2
$\begingroup$

Please consider the following dataset in SPSS:

v1  v2  weight
1   3   0,50
2   2   0,50
3   4   1,00
4   5   1,00
5   3   1,00
1   3   2,00
2   2   2,00
3   .   3,00
.   5   1,00
5   4   0,50

When I calculate correlations between v1 and v2 (pairwise missing deletion) with unweighted in SPSS, I get the following results:

Pearsons r: n=8; r=0.525
Kendalls tau: n=8, t=0,334

Suppose now, I'll apply case weights to the data, provided in the "weight" column. I now get the following results:

Pearsons r: n=8,5 (due to weighting); r=0.496
Kendalls tau: n=10 (weighted), t=0,274

My question/problem is that I don't understand why there are different weighted sample sizes for the two correlations. I understand SPSS weights as case weights, so for both correlations I was expecting a weighted sample size of 8,5, as it is indeed for Pearsons r.

Can anyone help me out here, why there is a different sample size for kendalls tau? If there is some scientific paper/website/whatsoever about that topic, I'm ready to read it on my own, but I didn't find anything useful on my own.

[Edit] Could it be that for Kendall's tau, SPSS is rounding the case weights to natural/whole numbers? In this case, we would indeed end up with n=10. But this then brings me to the question, why SPSS is doing so.

[Edit2] As suggested by Joel W., here's the SPSS syntax to reproduce what I'm doing.

Code for unweighted correlations:

CORRELATIONS
  /VARIABLES=v1 v2
  /PRINT=TWOTAIL NOSIG
  /MISSING=PAIRWISE.
NONPAR CORR
  /VARIABLES=v1 v2
  /PRINT=KENDALL TWOTAIL NOSIG
  /MISSING=PAIRWISE.

Code for weighted correlations:

WEIGHT BY weight.
CORRELATIONS
  /VARIABLES=v1 v2
  /PRINT=TWOTAIL NOSIG
  /MISSING=PAIRWISE.
NONPAR CORR
  /VARIABLES=v1 v2
  /PRINT=KENDALL TWOTAIL NOSIG
  /MISSING=PAIRWISE.

Thx, deschen2

$\endgroup$
  • $\begingroup$ Perhaps it would help if we were able to see your SPSS commands. $\endgroup$ – Joel W. Feb 4 '15 at 18:53
  • $\begingroup$ I could provide some syntax code, of course. But I'm not sure what's the point of doing so, because calculating correlations in SPSS is straightforward and relatively unambiguous. Nevertheless, I'll update my question. $\endgroup$ – deschen Feb 4 '15 at 22:06
  • $\begingroup$ It appears that 8.5 is correct. I don't think you can have a partial weight in a Kendall's tau calculation (because it's a ratio of counts, but maybe I'm wrong). Test it with some different weights. Change a weight from 0.5 to 0.01, and another from 1.0 to 1.1, and see what happens. $\endgroup$ – Jeremy Miles May 29 '15 at 17:17
  • $\begingroup$ You can also look at the SPSS algorithms. $\endgroup$ – Jeremy Miles May 29 '15 at 17:17
  • $\begingroup$ Finally, SPSS is unusual (unique?) in that it allows non-integer frequency weights. I don't know of another package that allows this. $\endgroup$ – Jeremy Miles May 29 '15 at 17:18
1
$\begingroup$

I just found an answer in a SPSS forum, just to let you know: http://www-01.ibm.com/support/knowledgecenter/SSLVMB_22.0.0/com.ibm.spss.statistics.algorithms/alg_nonpar_corr.htm

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.