Sorting box plots

[Konrad Zdeb]

Is there a way I could quickly sort box plots by interquartile range or some other measure reflecting dispersion? For presentational purposes, I have to generate a set of box plots that are sorted in a manner corresponding to dispersion. There are no sub-categories, I am using the simple code below:

graph box var1- var2, asyvars showyvars yvaroptions(label(angle(ninety) labsize(vsmall))) legend(off)

[Nick Cox]

If you are showing different variables, the quickest way is just to do something like

 
tabstat price-foreign, s(iqr) c(s)

and look at the results.

[Konrad Zdeb]

Nick, I was looking at tabstat results but I have over 50 indicators to graph. Is there a way to do it more automatically? In effect, sorted values in the matrix saved by tabstat correspond to my desired order variables. I just do not know how to impose one on another.

[Nick Cox]

You'd have to write a program, I think. I'd reshape the data to long, work out group iqr with egen and reshape back to wide.

[Konrad Zdeb]

Nick, that would be neat solution. In the meanwhile, I primitively exported the tabstat results to Excel, sorted by value, joined lines and exported sorted variable list to Stata.

[Konrad Zdeb]

Nick, that would be neat solution. In the meanwhile, I primitively exported the tabstat results to Excel, sorted by value, joined lines and exported sorted variable list to Stata, which I passed to graph box command.

[Nick Cox]

Thinking about it showed a quick-and-dirty fix

sysuse auto, clear

graph box price-foreign, name(g1)  

gen name = "" 
gen iqr = . 

local J = 0 

quietly foreach v of var price-foreign { 
     local ++J 
     su `v', detail 
     replace iqr = r(p75) - r(p25) in `J' 
     replace name = "`v'" in `J' 
} 

sort iqr 

forval j = 1/`J' { 
    local names `names' `=name[`j']' 
} 

graph box `names', name(g2)

This assumes that number of observations >= number of variables you want to show. If not, there is a cleaner Mata-based way of doing this.

[Konrad Zdeb]

Thank you very much for posting the code. As a matter of fact, I have 30 something observations and 50 something variables.

[Nick Cox]

You can fudge this by

set obs 60

[Konrad Zdeb]

Thanks, I'll fudge it accordingly.