geom_abline 3.3.4 odd y limits

[arnaud-feldmann]

Hi,

I am aware of #4024 and #3741 that modify geom_abline. Through, the abline modification in 3.3.4 has had some strange results for the default y limits, maintening my modest package disaggR and in particular its autoplot method for in_scatter. I can't decide if it's a feature or a bug, I'm sorry if I make a mistake to report that.

object <- in_scatter(twoStepsBenchmark(turnover,construction))
n <- nrow(object)
df <- data.frame(Time=as.numeric(time(object))[-1L],
                 `High-frequency serie` = object[-n,2L],
                 `Low-frequency serie` = object[-n,1L],
                 xend = object[-1L,2L],
                 yend = object[-1L,1L], check.names = FALSE)

ggplot() +
  geom_abline(intercept = attr(object,"coefficients")["constant"],
                       slope = attr(object,"coefficients")[names(attr(object,"coefficients")) != "constant"],
                       lty = "solid", colour = "red", size = 1)  +
  geom_segment(data=df,
               aes(x = `High-frequency serie`, y = `Low-frequency serie`,
                   xend = xend, yend = yend,
                   colour = Time),
               arrow=arrow(angle = 15,
                           ends = "last",
                           type = "closed",
                           length = unit(0.1,"inches")),
               na.rm = TRUE)

On 3.3.4 The default y limits seems to be very odd to me

on 3.3.3 the behavior was as expected

[clauswilke]

Could you please provide a reproducible example that depends on nothing but base R and ggplot2? And please simplify the example so all extraneous information is removed. This will make it easier to reason about the issue. Also, if you can, please use geom_point() instead of geom_segment(). However, if the problem only shows up for geom_segment() that would be good to know also.

[arnaud-feldmann]

Sorry for the bother,

@clauswilke Here is a very simpler example. Is this new default y limit intended ? Should we adapt our code ?

Thanks a lot for your work.

Arnaud

library(ggplot2)
df <- data.frame(X = c(670.964799810616, 738.482178701357, 774.152678327829,
                       816.022561905182, 877.503579439891, 933.507292448021),
                 Y = c(138.4, 143.9, 147.3, 155.9, 169.3, 181.4))
ggplot() +
  geom_abline(intercept = 44.28163,
              slope = 0.1410183)  +
  geom_point(data=df,
             aes(x = X, y = Y))

ggplot2 3.3.3 :

ggplot2 3.3.4 :

[clauswilke]

Looks like geom_abline() extends the line all the way to x = 0, no matter how far that is from the data. That definitely doesn't look like intended behavior to me.

library(ggplot2)
library(patchwork)

df <- data.frame(
  x = c(10000, 10001, 10002),
  y = c(500, 502, 501)
)

p <- ggplot(df) +
  geom_abline(
    intercept = -9500,
    slope = 1
  )  +
  geom_point(aes(x, y))

p | p + xlim(0, 10010)

^{Created on 2021-06-16 by the reprex package (v1.0.0)}

[ptoche]

Is it related to this? where the range of the line was extended to fix another issue:

e127cc9

[arnaud-feldmann]

after having tried with and without, it is actually linked to that commit 041e287

Seems like the intercept being told that way as a known aesthetic for y, it behaves as @clauswilke told

[clauswilke]

Yes, in hindsight, this was a problematic change. We want the intercept to get transformed but we don't want it to enter the limit calculations. We may need to separate out those two functions, or let the geoms/stats declare which aesthetics should be used for each of these calculations. The latter approach might be cleaner but is more work.

[thomasp85]

I'm leaning on actually rolling this change back and doing a quick patch release.. thoughts, @clauswilke and @yutannihilation ?

[clauswilke]

I agree. The newly introduced bug is more problematic than the bug that was fixed, in my opinion.

[andrefaure]

I am also having trouble with geom_abline 3.3.4, noticed this issue and think it could be related?
https://stackoverflow.com/questions/68068628/gglot2geom-abline-fails-with-default-settings-in-log-log-plot-ggplot2-v3-3-4

[yutannihilation]

It sounds good to me to roll the change back. Also, I think the fix for the ggsave() issues (i.e., #4524 and #4521) might be worth including in the patch release.

[thomasp85]

Agreed on the ggsave issues — was planning on including that as well. Will work on this today