Boxplots don't get drawn properly when there are zeros and it is log transformed

[ctSkennerton]

I came across an issue where some parts of the box plot don't get drawn when I log transform data that had a lot of zeroes in it. The image below illustrates what I mean. Having a quick look at the code I can see that cbook.boxplot_stats is returning -inf for some things like IQR or min values which I guess is causing weirdness.

import pandas as pd
from plotnine import *

reprex = pd.DataFrame({'value': [0,0,1,5,100000,4739273,11,0.0001,0, 0.00001,0.00001,1,5,100000,4739273,11,0.0001,0.00001], 
                       'cat':['1','1','1','1','1','1','1','1','1','2','2','2','2','2','2','2','2','2']})
ggplot(reprex, aes('cat', 'value')) + scale_y_log10() + geom_boxplot()

[has2k1]

I think you should get a warning about divide by zero operations. If that is the case then the problematic plot should not be a surprise, after-all log(0) is infinite.

However since visualisation is essentially an imprecise representation of data, for your case you can exploit this fact to obtain an interpretation of the data that is visually accurate with something like.

epsilon = np.finfo(float).eps
ggplot(reprex, aes('cat', 'value+epsilon')) + scale_y_log10() + geom_boxplot()

That should prevent the log(0) operations.

[ctSkennerton]

Thanks for the solution. I think I was more surprised, coming from ggplot2, where this seems to be handled internally.

[has2k1]

Well, I guess we could handle it too.