jmp-l
[Top] [All Lists]

Re: repeated measures: random effects versus manova

To: jmp-l@lists.biostat.wustl.edu
Subject: Re: repeated measures: random effects versus manova
From: Carl Schwarz <cschwarz@stat.sfu.ca>
Date: Thu, 1 Jun 2006 14:48:57 -0700
In-reply-to: <189B140B-ABDB-41EF-98E2-D3C055C0BCBA@northwestern.edu>
References: <189B140B-ABDB-41EF-98E2-D3C055C0BCBA@northwestern.edu>
There are several reasons why the results won't be identical but the most important one is likely the different (implicit) assumption about the covariance of the 4 repeated measures.

The MANOVA approach assumes an unstructured covariance matrix, i.e. the covariances among the 4 measurements have no structure with each measurement having its own variance and each covariance separately estimated.

If you use the REML approach with subject as a random factor, you are forcing compound symmetry on the covariance matrix, i.e. all variances down the diagonal are equal and all covariances are also forced equal.

In SAS, but not in JMP, you can use proc mixed to fit repeated measures in various ways. In SAS, the repeated statement = manova approach and you can specify different covariance structures and get it to match the random(suject) approach.


There is a long detailed explanation in the SAS for Mixed Models book by Littell et al .




I am familiar with two ways of analyzing repeated measures effects in jmp. One of them uses a vertically-organized file structure, with multiple lines per subject, and uses subject as a random factor. The other uses a horizontally-organized file structure, with MANOVA. (Fit Model: MANOVA Personality: Choose Response: Coompound; interaction effects) The former approach makes sense to me and I can get it to work. However, I have the same dataset organized both ways, and I cannot get the MANOVA approach to reproduce the results I obtain in the random effects approach. Maybe it's not supposed to? And I don't know MANOVA, so maybe I'm not looking in the right way. If anyone is sufficiently interested and knowledgeable, I'd appreciate your input. I'm attaching the two datasets in case you want to check 'em out.

In the horizontal file, the four variables, prefleft, prefright, nonleft, nonright represent the repeated measures. The two within-subjects variables are laterality (left versus right) and preference (pref versus nonpref). So laterality is the faster moving variable. Orientation is a between-subjects factor. As an example of a result I would like to find in the MANOVA, in the random effects analysis (with all main effects and interactions), I get a significant interaction for Prefnon*Orientation, F(1,59)=10.8, P=.0017.



Attachment converted: Macintosh HD:amygdalahorizontal.jmp (SGPD/SGP ) (0025C01A)
Attachment converted: Macintosh HD:amygdalavertical.jmp (SGPD/SGP ) (0025C01B)
Michael Bailey
jm-bailey@northwestern.edu


<Prev in Thread] Current Thread [Next in Thread>