different results for pwcorr and mkcorr

[Andreas Head]

Hello dear Stata friends,

I am confused by the fact that stata produces different results for pwcorr and mkcorr.
I checked some forums for information on that, however, I could only find one post (on stata.com) stating the same issue
(http://www.stata.com/statalist/archi.../msg00732.html).

Is that actually a well-known issue, or am I doing something wrong?
I am using the exact same variables (around 20 vars) for each command.

mkcorr X Y Z, log(rp1 corr) sig cdec(2) mdec(2) casewise

pwcorr X Y Z , star(.1) bon

Any ideas?

Thanks,
Andreas

[Andreas Head]

I kindly ask if there is anyone who could help me out here?

[Nick Cox]

I can think of two reasons why you didn't get more of a response.

1. Many people will have wondered what is mkcorr and will not have wanted to find out themselves.

2. There is no reproducible example here (we don't have your X Y Z) and no explanation even of what you mean by different results.

These points are covered by FAQ Advice #12. Sometimes people tell you what is wrong and sometimes they shrug their shoulders and get back to whatever they were doing.

[ben earnhart]

Sometimes I'll bite. Do you take into account the difference between pairwise and listwise (casewise) deletion? Do the matrices have the same sample size (or with pairwise, sample sizes)? From the partial information you provide, sounds like mkcorr is using a different sample than pwcorr.

[Andreas Head]

Thank you Nick for pointing out the inadequate communication within my initial post. That makes sense. I should have provided more background information. My apologies for that.

Thank you Ben for your comment. I tried mkcorr without the casewise option and got the same correlation matrix now as from pwcorr.

Hence, the problem is solved!

[Andreas Head]

I wonder when you use the Bonferroni correction in correlation matrices generated by the use of e.g. Stata's pwcorr command? That is is there a common threshold, e.g. a specific number of variables, that suggests when to consider the Bonferroni correction necessary? I read that you use it in case of large multiple comparissons, however, it remains unclear to me how many variables make up such large multiple comparissons. In my case, I run pwcorr with 15-25 variables, including continous, binary, and ordered variables.

Thanks in advance.