Possible bug in `stage()`/`after_stat()` with scale transformations.

[teunbrand]

I think I might have found a bug with either the scales or the stage()/after_stat() functions in combination with dates.
I came across this when making a custom stat where I wanted to put default_aes = aes(xmin = after_stat(start), xmax = after_stat(end)) in the ggproto methods, where start and end are computed variables.

What happens is that when I try to apply a stage() or after_stat() to a date variable, it gives me an error that the date transformation is invalid. I don't think it is a problem with stats, as it also shows up with stat_identity().
The attempted plot below is not very interesting, but it is the most minimal example I could come up with.

library(ggplot2)

ggplot(economics) +
  geom_point(aes(x = stage(date, after_stat = x), 
                 y = unemploy))
#> Error: Invalid input: date_trans works with objects of class Date only

The error above doesn't happen for integer or numeric data.

ggplot(economics) +
  geom_point(aes(x = stage(as.integer(date), after_stat = x), 
                 y = unemploy))

And also the error can be rescued by formatting the output as a date again, but this seems overly verbose and seems difficult to anticipate in a custom stat.

ggplot(economics) +
  geom_point(aes(x = stage(date, after_stat = structure(x, class = "Date")), 
                 y = unemploy))

The same is true for other geoms and aesthetics as well. Here are similar examples with after_stat() instead of stage().

ggplot(economics) +
  geom_segment(aes(x = date, xend = after_stat(x + 30),
                   y = unemploy, yend = unemploy))
#> Error: Invalid input: date_trans works with objects of class Date only

ggplot(economics) +
  geom_segment(aes(x = as.integer(date), xend = after_stat(x + 30),
                   y = unemploy, yend = unemploy))

ggplot(economics) +
  geom_segment(aes(x = date, xend = after_stat(structure(x + 30, class = "Date")),
                   y = unemploy, yend = unemploy))

^{Created on 2020-07-31 by the reprex package (v0.3.0)}

[yutannihilation]

I'm yet to figure out what's happening, but probably we need to somehow bypass this transformation in the end of map_statistic() because the non-stat value should be already transformed.

ggplot2/R/layer.r

Lines 326 to 330 in 3be0acc

	# Transform the values, if the scale say it's ok
	# (see stat_spoke for one exception)
	if (self$stat$retransform) {
	stat_data <- scales_transform_df(plot$scales, stat_data)
	}

Actually, abusing retransform param makes this work (this parameter seems meaningless as it is always TRUE in the current version, btw). I guess you are facing more complex case than this example. Could you try retransform <- FALSE works in your case?

library(ggplot2)

l <- geom_point(aes(x = stage(date, after_stat = x), 
                    y = unemploy))
l$stat$retransform <- FALSE

ggplot(economics) + l

^{Created on 2020-08-14 by the reprex package (v0.3.0)}

[teunbrand]

Yes, the retransform trick totally works in my case, see example below. Note that the after_stat() occurs in the default_aes ggproto member and not in the layer mapping argument.

I never noticed the retransform setting, but it seems useful to my problem. Bit selfish/off-topic question maybe, but would there be any downsides to setting retransform = FALSE at the ggproto level?

I still have the gut feeling that the after_stat() + date class should work by default though, but I guess it is not a common use case. Another idea might also be to adjust the scales::date_trans()$transform() method to accept any numeric input, but that might cause more problems than it fixes.

library(ggh4x) # https://github.com/teunbrand/ggh4x
#> Loading required package: ggplot2

p <- ggplot(economics, aes(date)) +
  stat_rle(
    aes(label = date > as.Date("1980-01-01") & date < as.Date("1990-01-01")),
    align = "center"
  ) +
  geom_line(aes(y = unemploy))

# Gives error
print(p)
#> Error: Invalid input: date_trans works with objects of class Date only

p$layers[[1]]$stat$retransform <- FALSE

# Output is fine
print(p)

^{Created on 2020-08-14 by the reprex package (v0.3.0)}

[yutannihilation]

I still have the gut feeling that the after_stat() + date class should work by default though

Yes, I totally agree with you! Sorry probably my comment was confusing. I meant

1. The staged values should be bypass the transformation as probably it is already transformed (I'm not 100% sure about this yet, though...).
2. Though we need to introduce a proper fix here, retransform, which seems an unused parameter, is a quick and dirty way to test if bypassing works. I didn't mean this is a workaround, but just wanted to see there's more thing to consider in more complex cases.

would there be any downsides to setting retransform = FALSE at the ggproto level?

I bet there's no downsides. It appears only the following two places:

ggplot2/R/stat-.r

Lines 52 to 56 in 112f960

	Stat <- ggproto("Stat",
	# Should the values produced by the statistic also be transformed
	# in the second pass when recently added statistics are trained to
	# the scales
	retransform = TRUE,

ggplot2/R/layer.r

Lines 328 to 330 in 3aa2937

	if (self$stat$retransform) {
	stat_data <- scales_transform_df(plot$scales, stat_data)
	}

[yutannihilation]

Another example why I think this should be fixed. The value becomes 4 because it gets sqrt-ed twice.

library(ggplot2)

d <- data.frame(value = 16)

ggplot(d) +
  geom_point(aes(value, 0), colour = "red", size = 10) +
  geom_point(aes(stage(value, after_stat = x), 0), colour = "black", size = 10) +
  scale_x_sqrt(limits = c(0, 16), breaks = c(0, 4, 16))

[teunbrand]

Ah yes that result indeed doesn't make sense, so the bug applies to non-date transformations too. I should probably generalise the title of the issue for any transformation. As an aside, shouldn't the 0-break/label have been rendered too?

[yutannihilation]

As an aside, shouldn't the 0-break/label have been rendered too?

Good catch, thanks! I've filed another issue here: #4193