geom_smooth custom linear model

Question

While looking at this issue, I couldn't specify a custom linear model to geom_smooth. My code is as follows:

example.label <- c("A","A","A","A","A","B","B","B","B","B")
example.value <- c(5, 4, 4, 5, 3, 8, 9, 11, 10, 9)
example.age <- c(30, 40, 50, 60, 70, 30, 40, 50, 60, 70)
example.score <- c(90,95,89,91,85,83,88,94,83,90)
example.data <- data.frame(example.label, example.value,example.age,example.score)

p = ggplot(example.data, aes(x=example.age,
                         y=example.value,color=example.label)) +
  geom_point()
  #geom_smooth(method = lm)

cf = function(dt){
  lm(example.value ~example.age+example.score, data = dt)
}

cf(example.data)

p_smooth <- by(example.data, example.data$example.label, 
               function(x) geom_smooth(data=x, method = lm, formula = cf(x)))

p + p_smooth 

I am getting this error/warning:

Warning messages:
1: Computation failed in `stat_smooth()`:
object 'weight' not found 
2: Computation failed in `stat_smooth()`:
object 'weight' not found 

Why am I getting this? And what is the proper method of specifying a custom model to geom_smooth. Thanks.

Answer 1

The regression function for a regression model with two continuous predictor variables and a continuous outcome lives in a 3D space (two for the predictors, one for the outcome), while the ggplot graph is a 2D space (one continuous predictor on the x-axis and the outcome on the y-axis). That's the fundamental reason why you can't plot a function of two continuous predictor variables with geom_smooth.

One "workaround" is to pick a few specific values of one of the continuous predictor variables and then plot a line for the other continuous predictor on the x-axis for each of the chosen values of the first variable.

Here's an example with the mtcars data frame. The regression model below predicts mpg using wt and hp. We then plot the predictions of mpg vs. wt for various values of hp. We create a data frame of predictions and then plot using geom_line. Each line in the graph represents the regression prediction for mpg vs. wt for different values of hp. You can, of course, also reverse the roles of wt and hp.

library(ggplot)
theme_set(theme_classic())

d = mtcars
m2 = lm(mpg ~ wt + hp, data=d)

pred.data = expand.grid(wt = seq(min(d$wt), max(d$wt), length=20),
                        hp = quantile(d$hp))
pred.data$mpg = predict(m2, newdata=pred.data)

ggplot(pred.data, aes(wt, mpg, colour=factor(hp))) +
  geom_line() +
  labs(colour="HP Quantiles")

Another option is to use a colour gradient to represent mpg (the outcome) and plot wt and hp on the x and y axes:

pred.data = expand.grid(wt = seq(min(d$wt), max(d$wt), length=100),
                        hp = seq(min(d$hp), max(d$hp), length=100))
pred.data$mpg = predict(m2, newdata=pred.data)

ggplot(pred.data, aes(wt, hp, z=mpg, fill=mpg)) +
  geom_tile() +
  scale_fill_gradient2(low="red", mid="yellow", high="blue", midpoint=median(pred.data$mpg))