Skip to content

MANOVA with random factor

4 messages · Helios de Rosario, Peter Dalgaard, David Costantini

Helios de Rosario
# 
Hi, after re-reading I think that I misunderstood your question. You
don't provide many details, but I suppose that the "brood" effect is
nested within the fixed effects, so you don't mean a multivariate
approach for a split-plot or a repeated-measures design, but the
analysis of a multivariate mixed effects model. Do you?

In that case, perhaps the package MCMCglmm may help, although I have
never used it, so I can't tell anything for sure.

Helios
<David.Costantini at glasgow.ac.uk> escribi?:
group*sex) and
does
INSTITUTO DE BIOMEC?NICA DE VALENCIA
Universidad Polit?cnica de Valencia ? Edificio 9C
Camino de Vera s/n ? 46022 VALENCIA (ESPA?A)
Tel. +34 96 387 91 60 ? Fax +34 96 387 91 69
www.ibv.org

  Antes de imprimir este e-mail piense bien si es necesario hacerlo.
En cumplimiento de la Ley Org?nica 15/1999 reguladora de la Protecci?n
de Datos de Car?cter Personal, le informamos de que el presente mensaje
contiene informaci?n confidencial, siendo para uso exclusivo del
destinatario arriba indicado. En caso de no ser usted el destinatario
del mismo le informamos que su recepci?n no le autoriza a su divulgaci?n
o reproducci?n por cualquier medio, debiendo destruirlo de inmediato,
rog?ndole lo notifique al remitente.
Peter Dalgaard
#
On May 18, 2012, at 16:15 , David Costantini wrote:

            

      
      
      
    

        
It might but it isn't easy. And especially in small samples it might not do it reliably.

First, is this a balanced design? I.e. for each brood, you have say two males and two females one of each allocated to each experimental group. In such models, the analysis can be partitioned in an analysis of brood averages and within-brood differences, where the latter can be obtained by adding brood as a fixed effect. In fact, you can do that if you just have the same number of observations per brood, but without balancing, you'll get information on some effects from multiple strata. All of this carries over to multivariate responses, but it isn't implemented for aov()/manova() so you'd have to do the averages yourself.

For arbitrary designs, lme() is probably the best/only route. The trick there is to pretend that the variables are repeated measures on the same unit and then specify an arbitrary within-unit covariance structure. It does get rather involved and I'm afraid I don't recall exactly how to do it --- it's something with correlation=, weights= for the finest-level effect (variable within individual within brood) and pdSymm for the coarser level (variable within brood). 

lme() is a bit old and has known issues with degrees-of-freedom calculations so works best if you have a sizable number of broods so that df are large enough not to matter.

Notice also, that this is the basic R-help group. There is also R-sig-ME which specializes in mixed effects, and it might be better to take the issue there. (In particular, some people there may have better memory than I do...)

-pd