How to log transform the y-axis of R geom_histogram in the right direction?

user_4507231 0
Viewed 681

I'm forgetting something very fundamental which would explain why I'm seeing very inflated y values after a log10 transformation of the y-axis.

I have the following stacked ggplot + geom_histogram. 

ggTherapy <- ggplot(genderTherapyDF, aes(freq, fill=name)) +
 geom_histogram(data=genderTherapyDF, binwidth = 1, alpha=0.5, color="black") + theme_bw() +
 theme(legend.position="none", axis.title = element_text(size=14), legend.text = element_text(size=14), axis.text.y = element_text(size=12, angle=45), axis.text.x = element_text(size=12), legend.background = element_rect(fill="transparent")) +
 ylab("No. of patients") + xlab("Events") + labs(fill="") +  ggtitle("Therapy")

enter image description here

The y-values are true to form, exactly what I expect. However, it's so skewed that to the naked eye I'm finding this very unsatisfying. I'd rather see a transformed plot.

I tried transforming x, quickly to realise that transforming along the binned axis was very difficult to interpret. So I transformed the frequency on the y axis:

ggTherapy <- ggplot(genderTherapyDF, aes(freq, fill=name)) +
 geom_histogram(data=genderTherapyDF, binwidth = 1, alpha=0.5, color="black") + theme_bw() +
 theme(legend.position="none", axis.title = element_text(size=14), legend.text = element_text(size=14), axis.text.y = element_text(size=12, angle=45), axis.text.x = element_text(size=12), legend.background = element_rect(fill="transparent")) +
 ylab("No. of patients") + xlab("Events") + labs(fill="") +  ggtitle("Therapy") +
scale_y_log10()

enter image description here

Visually, the plot makes sense. However, I'm struggling to come to terms with the y-axis labels! Why are they so huge after a log10 transformation?

ggplot2 histogram r
2 Answers

I'm going to make a case against using a stacked position on a log transformed y axis.

Consider the following data.

df <- data.frame(
  x = c(1, 1),
  y = c(10, 10),
  z = c("A", "B")
)

It's just two equal observations from two groups sharing an x position. If we were to plot this in a stacked bar chart, it would look like the following:

library(ggplot2)
ggplot(df, aes(x, y, fill = z)) +
  geom_col(position = "stack")

And this does exactly what you expect it would do. However, if we now transform the y-axis, we get the following:

ggplot(df, aes(x, y, fill = z)) +
  geom_col(position = "stack") +
  scale_y_continuous(trans = "log10")

In the plot above, it seems that group B has the value 10, which is correct and group A has the value 90, which is incorrect. The reason this happens is because position adjustments happen after statistical transformation, so instead of log10(A + B), you are getting log10(A) + log10(B), which is the same as log10(A * B), as top height.

Instead, I'd recommend to not stack histograms if you plan on transforming the y-axis, but use the fill's alpha to tease them apart. Example below:

df <- data.frame(
  x = c(rnorm(100, 1), rnorm(100, 2)),
  z = rep(c("A", "B"), each = 100)
)

ggplot(df, aes(x, fill = z)) +
  geom_histogram(position = "identity", alpha = 0.5) +
  scale_y_continuous(trans = "log10")
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
#> Warning: Transformation introduced infinite values in continuous y-axis

Yes, the 0s will become -Inf but at least the y-axis is now correct.

EDIT: If you want to filter out the -Inf observations, one nice thing in the scales v1.1.1 package is the oob_censor_any() function used as follows:

scale_y_continuous(trans = "log10", oob = scales::oob_censor_any)
user_113748270

I'm guessing that you should transform the data manually as described here https://ggplot2-book.org/scales.html#continuous-position-scales:

"Note that there is nothing preventing you from performing the transformation manually. For example, instead of using scale_x_log10() to transform the scale, you could transform the data instead and plot log10(x). The appearance of the geom will be the same, but the tick labels will be different. Specifically, if you use a transformed scale, the axes will be labelled in the original data space; if you transform the data, the axes will be labelled in the transformed space. Regardless of which method you use, the transformation occurs before any statistical summaries. To transform after statistical computation use coord_trans(). See Section 14.1 for more details."

user_136799110
Related