fix alignment problem

[javlon]

Fix #4350

This PR fixes the alignment problem. The problem is fixed by creating normalized fake width.

[thomasp85]

@karawoo would you be able to take a look at this?

also, @javlon have you fixed the issue pointed out by @smouksassi in #4350?

[javlon]

also, @javlon have you fixed the issue pointed out by @smouksassi in #4350?

Sorry I haven't seen this issue. The only way to fix mentioned issue requires changes in the position_dodge algorithm. Changes include information about the width of the boxplot (or any other object) from position_dodge2. It means position_dodge should know information about other layers which might lead to more trouble in the future than the proposed solution.

Above PR include changes only in the position_dodge2 algorithm, and now it has the same algorithm for position selection as position_dodge.

Yes, the distance between boxplots has changed slightly, but the overall width for each group is constant and hasn't changed.

[karawoo]

This PR does fix the alignment of points and box plots in some cases, but I think the issue that @smouksassi mentioned about the distance between boxes is a problem. Bar charts really take a hit with this approach. Taking the example from the comments of #4350, currently they look like this:

library("tidyverse")

set.seed(123)

df <- bind_rows(lapply(1:7, function(i) {
  data.frame(
    x = i,
    y = rnorm(10*i),
    g = sample(LETTERS[seq_len(i)], 10*i, replace = T)
  )
}))

ggplot(df, aes(as.factor(x), fill = g)) +
  geom_bar(position = position_dodge2(preserve = "single"))

^{Created on 2022-08-23 with reprex v2.0.2}

With this PR the spacing changes quite a bit and doesn't look very good IMO:

And there are still other cases where box alignment won't work with points, e.g. when varwidth = TRUE

ggplot(df, aes(as.factor(x), y, fill = g)) +
  geom_boxplot(position = position_dodge2(0.75, preserve = "single"), varwidth = TRUE) +
  geom_point(position = position_dodge(0.75, preserve = "total"), shape = 4)+
  xlab("number of groups")

I don't think there's any way to address all the edge cases of alignment without a rewrite that lets layers know more about each other. And even though point/box alignment is a frequently raised issue, I'm not sure it's worth the cost of the spacing changes.

[karawoo]

I should say @javlon that I really appreciate you working on this and your solution does look really good for basic boxes + points. It's just a tricky problem with a lot of edge cases, unfortunately 😕

[javlon]

Thank you, @karawoo, for your clear explanation of the problem. I forgot about this kind of usage of boxplots.