`geom_box()` and `scale_x_log10()` / `scale_y_log10()` produce incorrect whiskers/outliers.

[ja-ortiz-uniandes]

Description

When using geom_box() to plot data distributions, the whiskers appear to be computed as a linear graphical multiple of the IQR, regardless of the axis scale. As a result, applying scale_x_log10() or scale_y_log10() produces incorrect whiskers and incorrect outliers.

Reproducible example

R and ggplot version

> R.version.string
[1] "R version 4.5.1 (2025-06-13 ucrt)"
> 
> # Print ggplot2 version
> packageVersion("ggplot2")
[1] ‘4.0.0’

Demonstration of the issue

library(ggplot2)

set.seed(123)

# Generate exponential data
n <- 2000
rate <- 1
x <- (rexp(n, rate = rate) + 1) * 1e4

# Compute percentiles and IQR-based bounds
qs <- quantile(x, probs = c(0.25, 0.50, 0.75), names = FALSE)
q1  <- qs[1]
q2  <- qs[2]
q3  <- qs[3]
iqr <- IQR(x)
lower_bound <- q1 - 1.5 * iqr
upper_bound <- q3 + 1.5 * iqr

# Inspect computed values
stats <- data.frame(
  p25 = q1,
  p50 = q2,
  p75 = q3,
  IQR = iqr,
  lower_1p5_IQR = lower_bound,
  upper_1p5_IQR = upper_bound
)
stats

       p25     p50      p75      IQR lower_1p5_IQR upper_1p5_IQR
1 12847.19 17139.3 24405.04 11557.84      -4489.57       41741.8

# Plotting
base_plot <- ggplot(data.frame(x = x), aes(x = x)) +
  geom_boxplot() + labs(title = "Exponential sample (linear scale)")

p_log <- base_plot + scale_x_log10() + labs(title = "Exponential sample (log10 scale)")

base_plot

p_log

In the example above, the theoretical upper whisker should extend to the most extreme point within the range [p75, p75 + 1.5 × IQR]. In this case 1.5 × IQR = 41741.8, the whisker should end at the most extreme point below or equal to that value. This behaviour seems correct in the linear-scale plot.

However, when using a logarithmic scale, the whisker incorrectly extends beyond 50,000, which is outside the valid range. It appears that the whisker’s position is determined by a linear pixel distance corresponding to 1.5 × IQR, rather than using the correct scale transformation.

Expected behaviour

• The upper whisker should extend to the most extreme data point within [p75, p75 + 1.5 × IQR].
• The lower whisker should extend to the most extreme data point within [p25 - 1.5 × IQR, p25].
• “Extreme point” refers to the observation with the highest absolute value from the median.
• The whisker range should respect the scale of the axis on which it is plotted (e.g., logarithmic or linear), not a linear pixel distance.

[teunbrand]

This is the expected interaction between stats, like the boxplot stat, and scale transformations. Data are transformed by the scale first, and then the stat is computed on the scale-transformed data. You can use coord_transform(x = "log10") if transformation after the stat computation is needed.