geom_jitter can produce the same jitter across facets

[mjsulik]

ggplot2 version 4.0.3

Example:

library(ggplot2)
test <- data.frame(
  group = c(1, 2, 1, 2, 1, 2), 
  x = c(1, 2, 3, 4, 5, 6), 
  y = c(0, 0, 0, 0, 1, 1))
ggplot(test, aes(x = x, y = y)) + 
  geom_jitter(height = .2, width = 0) + 
  facet_wrap(. ~ group)

Is this the intended behavior? I expected jitter values to be independently generated across facets.

[teunbrand]

The seed is setup once and re-used in every panel. I don't think that behaviour is intentional. Do you think it is incorrect?

[mjsulik]

Thanks for asking!

In my opinion, it would make sense for the default behavior to add jitter to each observation independent of facet. The following example shows how I thought this would work by default. Sorry for my ignorance, but is there already a way to use geom_jitter() to apply jitter to each observation independent of facet?

library(ggplot2)

manual <- data.frame(
  "group" = c(1, 1, 1, 2, 2, 2),
  x = c(1, 2, 3, 1, 2, 3),
  y = c(1, 2, 3, 1, 2, 3))
manual$y <- manual$y + runif(6, min = -.2, max = .2)
ggplot(manual, aes(x = x, y = y)) + 
  geom_point() + 
  facet_wrap(. ~ group)

I can imagine that having consistent jitter across facets could potentially be useful, but in that case how the jitter is mapped to observations across facets would be important. I created a couple of additional examples to try to understand that mapping, and manually setting the random number seed unexpectedly changed the function behavior, which seems undesirable.

library(ggplot2)

# 2nd group (x, y) values are in reverse order of 1st group
# it looks like jitter amounts are applied sequentially in the 
# order of the rows, not the values of the x or y aesthetics
# this makes sense (especially because those values may repeat)

test2 <- data.frame(
  group = c(1, 1, 1, 2, 2, 2),
  x = c(1, 2, 3, 3, 2, 1),
  y = c(1, 2, 3, 3, 2, 1))
ggplot(test2, aes(x = x, y = y)) + 
  geom_jitter(width = 0, height = .4) + 
  facet_wrap(. ~ group)

# setting the seed changes the results so now the jitter amounts 
# appear to be sorted by x value instead of row? 
set.seed(1)
ggplot(test2, aes(x = x, y = y)) + 
  geom_jitter(width = 0, height = .4) + 
  facet_wrap(. ~ group)

# how does missing data and unequal group sizes affect results?
# jitter values appear to be applied sequentially within the first facet, 
# including for missing values, and then recycled element-by-element 
# in the 2nd facet. This seems good! 
test3 <- data.frame(
  group = c(1, 1, 1, 1, 2, 2, 2),
  x = c(NA, 2, 3, 4, 1, 2, 3),
  y = c(NA, 2, 3, 4, 1, 2, 3))
ggplot(test3, aes(x = x, y = y)) + 
  geom_jitter(width = 0, height = .4) + 
  facet_wrap(. ~ group)

[teunbrand]

I guess the notion the panels should be jittered with different seeds hasn't come up a lot, or at least I can't find any history of similar remarks. I think manually adding jitter with + runif(n, -.2, .2) is what I'd recommend as a workaround as well.

If this workaround were to be applied to x and you want to preserve any potential discreteness of x, you could use a stage() trick like so: aes(x = stage(x, x + runif(length(x), -.2, .2)), y).

Alternatively, one could also use position_nudge() to do the jittering for you.

ggplot(test, aes(x, y, nudge_y = runif(nrow(test), -0.2, 0.2))) + 
  geom_point(position = "nudge") + 
  facet_wrap(. ~ group)

I don't think this is a high priority issue for us, but it might be nice to add in the next time we touch the jitter position.