Non-parametric partial/semipartial correlation

[Piotr Lewczuk]

Good morning,
Is there a way in Stata to perform non-parametric (Spearman) partial correlation? Something analogous to . pcorr x y z -?
Thank you in advance for a feedback.
Best greetings,
Piotr Lewczuk

[ericmelse]

Dear Piotr,

You can use the Stata function spearman
Type help spearman in the command window.

Possibly this webpage offers you some more background information.

[Piotr Lewczuk]

Good morning,
Are you sure one can run P A R T I A L correlation with . spearman?
One of us have obviously not read what the other writes
So, I'm still waiting for an answer, greetings,
Piotr

[ericmelse]

Dear Piotr,
I suppose what you are looking for is possible with Richard Williams' PCORR2: Stata module to display partial and semipartial correlation coefficients.

ssc install pcorr2

That provides partial and semi-partial statistics, as discussed by Richard Williams in Semipartial (Part) and Partial Correlation.

Next, I follow the recommendation of Nick Cox:

I can't speak for some alternative software that you are invoking here, but -egen, rank()- will rank your variables, after which partial correlations can be calculated in the usual way.

Some example of how to proceed:

use https://stats.idre.ucla.edu/stat/stata/notes/hsb2 , clear

foreach var in read write math science socst {
egen rnk_`var' = rank(`var')
}

* Assuming that a students' science performance is somehow determined, in part, by performance at other subjects:
pcorr2 rnk_science rnk_read rnk_math rnk_write rnk_socst

* result omitted

I hope this is helpfull.

[Piotr Lewczuk]

Dear Eric,
Thanks for pointing at pcorr2 and the paper; I am wondering what it means that pcorr2 "is a modified and enhanced version of Stata's pcorr command." (as it stays in the help for pcorr2).

Yes, I was aware of the Dr. Cox's suggestion to first rank the variable and then to run "usual" parametric partial correlation with pcorr (or pcorr2), I was just wondering, given that this suggestion is already a couple of years old, if there are any more direct approaches available meanwhile.

Anyway, thanks a lot for your feedback.

Regards,
Piotr Lewczuk

PS: The question is still open, if somebody on the forum knows a more direct approach than by the ranking of a variable.