[R-sig-ME] differing variances within different random effects levels

Ben Bolker bbolker at gmail.com
Mon Feb 14 21:53:52 CET 2011

  This is a question that's been bothering me for a while.  The answer
might be "read sections 4.2.2 and 4.2.3 of Pinheiro and Bates 2000
again, more carefully", but if anyone has done this before I wouldn't
mind some hints ...

  Consider a simplified version of the previous poster's example.
Suppose I have a randomized block design where T treatments are done for
each of N individuals within each of S species. My basic model for this
would be


(or fixed=response~treatment, random=~treatment|species/indiv in lme)

which would give me random effects


(and correlations between intercept and treatment effects at species and
indiv:species level).

  Suppose now that I think the *variance* among individuals varies among
species.  As far as I can tell, lme/lmer assume that this variance is
constant across species.  If I wanted variance across *observations*
(i.e. residual variance) to vary across species I could do it via
something like weights=varIdent(~species), but this only works for
residual variance.

  I suspect I could work out some way to do this with MCMCglmm, and I
don't have an immediate need for it, but I thought I would ask ...

    Ben Bolker

More information about the R-sig-mixed-models mailing list