small update to position-dodge2.r

[frostell]

Small update to position-dodge2.r to enable alignment of geom_boxplot() and geom_point() with NA's through position_dodge2(), possible patch for #2480.

[frostell]

Hmm, I realise reading the logs for the failed checks that my suggested update breaks functionality as defined in test-position-dodge2.r (it gives a different result for geom_col() and geom_rect() when there is x-overlap, but no grouping).

• changed one word in a new commit and I hope the tests will pass now!

[frostell]

I don't think the travis-ci check broke because of my change, it passes all your tests scripts @karawoo, what do you think?

[karawoo]

Thanks for looking into this @frostell. When I try your update I still get misaligned points and boxes on the left, does this match what you see?

# using example from #2480
set.seed(5820)
dat <- data.frame("value" = rnorm(n=30, mean=2, sd=0.5),
                  "group" = LETTERS[1:3],
                  "x" = factor(1:2))

dat$value[dat$group=="A"&dat$x=="1"] <- NA

ggplot(dat, aes(x=x, y=value)) +
  geom_boxplot(aes(fill=group), alpha=0.3) +
  geom_point(aes(colour=group), position=position_dodge2(width=0.75), size=3, alpha=0.5)
#> Warning: Removed 5 rows containing non-finite values (stat_boxplot).
#> Warning: Removed 5 rows containing missing values (geom_point).

[frostell]

Hello @karawoo!
Hmm, the exact same code gives the correct alignment on my installation...

Which version of ggplot are you using? I have 2.2.1.9000 installed, maybe that's the problem?
In your example, it looks like geom_boxplot is using position="dodge" and not position="dodge2".

I think you will get the correct alignment if you write:

ggplot(dat, aes(x=x, y=value)) +
  geom_boxplot(aes(fill=group), position="dodge2", alpha=0.3) +
  geom_point(aes(colour=group), position=position_dodge2(width=0.75), size=3, alpha=0.5)

Does that work for you?

The other possibility is that the default preserve argument differs, then you could try:

ggplot(dat, aes(x=x, y=value)) +
  geom_boxplot(aes(fill=group), position=position_dodge2(width=0.75, preserve = "single"), alpha=0.3) +
  geom_point(aes(colour=group), position=position_dodge2(width=0.75, preserve = "single"), size=3, alpha=0.5)

(or "total" instead of "single")

[karawoo]

Ah I see what's happened. I applied your changes onto the current master branch. position_dodge2() was updated in #2386 to make preserve = "total" the default. I realized in looking into this that I neglected to make one change in that PR that caused it to revert back to "single" for the points but not the boxes, hence the mismatch. This is addressed in #2545.

[hadley]

Can you start by describing the key problem with the existing algorithm and how your code fixes it? The current logic is not obvious to me.

[frostell]

With this small update pos.dodge2() works with geom_point too and it's possible to align boxplots and points even in trickier cases when some unique combinations are all NA's.

Background:
Displaying raw data points together with box plots can be useful. When using a fill aesthetic for the box plots, position.dodge2() gives great positioning options. Raw data can be plotted with geom_point() and position.dodge() for alignment.

Problem:
When one of the unique combinations of a category and fill/color is all NA's, position.dodge() and position.dodge2() gives different x-positions and the boxes and points end up misaligned. For geom_point, the current position.dodge2() does not work because it treats every unique point as a category that should be dodged from all the other points. See #2480 for reprex.

Solution:
Change pos.dodge2() so that:
- It evaluates what kind of data is fed into it by checking for existing column names
- If data is not a stat: save the full df and the reduce it to just the unique combinations of fill/color and x-category (the rest of pos.dodge2() runs like before and calculates appropriate x-positions)
- at the end, repopulate the full df with the x-positions calculated

[hadley]

Your logic doesn't seem quite general enough to me. I'll take a stab at it myself.

[frostell]

ok, I admit it was patchy even though it worked for the specific problem :)

I guess ideally there would be just one position_dodge function to handle all geoms and stats with appropriate options to choose styling of the "dodging"... Looking forward to see your solution!

[lock]

This old issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with reprex) and link to this issue. https://reprex.tidyverse.org/