It is often helpful to plot a line (e.g., a fitted regression line or a non-parametric LOWESS line) over data. Likewise, when variables are confounded, it is often helpful to plot x1, y data stratified by x2. However, I can't quite figure out how to combine both using ?coplot. This ought to be quite natural in lattice, but I can't seem to figure out how to do it there, either.
The following provides some data to play with:
library(MASS)
set.seed(7422)
X <- mvrnorm(100, mu=c(5,5), Sigma=rbind(c(2.0, 1.9),
c(1.9, 2.0) ) )
x1 <- X[,1]; x2 <- X[,2]
y <- 3 + .6*x1 - .4*x2 + rnorm(100)
m <- lm(y~x1+x2)
summary(m)$coefficients
# Estimate Std. Error t value Pr(>|t|)
# (Intercept) 3.8281801 0.4022204 9.517619 1.476519e-15
# x1 0.5466495 0.2060269 2.653292 9.314673e-03
# x2 -0.4835804 0.2139542 -2.260205 2.604451e-02
windows()
pairs(cbind(y, x1, x2))
xs <- seq(1.6, 9.2, by=.1)
windows()
plot(y~x2)
lines(xs, predict(m, data.frame(x1=5.2, x2=xs)), col="blue")
windows()
coplot(y~x2|x1)
## doesn't work (lines on plot were drawn by hand)
# lines(xs, predict(m, data.frame(x1=mean(x1[subscripts], x2=xs)), col="blue")
# lines(lowess(y~x2, subset=x1[subscripts]), col="gray")
Scatterplot matrix of the raw data:
Here is a marginal plot of x2
vs y
, with fitted the regression line and a marginal LOWESS line plotted on top of the data. This will look weird to people who aren't statistically savvy.
This is what I have in mind. It's a coplot, with the regression model and a LOWESS line fitted to each stratum of the data plotted inside each panel.