Vertical line and crossbar jitter being applied differently.

[rharfoot]

I have this issue - basically the different components of the geom_errorbar have jitter being applied to them differently. I posted this on stack-exchange and someone there seemed to think it might be a bug.

library(tidyverse)

Basic data set:

df <- structure(list(
Test = c("A", "B", "C", "D", "A", "C", "D"), 
mean = c(1, 100.793684, 1, 1, 51.615601, 1, 2.456456), 
sd = c(1, 2.045985, 1, 1, 4.790053, 1, 4.250668), 
lower = c(2, 102.839669, 2, 2, 56.405654, 2, 6.707124), 
upper = c(0, 98.747699, 0, 0, 46.825548, 0, -1.79421)), 
row.names = c(NA, -7L), class = c("tbl_df", "tbl", "data.frame"))

Now the code I am using:

subplot <- ggplot(df, aes(x = Test, y = mean)) +
  geom_point(aes(x= Test, y = mean), 
                position = position_jitter(width = 0.2, height = 0.2))+
  geom_errorbar(aes(ymin = lower, ymax = upper),
                width = 0.1,
                position = position_jitter(width = 0.2, height = 0.2)) 
subplot

Thanks.

[clauswilke]

There are two separate issues here. First, you need to set the seed if you need two position_jitter() calls to produce the same jitter. Second, however, there is indeed something wrong with geom_errorbar(). The vertical lines are jittered correctly but the horizontal lines are not. geom_linerange() works fine.

library(ggplot2)

df <- structure(list(
  Test = c("A", "B", "C", "D", "A", "C", "D"), 
  mean = c(1, 100.793684, 1, 1, 51.615601, 1, 2.456456), 
  sd = c(1, 2.045985, 1, 1, 4.790053, 1, 4.250668), 
  lower = c(2, 102.839669, 2, 2, 56.405654, 2, 6.707124), 
  upper = c(0, 98.747699, 0, 0, 46.825548, 0, -1.79421)), 
  row.names = c(NA, -7L), class = c("tbl_df", "tbl", "data.frame"))

# works fine once seed is set
ggplot(df, aes(x = Test, y = mean)) +
  geom_point(aes(x= Test, y = mean), 
             position = position_jitter(width = 0.2, height = 0.2, seed = 123))+
  geom_linerange(aes(ymin = lower, ymax = upper),
                 position = position_jitter(width = 0.2, height = 0.2, seed = 123)) 

# doesn't work because the crossbars don't get jittered correctly
ggplot(df, aes(x = Test, y = mean)) +
  geom_point(aes(x= Test, y = mean), 
             position = position_jitter(width = 0.2, height = 0.2, seed = 123))+
  geom_errorbar(aes(ymin = lower, ymax = upper),
                width = 0.1,
                position = position_jitter(width = 0.2, height = 0.2, seed = 123)) 

^{Created on 2018-10-15 by the reprex package (v0.2.1)}

[rharfoot]

Thanks. I always forget to set the seed, I was initially using a jitter set before the geom_ calls, but this was having the same odd errorbar problem as in the examples above. I'll look at a work-around using geom_linerange

[clauswilke]

I have found out what the problem is.

The way the jittering code is set up, it actually jitters different x or y aesthetics separately, i.e., x, xmin, and xmax all get individually jittered:

ggplot2/R/position-jitter.r

Lines 76 to 79 in 3550772

	trans_x <- if (params$width > 0) function(x) jitter(x, amount = params$width)
	trans_y <- if (params$height > 0) function(x) jitter(x, amount = params$height)
	with_seed_null(params$seed, transform_position(data, trans_x, trans_y))

The solution could be something like the following, though I don't particularly like it:
https://github.com/clauswilke/ggplot2/blob/ed9be70c14e67e5448ac9f11aadce2cfb868764d/R/position-jitter.r#L76-L89

Also, it looks like position_jitter_dodge() may suffer from the same problem:

ggplot2/R/position-jitterdodge.R

Lines 67 to 70 in 3550772

	trans_x <- if (params$jitter.width > 0) function(x) jitter(x, amount = params$jitter.width)
	trans_y <- if (params$jitter.height > 0) function(x) jitter(x, amount = params$jitter.height)
	with_seed_null(params$seed, transform_position(data, trans_x, trans_y))

[rharfoot]

That's interesting. I have found that if I use set.seed(123) before the plotting call it fixes the problem.
I think it would still be useful to have a fix for this inside the plotting code, as it would make sense to have all the geoms following the same jitter from within the call.

[rharfoot]

Actually, ignore that - I just noticed I had changed to position_dodge and forgot to reset to position_jitter when testing. Adding a pre-called set.seed() does not fix the problem.

[paleolimbot]

This is just quick note that the aes_to_scale() function might be useful here if it the desired behaviour is to apply a single jitter per axis:

ggplot2:::aes_to_scale(c("x", "xmin", "xmax"))
#> [1] "x" "x" "x"

^{Created on 2019-05-23 by the reprex package (v0.2.1)}