Needle plots

[ggrothendieck]

It seems there is no nice way to create needle plots in ggplot2 whereas it is very easy in classic graphics. geom_linerange can work but seems too involved and non-obvious. A geom_needle wrapper around geom_linerange would be nice.

x <- 1:6
y <- x * c(1, -1)

plot(y ~ x, type = "h")

ggplot(data.frame(x, y), aes(x = x, ymax=pmax(0, y), ymin = pmin(0, y))) + 
  geom_linerange()

[clauswilke]

Reprex with output:

library(ggplot2)
x <- 1:6
y <- x * c(1, -1)

plot(y ~ x, type = "h")

ggplot(data.frame(x, y), aes(x = x, ymax=pmax(0, y), ymin = pmin(0, y))) + 
  geom_linerange()

^{Created on 2019-11-16 by the reprex package (v0.3.0)}

[clauswilke]

How exactly is a needle plot defined? Would geom_col() fit the bill?

library(ggplot2)
x <- 1:6
y <- x * c(1, -1)

ggplot(data.frame(x, y), aes(x = x, y = y)) + 
  geom_col()

ggplot(data.frame(x, y), aes(x = x, y = y)) + 
  geom_col(width = 0.05)

^{Created on 2019-11-16 by the reprex package (v0.3.0)}

[ggrothendieck]

It should act identically to type = "h" and should be as simple to use as in classic graphics and lattice.

I don't really think geom_linerange and geom_col are sufficient. geom_linerange can implement it exactly but is too complex and geom_col requires that you specify the width which seems problematic if you are generating reports automatically and don't know the range of the data. Neither seem obvious.

Also, it seems that the idea of ggplot2 is that there is a geom for each type so type = "p" corresponds to geom_point, type = "l" corresponds to geom_line and so on so it would seem to be most consistent to have a geom for each additional type not already directly supported with one.

[ggrothendieck]

By the way, I did some further googling and there are several questions on SO about this in relation to ggplot2 so it seems a number of people have found this confusing. One answer came up with geom_linerange but they implemented it only partially assuming non-negative numbers.

https://stackoverflow.com/questions/26139878/needle-plot-in-ggplot2
https://stackoverflow.com/questions/25952270/code-for-type-h-in-ggplot2

Also with regards to lattice in addition to type = "h" the caret package defines panel.needle which is like type = "h" except it is rotated 90 degrees and there is a dot at the end of each needle.
https://rdrr.io/cran/caret/man/panel.needle.html

For example, try

library(caret)
library(lattice)
xyplot(demand ~ Time, BOD, panel = panel.needle)

Finally it should be pointed out that type = "h" plots are widely used, particularly in economics and finance. Stock price graphs typically have a type = "h"chart at the bottom of the graph showing the volume.

[thomasp85]

ggalt::geom_lollipop() seems to fit the bill. If you don’t want the terminal point you can just set size to zero

[yutannihilation]

geom_col requires that you specify the width which seems problematic if you are generating reports automatically and don't know the range of the data.

If width is the only problem, you can set it to zero. But, I agree it's nice if we can draw this without this kinds of workaround. I'm not yet sure if ggplot2 should have the dedicated geom for this, though....

library(ggplot2)
x <- 1:6
y <- x * c(1, -1)

ggplot(data.frame(x, y), aes(x = x, y = y)) + 
  geom_col(width = 0, colour = "black")

^{Created on 2019-11-18 by the reprex package (v0.3.0)}

[clauswilke]

We could implement a geom_needle() that works just like geom_col() but uses line segments instead of rectangles for drawing. The use of geom_col(width = 0) may have unintended consequences, since it draws two lines (one up, one down) per column. Also, it doesn't apply the alpha setting to the lines.

[ggrothendieck]

Regarding the idea of using geom_col(width = 0) it seems that lwd does work with it so one has a range independent way of making it wider, if needed althoug perhaps less than obvious. At any rate it seems the basic functionality is there with geom_col or geom_linerange and it just needs to be wrapped nicely to make it easier or more obvious to use.

library(ggplot2)
x <- 1:6
y <- x * c(1, -1)

ggplot(data.frame(x, y), aes(x = x, y = y)) + 
  geom_col(width = 0, colour = "black", lwd = 3)

[yutannihilation]

uses line segments instead of rectangles for drawing.

Agreed. I guess a naive implementation would be like this:

library(ggplot2)

x <- 1:6
y <- x * c(1, -1)

GeomNeedle <- ggproto("GeomNeedle", GeomBar,
  # derived from GeomLinerange
  default_aes = aes(colour = "black", size = 0.5, linetype = 1, alpha = NA),
  
  draw_panel = function(self, data, panel_params, coord, width = NULL, flipped_aes = FALSE) {
    GeomLinerange$draw_panel(data, panel_params, coord)
  }
)

geom_needle <- function(...) {
  l <- geom_col(...)
  l$geom <- GeomNeedle
  l
}

ggplot(data.frame(x, y), aes(x = x, y = y)) + 
  geom_needle(alpha = 0.5)

^{Created on 2019-11-18 by the reprex package (v0.3.0)}

I think this is fine. What I was wondering is that if what GeomBar's setup_data() does can be generalized as something like StatDistanceFromZero so that we can use geom_linerange(stat = "distance_from_zero") instead of implementing a new one...

ggplot2/R/geom-bar.r

Lines 136 to 146 in cb0eaa0

	setup_data = function(data, params) {
	data$flipped_aes <- params$flipped_aes
	data <- flip_data(data, params$flipped_aes)
	data$width <- data$width %||%
	params$width %||% (resolution(data$x, FALSE) * 0.9)
	data <- transform(data,
	ymin = pmin(y, 0), ymax = pmax(y, 0),
	xmin = x - width / 2, xmax = x + width / 2, width = NULL
	)
	flip_data(data, params$flipped_aes)
	},

[clauswilke]

@ggrothendieck FYI, ggplot2 uses size instead of lwd, so making the lines thicker does work. However, I think this is only a workaround, not a proper solution, if only due to the alpha issue I pointed out.

library(ggplot2)
x <- 1:6
y <- x * c(1, -1)

ggplot(data.frame(x, y), aes(x = x, y = y)) + 
  geom_col(width = 0, colour = "black", size = 3)

^{Created on 2019-11-18 by the reprex package (v0.3.0)}

[thomasp85]

If we end up adding it (and I’m not convinced it is warranted) we should not implement it through a stat. As long as stats are not composable reserving the stat slot for something as simple as setting ymin to zero is needlessly limiting the use. Rather, it should just subclass GeomLinerange and set ymin in setup_data() A’s is done in GeomArea etc

[yutannihilation]

Thanks, makes sense.