Suggested improvments to `plot_*()` functions

[mattansb]

Some thoughts for improving the current default behavior of the plot_*() functions' construction of ggplots. I would be happy to make a PR for any of these you find appropriate.

1. Use default aes()

Currently plot_build() specifies the aesthetics for each layer separately instead of setting them globally for the plot. This means that instead of writing something like:

plot_predictions(mod, condition = c("wt", "hp")) +
  geom_point(aes(y = mpg), data = mtcars)

We need to write out

plot_predictions(mod, condition = c("wt", "hp")) +
  geom_point(aes(wt, mpg, color = hp), data = mtcars)

# or:
plot_predictions(mod, condition = c("wt", "hp")) +
  aes(wt, color = hp, fill = hp) +
  geom_point(aes(y = mpg), data = mtcars)

2. Options for plotting numeric second condition / by

Current behavior: if the second argument to condition / by is a numeric variable, it is converted into a factor when passed to aes(). Additionally, for plot_predictions(points = <not 0>) the points aren't colored.

This means that when building on such plots with the original numeric variable we get an error:

library(marginaleffects)
library(ggplot2)

mod <- lm(mpg ~ wt * hp, data = mtcars)

plot_predictions(mod, condition = c("wt", "hp")) +
  geom_point(aes(wt, mpg, color = hp), data = mtcars)
#> Error in `scale_colour_discrete()`:
#> ! Continuous value supplied to a discrete scale.
#> ℹ Example values: 110, 93, 175, 105, and 245.

^{Created on 2026-01-12 with reprex v2.1.1}

I suggest two changes:

1. Change default behavior to convert numeric second condition / by variable to an ordered factor.
2. Allow user to ask for the variable to stay numeric.

3. autolayer() functions?

We can also go in the opposite direction - instead of adding ggplot elements to plot_*() outputs, we can add marginaleffects-derived layers to existing ggplots with some clever autolayer() methods?

[vincentarelbundock]

Thanks for opening this!

1. Use default aes()

I'm open to this. I feel like this was close to the original implementation but that I changed it to fix a problem at some point. Unfortunately, I can't remember the details and can't find a relevant git commit message. I might be inventing this...

If you can make this work and all the tests pass (without changing the test suite at all), I suppose we should be able to merge without too much risk.

2. Options for plotting numeric second condition / by

plot_predictions(mod, condition = c("wt", "hp")) +
geom_point(aes(wt, mpg, color = hp), data = mtcars)
#> Error in scale_colour_discrete():
#> ! Continuous value supplied to a discrete scale.
#> ℹ Example values: 110, 93, 175, 105, and 245.

You can try, but I don't think that ordered=TRUE will fix this particular problem. If you try and it works, I'll be happy to consider a merge.

A cleaner solution might be to avoid converting variables to factors altogether. This feels like a terrible hack anyway, and something like this might be cleaner:

ggplot(mtcars, aes(hp, mpg, color = factor(cyl))) + geom_point()

That makes the legend less pretty, but perhaps we can auto-rename it.

Allow user to ask for the variable to stay numeric.

I do not want to add a new argument or option for this. Let's find a solution that does not add clutter to the user interface.

3. autolayer() functions?

I am not very familiar with this and would prefer to avoid learning it unless there's no "clean" way to do this with simpler ggplot2 calls.