Whiskers and outliers for geom_boxplot with transformed scale

[mbertolacci]

Boxplot whiskers and outliers seem to be incorrect when using geom_boxplot with a transformed scale. It probably affects notches too, but I haven't checked.

The hinges are fine, which makes sense—monotonic transformations preserve quantiles.

library(ggplot2)
df <- data.frame(x = 1, y = c(1, 4, 5, 6, 9))
print(ggplot(df, aes(x, y)) + geom_boxplot())

print(ggplot(df, aes(x, y)) + geom_boxplot() + scale_y_sqrt())

^{Created on 2019-05-20 by the reprex package (v0.3.0)}

I'm using ggplot2 3.1.1.

[paleolimbot]

You're right, boxplots are calculated based on the transformed values, which affects the calculation of outliers. However, StatBoxplot$compute_group() has access to the x and y scales (and their transformations), so it is possible to change this.

https://github.com/tidyverse/ggplot2/blob/master/R/stat-boxplot.r#L82-L85

[paleolimbot]

The only other place I can find a Stat that backtransforms a range is stat_function(): https://github.com/tidyverse/ggplot2/blob/master/R/stat-function.r#L93-L111

https://github.com/paleolimbot/ggplot2/blob/1f6f0cb4462aa9e3b5f51dc7fd594073e78351da/R/stat-function.r#L93-L111

[paleolimbot]

It's worth noting that this (surprise at Stats being calculated on transformed values) has been reported several times (#2880, #2804), and at least once it was deemed as a won't-fix: #2804 (comment)

[mbertolacci]

Ah, sorry I missed those issues—I only searched for boxplot related ones.

Computing stats on the transformed values is potentially quite reasonable from a statistical perspective. If that's the way this goes, I would then just suggest a documentation update to note the behaviour since it's a bit surprising.

[clauswilke]

This topic came up at Lee Wilkinson's talk yesterday at SDSS 2019: Stats are computed on the transformed values. If you want the stats to be calculated on the original values then you need to work with a transformed coordinate system: https://ggplot2.tidyverse.org/reference/coord_trans.html

This is common confusion in ggplot2: Many people try to use scales for things that should be done in coords. Setting axis limits without dropping data points is another common issue.

[paleolimbot]

If that's true, it should be much easier to make a pretty looking plot with coord_trans() - right now the defaults don't produce the plot most people are looking for, xlim and ylim parameters aren't named the same as coord_cartesian(), non-default axis expansion isn't possible (#3338 and #2990), second axes aren't well-supported (#2990), and some of the nice things you can do like use NA as one of the limits aren't possible in coords (#2907). Even using breaks = scales::log_breaks() doesn't really work in coord_trans().

library(ggplot2)

p <- ggplot(diamonds, aes(cut, price)) +
  geom_boxplot() 

p + aes(y = log10(price))

p + scale_y_log10()

p + coord_trans(y = "log10")

p + coord_trans(y = "log10") + scale_y_continuous(breaks = scales::log_breaks())

^{Created on 2019-06-02 by the reprex package (v0.2.1)}

[lock]

This old issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with reprex) and link to this issue. https://reprex.tidyverse.org/