[R-sig-ME] need help with mixed effects model

[Mark W Kimpel]

First, I want to thank everyone on this SIG list for their generous time 
and thoughts on my problem. I am particularly grateful for learning that 
my problem is "multivariate" and that lmer is not (yet) set up for that. 
I will certainly be paying close attention to the list for the day when 
it is I think it is a terrific package.

Upon learning that what I am interested in is multi-variate correlation, 
not multi-factorial, a search of R packages turned up CORREP, which 
seems tailor made to my situation. In fact, perhaps once again a Nobel 
prize has slipped through my grasp as I find that yet another great idea 
of mine is not very original after all ;)

So, I'll give lmer a rest for now and see how CORREP does, but I do look 
forward to hearing about lmer's multivariate capabilities in the future.

Mark

Mark W. Kimpel MD  ** Neuroinformatics ** Dept. of Psychiatry
Indiana University School of Medicine

15032 Hunter Court, Westfield, IN  46074

(317) 490-5129 Work, & Mobile & VoiceMail
(317) 204-4202 Home (no voice mail please)

mwkimpel<at>gmail<dot>com

******************************************************************


David Duffy wrote:
> On Wed, 5 Mar 2008, Nick Isaac wrote:
> 
>> You can do SEM-type models using the smatr package. Only drawback is that
>> you'd have to treat Rat as a fixed effect. This is a general class of
>> problem that afflicts several of my current projects, and I'm having a tough
>> time choosing between mixed effect and structural equation models. The
>> former is most appropriate for partitioning the variance, but the latter is
>> most appropriate for modelling error variance in the observations. I don't
>> see an obvious solution with the available tools and would appreciate any
>> general insights.
>>
>> Cheers, Nick
>>
> 
> I don't know if smatr supports multiple groups (so you can have unbalanced 
> type data).  We generally use the Mx SEM package, which allows both 
> multiple groups as well as "irregular" data, and our main interest is always
> in partitioning the variance.  Every mixed model (indeed 
> every linear model) is expressible as a SEM (it just runs slower ;)).
> 
> Generally, if you want to include a random effect you set up a latent 
> variable with appropriate path(s) to the observed value(s).  Your
> relationships will be between the latent variables.  Often, you'll fix
> the path coefficient to unity.  Have a look at Mx book/manual (Mike Neale and
> Hermine Maes) and you'll get some ideas.
> 
> For the expression example, one model would be something like
>                               +--------------------------+
>                  Strain1-Rg1  |                          |
>               --------------------------    Rat16        |
>               |               |        |                 |
>                     r1        |                 r1       |
> Latent    Gene111 <-->  Gene112       Gene161 <-->  Gene162
>           /b1   | \b3   b4/ | \       /b1    | \ 
> Manif  Tissue1 T2 T3               Tissue1 T2 T3
> 
> With appropriate equality constraints (r1=r1..=r1, b1=b1..) or 
> random regressions on the path coefficientsi and covariance matrix of latent
> variables.  So this model states (and
> can test) that tissue expressions are imperfect measures of
> overall gene expression in the ith animal.
> 
> But after saying that, for this particular example, I would use a
> program such as ASREML or Wombat or MENDEL that easily fit multivariate
> mixed models.  Our group has used ASREML to fit mixture models to
> deal with error in DNA pooling experiments for 50000 SNPs.
> 
> Hope that's clear (and even right), David Duffy.
>