geom_bar()/geom_col() should provide clearer warning when dropping data due to axis limits and column width

[mkoohafkan]

Consider the following plot:

library(ggplot2)

ggplot(mtcars, aes(x = carb)) + 
  geom_bar(stat = "count") +
  scale_x_continuous(limits = c(2, 4))

The bars centered at 2 and 4 are dropped. As documented in #2887, this is because the limits are not adjusted for the width of the bars. The bar centered on 2 actually extends back to ~1.5, and the bar centered on 4 actually extends to 4.5. Users do get some warnings:

Warning messages:
1: Removed 9 rows containing non-finite values (stat_count).
2: Removed 2 rows containing missing values (geom_bar).

The second warning alludes to the bars being dropped because they extend outside the specified axis limits. However, this isn't a sufficient explanation of what is happening because users may be specifying limits because they want bars outside of the axis limits to be dropped (and thus they expect to see the warnings), but don't understand why columns that appear to be inside the axis limits are also dropped.

suggestion: Improve the error message to note when columns are dropped because the column width is outside of the range (e.g. the column centered on 2), as opposed to when the entire column is outside of the range (e.g. the column centered on 1).

Finally, consider this solution to the issue:

ggplot(mtcars, aes(x = carb)) + 
  geom_bar(stat = "count") +
  # expand axis limits to accommodate column width
  scale_x_continuous(limits = c(1.5, 4.5))

Warning message:
Removed 9 rows containing non-finite values (stat_count).

Is this solution the preferred approach to resolving the issue, or is this relying on undocumented/unreliable behavior?

[yutannihilation]

I've filed a similar issue (#2887) before. I agree it's nice if we can show more clearer warning, which probably recommends to use coord_cartesian(xlim = , ylim = ).

[mkoohafkan]

I don't think the warning needs to even specify the solution---the unclear part is that columns are removed based on the column edges rather than the column centers. Adding something to the warning message like

Columns with width that extends beyond the specified axis limits were dropped.

Would cue users in to why it's happening, and would easily be able to identify the best solution for their use case (e.g. using coord_cartesian() or simply expanding their specified limits).

[mkoohafkan]

Also, here's an example of why I think the behavior is inconsistent, although it was discussed at length in #2887:

library(ggplot2)
 ggplot(mtcars) + aes(x =carb, y = mpg) + geom_point(size = 40) + xlim(c(3,4))

The points are clearly wide enough to extend past the axis limits, but they are not removed.

[yutannihilation]

Your example makes me think that we might need some custom grob that is represented in 2D (i.e. x, ymin, ymax and width)...

[teunbrand]

The warning is now clearer:

Removed 2 rows containing missing values or values outside the scale range (geom_bar()).

Throwing a more specific warning for geom_bar() is hard with the current setup. The scales who handle out-of-bounds values are agnostic of layers. Building a layer awareness in scales is not straightforward and I don't think it is worth it for just this one message.

A second option would be to always add this suggestions to the warning, regardless of whether out-of-bounds censoring happened. This might raise a bunch of false positives, so I think we shouldn't go this route either.

With the situation improved by a clearer message, no viable paths (that I can see) to further the specificity of the message and this issue being dormant for the past 4 years, I think we should close this issue, and consider reopening it when a reasonable path has been laid out.