Challenge with getting nice logarithmic scale for box plot.

[Konrad Zdeb]

Colleagues,

I'm trying to generate a presentable box plot for the attached variable. yscale(log) produces rather meagre results:



I'm trying to follow the FAQ on creating box plots with logarithmic scale, and I'm using the mylabels (SSC). I drafted the following code:

// Create nice log scale
clonevar logmeasure = measure
replace logmeasure = log10(measure)

// Define labels
mylabels 0(20)100 500(2000)10000 15000(30000)115963, ///
    myscale(log10(@)) local(labels)

// Graph box plot with population and area size
graph box logmeasure, ///
    ytitle("Measure", size(small) margin(vsmall)) ///
    ylabel(, labsize(vsmall) angle(horizontal)) ///
    plotregion(lstyle(none)) ///
    lines(lwidth(vthin)) ///
    ylabel(`labels', angle(h)) ///
    title("mylabels + log10") ///
    name(box_measure, replace)


// Drop log scale var
drop logmeasure

which produces:


The obtained box plot looks slightly more like the one suggested in the FAQ but it's far from perfect. Consequently, I was wondering if someone could advise me how should I adjust the mylabels syntax in order to obtain nice 8 - 10 labels that would highlight the distribution of the attached data and spread nicely across the graph. With respect to the attachment, the data is in tab delimited format as, oddly, I couldn't attach dat or csv file.

[Nick Cox]

mylabels (SSC) can be useful in this and other problems, and I am as fond of it as anybody else, but it's a distraction here.

First off note that

1. Konrad's data (which define a single variable measure) range from about 1 to about 100000, so immediately 6 labels for the powers of ten 1, 10, 100, 1000, 10000, 100000 are natural.

2. Attempting to put 0 on a log scale is doomed.

3. Despite that FAQ, I would argue that using 1.5 IQR to determine what to plot outside the box is especially awkward on a log scale. Many readers would guess wrong what was being shown, and explaining it simply is a challenge. There is an easy alternative: draw whiskers to specified quantiles or percentiles, as the commutative property log of quantiles = quantile of logs is exact in principle and in practice compromised only slightly by any use of averaging adjacent values to get quantiles at specific % points.

4. Often with box plots what they leave out can be as interesting or important as what they show. So, showing a box together with something more detailed is usually a good idea.

For these and other reasons I would suggest stripplot (also SSC) as in this respect more versatile than graph box.

If a log scale is natural, or at least convenient, binning is ipso facto natural, or at least convenient, on that scale. That being so, it is a good idea to use

 
gen log_measure = log10(measure)

After some playing around here is one possibility:

stripplot log_measure , box(barw(0.1))  boffset(-0.1) pctile(1)  xla(0 "1" 1 "10" 2 "100" 3 "1000" 4 "10000" 5 "100000") width(0.01) ms(sh) msize(*.25) stack xtitle(measure)

Here the whiskers are drawn to 1% and 99% percentiles. What is more intriguing is the apparent secondary mode on the log scale.

If you wanted more axis labels, then variations on 1 3 10 30 100 ... or 1 2 5 10 20 50 100 ... spring to mind, with the risk that the graph gets too busy. For those mylabels might indeed be useful.

[Konrad Zdeb]

Nick,

Thanks, stripplot comes often as very nice alternative. I have this and some other variables that are pronounced by similar distributions my ambition was to box plot them on log scales and then contrast on one chart. Having said that, I agree that this is not an easiest distribution to show. What I'm trying to do here is to say:

• vast majority of values clusters around those values
• but there are some values that are way off from what we would normally expect
• some other variables are pronounced by similar characteristics (but other don't) - here similar box plots with different scales would follow

Having said that I can combine some stripplots with box plots and it will work as well, thanks for the useful suggestion.

[Nick Cox]

Also, check out multqplot (SJ).

[Konrad Zdeb]

Also, check out multqplot (SJ).

I tried an it returned the following picture.


I think that I will go with the first chart that was suggested as it's most readable to me. On the other matter it occurred to me that qplot is required to run this:

. multqplot measure
unrecognized command:  qplot
r(199);