Need position_stack_line()

[mikmart]

geom_area() doesn't play well with duplicated x values. The docs state:

geom_area is a special case of geom_ribbon, where the ymin is fixed to 0.

But when there are multiple values of y for a given value of x, the two are not equivalent. The geom_ribbon() result matches the intuition of "geom_line() but with the area under the curve filled in":

library(tidyverse)
df <- data.frame(x = c(1, 2, 2, 3), y = 1:4)

ggplot(df, aes(x, y)) + geom_ribbon(aes(ymin = 0, ymax = y))

However, it seems this scenario results in geom_area() having the equivalent of ymin = min(y) and ymax = sum(y) in geom_ribbon().

ggplot(df, aes(x, y)) + geom_area()

I'm happy to dig a bit further and see if I can submit a PR for this.

Created on 2018-09-06 by the reprex package (v0.2.0.9000).

[smouksassi]

This is because (as per the docs) the default position for geom_area is "stack" if you specify "identity" you get the same results.

df <- data.frame(x = c(1, 2, 2, 3), y = 1:4)
ggplot(df, aes(x, y)) +
geom_point(color="red")+
  geom_area(alpha=0.5,aes(fill="geom_area") ,
            position = "identity")+
  geom_ribbon(alpha=0.5,aes(ymin = 0, ymax = y,fill="geom_ribbon"))

[mikmart]

That's true, and perhaps points to a potential issue with position_stack() instead:

library(tidyverse)

df <- data.frame(x = c(1, 2, 2, 3), y = 1:4)
df2 <- bind_rows(a = df, b = df, .id = "g")

ggplot(df2, aes(x, y, fill = g)) + geom_area()

ggplot(df2, aes(x, y, fill = g)) +
  geom_ribbon(aes(ymin = 0, ymax = y), position = "stack")

Created on 2018-09-06 by the reprex package (v0.2.0.9000).

[mikmart]

Then again, position_stack() is just doing what it does, stacking y values with a common x.. i.e. the above would look entirely reasonable with e.g. points. But this doesn't really match the intuition (or mine, at least) for area charts.

[smouksassi]

what kind of results you are expecting the above are identical

[mikmart]

Yes, that's why this is probably an issue with position_stack(). (If it is an issue at all.)

I would expect to not have the gaps in between the filled in areas.

[mikmart]

Essentially I would want to get this:

library(tidyverse)

df3 <- data.frame(
  x = c(1, 2, 2, 3, 1, 2, 2, 3),
  y = c(1, 2, 3, 4, 2, 4, 6, 8),
  g = rep(c("b", "a"), each = 4)
)

ggplot(df3, aes(x, y, fill = g)) +
    geom_area(position = "identity")

This matches my intuition of area stacking, but doesn't really fit all that well with how position_stack() works. Also, this is not really compatible with how e.g. point stacking should (and does) work..?

Created on 2018-09-06 by the reprex package (v0.2.0.9000).

[mikmart]

The more I think about this, the more it feels like perhaps this is not an issue with position_stack(), but rather area stacking in the above manner would be the job for a different position adjustment.

When presenting data like this, only the maximum of the y values in an "x group" is really a data point: the minimum y is really just a bounding point for a polygon, and doesn't actually represent data (okay it can be data, too: the minimum and maximum together define a range that is the data); and any other y values in between don't do anything.

Perhaps some sort of position_stack_range() might be useful? But I assume out of scope for ggplot2.

[yutannihilation]

I too feel geom_area() seems against the "intuition." I think this feeling comes from the fact that, while we imagine we are stacking areas, we actually are playing with points. For example, if we add a slight jitter to x axis, they are not stacked as expected. In the following, you can see "a" sinks into "b" because it fails to be stacked.

library(tidyverse)

df <- data.frame(x = rep(1:5, each = 2),
                 y = rep(10 + 1:5, each = 2),
                 g = rep(c("a", "b"), times = 5))
df$x_jitter <- df$x + 10e-4 * runif(10)
df
#>    x  y g x_jitter
#> 1  1 11 a 1.000296
#> 2  1 11 b 1.000814
#> 3  2 12 a 2.000061
#> 4  2 12 b 2.000527
#> 5  3 13 a 3.000078
#> 6  3 13 b 3.000748
#> 7  4 14 a 4.000230
#> 8  4 14 b 4.000891
#> 9  5 15 a 5.000595
#> 10 5 15 b 5.000664

library(egg)
#> Loading required package: gridExtra
#> 
#> Attaching package: 'gridExtra'
#> The following object is masked from 'package:dplyr':
#> 
#>     combine

ggarrange(
  ggplot(df, aes(x, y, fill = g)) + geom_area(),
  ggplot(df, aes(x_jitter, y, fill = g)) + geom_area()
)

Created on 2018-09-06 by the reprex package (v0.2.0.9000).

[mikmart]

@yutannihilation yeah I think you're right. Both examples make perfect sense with points, bars, etc. but don't quite "work" with areas (even though when reasoning about it, the behaviour is logical).

I read some other issues related to this before posting (#2720, #2802; also just now noticed that this is a duplicate from way back in 2013: #795) and thought this was separate; perhaps not.

Maybe it would make sense to have a position adjustment (somewhere) that would handle all of these problems, sort of simulating physical area stacking. That would require automatic interpolation between groups, too, to address the misaligned x values.

[smouksassi]

yes stacking time series can be done in several ways for example:
there is : https://github.com/AtherEnergy/ggTimeSeries
and this paper.
http://leebyron.com/streamgraph/stackedgraphs_byron_wattenberg.pdf
it may be a special stat on how to handle duplicate x

[mikmart]

The Byron & Wattenberg paper is really cool, but I'm not sure it applies to duplicated y's within x's, or "physical" area stacking, though. It's essentially a stacked position adjustment with a general baseline function rather than a fixed 0 baseline + smart ordering.

[ptoche]

@mikmart, is this a new issue or was introduced by an update? I ask out of curiosity because when I filed 2803, I noticed the stacking behaviour had changed over the previous approx 2 years.

[mikmart]

I'm not sure, but I would assume it's been like this for quite a while. The way the stacking algorithm works (replacing y values that share an x value with their cumulative sum) seems pretty fundamental.

[mikmart]

I had a look and I think what caused the change in #2803 was cf716a3, from which onward positive and negative values have been stacked separately. These seem separate to me.

[yutannihilation]

I've played around this and now feel there's a lot to consider to stack areas properly. I'm afraid this is a bit too complex to implement in ggplot2...

https://gist.github.com/yutannihilation/2d3851adc874a02f42914f1655329c71