[Gretl-devel] different treatment of d.o.f. correction between sd()/sdc() and corr()/mcorr()

Allin Cottrell cottrell at wfu.edu
Wed Jul 1 13:21:52 EDT 2015


On Wed, 1 Jul 2015, Sven Schreiber wrote:

> I noticed that corr() and mcorr() produce the same results, but sd() 
> and sdc() seem to use different divisors.

True: both corr() and mcorr() use sample standard deviations in 
computing the correlation coefficient, but sd() and sdc() differ in 
that sd computes the sample standard deviation while sdc computes the 
MLE by default (with the option of supplying a df value).

> This "diff-in-diff" seems to me a bit arbitrary, or is there some 
> background story? Otherwise I'd say the functions should all use the 
> same convention. (which one I don't care)

Well, there's some sort of background here. (1) AFAIK, correlation 
coefficients are almost always computed using the sample sd. (2) The 
sd() function is intended for series and I'd guess that in that 
context most people want the sample sd most of the time. (3) In the 
matrix context of sdc() users are somewhat more likely to be doing 
something asymptotic.

This may not provide a very compelling rationale for what we do now, 
but the difference between sd() and sdc() is documented in the help 
for those functions.

Allin


More information about the Gretl-devel mailing list