topic Re: COVARIANCE STRUCTURE PROC MIXED in Statistical Procedures

COVARIANCE STRUCTURE PROC MIXED

L_L — Tue, 22 Nov 2011 10:51:01 GMT

Dear all

I'm using proc mixed for repeated mesaures with the code below:

Proc mixed data=ds_input;

class id m_time;

model y=x1 x2 x3 /solution;

repeated m_time / type=un subject=id;

run;

Could anyone explain me the difference between the sintax type=un and type=un(1)? What's the meaning of Banded Covariance structure?

When is it recommend to use a banded covariance structure? and an autoregressive covariance structure?

Thanks in advance for any help

COVARIANCE STRUCTURE PROC MIXED

lvm — Tue, 22 Nov 2011 13:45:14 GMT

Big difference. With UN(1), there is a separate variance for each level of time (assuming your repeated measure is time), and 0 covariance (correlation) between times. With UN, there is a separate variance for each, and separate covariances for each pair of times. UN is the most general structure, but it can be difficult to fit (there can be many variance-covariance parameters to estimate). With a repeated measure, I would expect some correlation between times; thus, UN(1) is usually not realistic. AR(1) should also be tried. There are many possible structures that are simpler than UN, but still account for correlations. I recommend you read the repeated measures chapter in the book SAS for Mixed Models, 2nd Edition, by R. Littell et al. (SAS Press).

Re: COVARIANCE STRUCTURE PROC MIXED

stan — Tue, 06 Aug 2013 19:34:53 GMT

Dear lvm, what is the difference between UN(1) and CSH then ? Thank you.

Re: COVARIANCE STRUCTURE PROC MIXED

lvm — Tue, 06 Aug 2013 19:43:09 GMT

Big difference. With UN(1), all the covariances (and thus correlations) within subjects are 0. The variances are all different (potentially) for each group. With CSH, there is also a separate variance for each group, but there are nonzero covariances for the pairs of times. With CSH, there is a single correlation for all possible pairs, but because the variances can be different, the covariances will also be different. The User's Guide for MIXED has a nice table showing the covariance matrix for many of the different structures.