Categorical Overlay on Bar Graphs

[Joshua Matthews]

I'm having some trouble finding the right way to make this graph. I want to combine two bar charts and have them overlayed on each other, with the more general categorical variable overlaying the more specific one.

Here are the variables I am working with

* Example generated by -dataex-. For more info, type help dataex
clear
input double Q_RecaptchaScore float(RecaptchaFails recaptchafails)
 .4000000059604645 1 1
 .8999999761581421 0 0
 .8999999761581421 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
                 . . .
  .699999988079071 0 0
 .8999999761581421 0 0
.10000000149011612 1 1
  .800000011920929 0 0
 .8999999761581421 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
                 1 0 0
  .800000011920929 0 0
                 . . .
  .800000011920929 0 0
                 . . .
 .8999999761581421 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
.30000001192092896 1 1
 .8999999761581421 0 0
  .699999988079071 0 0
 .4000000059604645 1 1
  .800000011920929 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
 .4000000059604645 1 1
.30000001192092896 1 1
 .8999999761581421 0 0
                 1 0 0
  .800000011920929 0 0
 .8999999761581421 0 0
 .6000000238418579 0 0
  .800000011920929 0 0
.10000000149011612 1 1
  .699999988079071 0 0
  .699999988079071 0 0
 .8999999761581421 0 0
                 1 0 0
                 . . .
 .8999999761581421 0 0
 .4000000059604645 1 1
  .800000011920929 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
 .4000000059604645 1 1
                 . . .
  .800000011920929 0 0
                 . . .
 .8999999761581421 0 0
  .800000011920929 0 0
  .800000011920929 0 0
.10000000149011612 1 1
                 . . .
.10000000149011612 1 1
                 . . .
 .8999999761581421 0 0
 .8999999761581421 0 0
  .699999988079071 0 0
.30000001192092896 1 1
 .8999999761581421 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
.30000001192092896 1 1
  .800000011920929 0 0
                 1 0 0
  .800000011920929 0 0
 .4000000059604645 1 1
  .800000011920929 0 0
 .8999999761581421 0 0
.10000000149011612 1 1
 .4000000059604645 1 1
  .699999988079071 0 0
                 1 0 0
                 . . .
                 . . .
  .699999988079071 0 0
  .800000011920929 0 0
                 . . .
  .800000011920929 0 0
 .8999999761581421 0 0
 .8999999761581421 0 0
 .8999999761581421 0 0
  .800000011920929 0 0
 .8999999761581421 0 0
 .8999999761581421 0 0
.30000001192092896 1 1
 .8999999761581421 0 0
  .800000011920929 0 0
  .800000011920929 0 0
  .800000011920929 0 0
 .4000000059604645 1 1
 .8999999761581421 0 0
                 . . .
 .8999999761581421 0 0
  .800000011920929 0 0
  .800000011920929 0 0
  .800000011920929 0 0
end
label values RecaptchaFails FraudScoreFails
label def FraudScoreFails 0 "Valid", modify
label def FraudScoreFails 1 "Invalid", modify

Using

graph bar (count) Q_RecaptchaScore, over(RecaptchaFails, label)

I get a graph that looks like this:

Then using this

graph bar (count), over(Q_RecaptchaScore, label)

I get a graph that looks like this:

I've tried using

 graph bar (count), over(Q_RecaptchaScore, gap(*0))  over(RecaptchaFails, gap(*0))  stack

but that leaves the empty values on the valid and invalid sides like this:

I've also tried

 graph bar (count), over(recaptchafails, gap(*0))  over(Q_RecaptchaScore, gap(*0))  over(RecaptchaFails, gap(*0))
//and
graph bar (count), over(Q_RecaptchaScore, gap(*0))  over(RecaptchaFails, gap(*0))  over(recaptchafails, gap(*0))
//and
graph bar (count), over(Q_RecaptchaScore)  over(recaptchafails, gap(*0))  by(RecaptchaFails) asyvars

but the results are similarly ineffective.

[Nick Cox]

I don't follow exactly what you want, but what seems clear to me is that graph bar is quite wrong for your second graph. graph bar has ideas of order but its axis is categorical, not measured.

It seems that the values that occur are all multiples of 0.1. That being so, I would start with something like

histogram Q_RecaptchaScore, discrete barw(0.08) by(RecaptchaFails) freq fcolor(blue*0.3) lcolor(blue)

[Joshua Matthews]

That's closer, but I'm looking for something more like this where RecaptchaFails isn't split into two different graphs and the two bar graphs are superimposed on each other.

[Nick Cox]

You can make progress with superimposed histograms.

twoway histogram Q if R == 1, discrete barw(0.08) frequency fcolor(red*0.3) lcolor(red) || histogram Q if R == 0, discrete barw(0.08) frequency fcolor(blue*0.3) lcolor(blue) legend(order(1 "Fail" 2 "Pass")) scheme(s1color)

from which you can see that to get what you want you need to map your indicator to a new variable, 1 to 0.25, 0 to 0.75, and draw bars in the background with width 0.5 or so. Unless you are being told to do what you sketch, I suspect that readers will find something like the design above easier to follow.

The canonical principle here is that bar area represents frequency. To be compatible, total and component bars drawn with different bar widths would need to be on consistent scales. Again, the design just above is automatically consistent.

[Joshua Matthews]

Thank you! That's what I was looking for