Having recently discovered the R package denstrip, I was struck by intuitive it is to use shading to express uncertainty. If you need convincing, check out the paper by Jennifer Lundquist and Ken Lin linked below.

I wondered whether this would also work well for generic coefficient plots to examine regression results. The proof is in the pudding, so here’s some quickly thrown together code (also on Gist).

`library(denstrip)`

fade <- function(x, labels=names(coef(x)), expo=FALSE, xlab="", ylab="", bty="n", ...) {

# argument: a regression, additional arguments passed on to plot() and text()

coe <- summary(x)$coefficients[,1] # extract coefficients

cse <- summary(x)$coefficients[,2] # standard errors

len <- length(coe) # how many coefficients (without intercept)

if(expo == TRUE) { # exponential form

coe <- exp(coe)

cse <- exp(cse)

}

ran <- c(min(coe)- 3*max(cse), max(coe)+3*max(cse)) # range of values

plot(0, ylim=c(1.5,len+0.5), xlim=ran, xlab=xlab, ylab=ylab, bty=bty, type="n", ..., axes=FALSE) # empty plot

# title, xlab, ylab, etc. can be used here.

for(i in 2:len) {

dens <- rnorm(mean=coe[i], sd=cse[i], n=1000)

denstrip(dens, at=i) # passing arguments conflicts with plot()

text(ran[1],i, labels[i], adj = c(0,0), ...) # adj to left-align

# cex, col can be used here

}

axis(1)

}

Well, it does work (see figure above). If we add an `abline(v=0, lty=3)`

, we can easily highlight the zero line.

However, intuitive as it is, I’m not really convinced. Here’s the coefficient plot the arm package produces. While it doesn’t use shading, lines of different thickness are used to indicate the standard errors — nicely de-emphasizing the end-point. There’s more emphasis on the point estimate (i.e. the coefficient itself: the square), while my code hides this. I did try adding a white line at the point estimate (third figure, below), but this doesn’t resolve my biggest worry: the amount of ink that is used to convey the message… I’m just not convinced that the shading adds that much to the relatively simple coefficient plots in the package arm.

(There’s an argument `width`

that could be added to line calling `denstrip`

, like `width=0.2`

, but it hasn’t convinced me enough.)

Jackson, Christopher H. 2008. “Displaying Uncertainty with Shading.” *The American Statistician* 62 (4): 340–47. doi:10.1198/000313008X370843.

Lundquist, Jennifer H., and Ken-Hou Lin. 2015. “Is Love (Color) Blind? The Economy of Race among Gay and Straight Daters.” *Social Forces*, March, sov008. doi:10.1093/sf/sov008.

## One Reply to “Using Shading in Coefficient Plots”