scale_*_continuous trans and limits combination leading to blank plot

[igordot]

I noticed an odd behavior for scale_*_continuous limits parameter. I think it would be easiest to show an example rather than try to explain.

Set up the plot.

library("ggplot2")
df <- data.frame(trt = c("a", "b", "c"), outcome = c(2.3, 1.9, 3.2))
p <- ggplot(df, aes(trt, outcome)) + geom_col()

Start adjusting scale_y_continuous().

p + scale_y_continuous(trans = "log", limits = c(1, 5))

p + scale_y_continuous(trans = "log1p", limits = c(1, 5))

p + scale_y_continuous(trans = "log1p", limits = c(0, 5))

The first and third plots are expected, but why is the middle empty? There is no error or warning.

I am using R 3.4.4 with ggplot2 2.2.1.

[clauswilke]

Bars are not shown unless they fit entirely within the specified limits. This is expected behavior. However, I agree that there probably should be a warning.

Simpler reprex (the transformation is irrelevant):

library(ggplot2)
df <- data.frame(trt = c("a", "b", "c"), outcome = c(2.3, 1.9, 3.2))
ggplot(df, aes(trt, outcome)) + geom_col() + scale_y_continuous(limits = c(1, 4))

Created on 2018-06-25 by the reprex package (v0.2.0).

[igordot]

@clauswilke Isn't the transformation still relevant? My first two examples have the same limits, but only one is visible.

[clauswilke]

I think the question is why are bars shown in the first plot, not why are they absent in the second. log(0)=-Inf, so the bars should extend to minus infinity (and hence outside the set limits).

[igordot]

@clauswilke Good point. I guess what I really meant is there is a discrepancy between the first and second.

Regardless, there should probably be a warning if the data is outside the limits on either end. There is one if the upper bound is lowered:

> p + scale_y_continuous(trans = "log1p", limits = c(0, 5))
> p + scale_y_continuous(trans = "log1p", limits = c(0, 1.5))
Warning message:
Removed 3 rows containing missing values (position_stack). 

[clauswilke]

I understand what is happening now. First the data are transformed using the transformation provided to the scale, and then geom_col() draws a column from 0 in the transformed space to whatever value is given in the transformed space. This explains why the bars start at 1 in the case of the log transform and at 0 in the case of the log1p transform, and this behavior is correct. (In log space, bars should start at 1.)

The only remaining question is where in the code bars are excluded when they don't fit entirely within the scale limits and how we could produce a warning for this case. Maybe @hadley can point me in the right direction here.

[hadley]

I have a vague recollection that we've tried to make this an error or warning in the past, and the obvious thing doesn't work. That's probably not much help 😞

[clauswilke]

Yes, the problem is that warnings are generated early in the processing pipeline and here the zero value is injected at a later stage. I'll see if I can come up with something.

[clauswilke]

I've figured it out. The issue is that the required_aes values for GeomCol() (and GeomBar(), which has the same issue) are different from those of GeomRect(), but GeomRect() is used for drawing. Therefore, the coordinates created by GeomCol() are ignored when missing values are evaluated. The fix is to define non_missing_aes appropriately. Will do a PR momentarily.

library("ggplot2")
df <- data.frame(trt = c("a", "b", "c"), outcome = c(2.3, 1.9, 3.2))
p <- ggplot(df, aes(trt, outcome)) + geom_col()

p + scale_y_continuous(trans = "log1p", limits = c(1, 5))
#> Warning: Removed 3 rows containing missing values (geom_col).

Created on 2018-06-26 by the reprex package (v0.2.0).

[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/