattice example data: heights of singing voice types

bwplot(voice.part ~ height, data=singer, xlab="Height (inches)",
       panel=panel.violin, xlim=c(50,80))
bwplot(voice.part ~ height, data=singer, xlab="Height (inches)",
       panel = function(x, y) {
           xlist <- split(x, factor(y))
           for (i in seq(along=xlist))
               panel.denstrip(x=xlist[[i]], at=i)
       },
       xlim=c(50,80)
       )

Run the code above in your browser using DataLab