turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- How to show degrees of feedom for random component...

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

12-22-2015 01:05 PM

How do I show the degrees of freedom for random components of a mixed model in proc mixed?

Is there a way to show them in the table of the output of the covariance parameter estimates? Does anyone know????

Thanks a lot

Antoine

Here is my syntax:

proc mixed data = work.import method=reml;

class ID A B C D E;

model outcome = A B C D A*B /s ddfm = kr;

random intercept A B E / subject = ID v = 1;

run;

repeated / subject = ID r = 1;

run;

Accepted Solutions

Solution

12-22-2015
05:07 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

Posted in reply to Antoine37

12-22-2015 01:31 PM

There are no degrees of freedom for an estimated variance or covariance based on likelihood (restricted likelihood) analysis. That's the short answer. It can be more complicated (for different purposes). An estimated variance asymptotically has a normal distribution (i.e., a t distribution with infinite df). But asymptotic results kick in here only for VERY large sample size. It is rare that one can rely on large-sample theory to use the normal approximation (although you can request this with a COVTEST statement in GLIMMIX). One can also test an estimated variance based on a likelihood ratio test. This still relies on asymptotic theory, but simulations generally show that one can use this approach at much smaller and more reasonable sample sizes. If you have one variance term (in addition to a residual), you can use:

COVTEST 0;

to test for the equality of the variance to 0. The df nominally is 1 for this (basically 1 for the numerator df and infinity for the denominator df), and a chi-squared test statistic (with 1 df) is used. However, for certain common situations, other work shows that the test statistic is more complicated than that, being a mixture of two chi-squared distributions (with 1 df and 0 df for the two components). The COVTEST statement figures this out automatically.

Anyway, I go back to the start of the message. I would not report any df for your estimated variances.

All Replies

Solution

12-22-2015
05:07 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

Posted in reply to Antoine37

12-22-2015 01:31 PM

There are no degrees of freedom for an estimated variance or covariance based on likelihood (restricted likelihood) analysis. That's the short answer. It can be more complicated (for different purposes). An estimated variance asymptotically has a normal distribution (i.e., a t distribution with infinite df). But asymptotic results kick in here only for VERY large sample size. It is rare that one can rely on large-sample theory to use the normal approximation (although you can request this with a COVTEST statement in GLIMMIX). One can also test an estimated variance based on a likelihood ratio test. This still relies on asymptotic theory, but simulations generally show that one can use this approach at much smaller and more reasonable sample sizes. If you have one variance term (in addition to a residual), you can use:

COVTEST 0;

to test for the equality of the variance to 0. The df nominally is 1 for this (basically 1 for the numerator df and infinity for the denominator df), and a chi-squared test statistic (with 1 df) is used. However, for certain common situations, other work shows that the test statistic is more complicated than that, being a mixture of two chi-squared distributions (with 1 df and 0 df for the two components). The COVTEST statement figures this out automatically.

Anyway, I go back to the start of the message. I would not report any df for your estimated variances.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Permalink
- Email to a Friend
- Report Inappropriate Content

12-22-2015 01:48 PM

Thank you very much for your support in helping me. I think your answer is very clear !

Regards