[R-sig-ME] starting values in glmmadmb and / or covariance specification in glmer
paul debes
paul.debes at utu.fi
Tue Aug 18 13:50:46 CEST 2015
Dear list members,
Does anybody know either how to use starting values for the variance
components using glmmadmb (glmmADMB package) or how I can specify a common
covariance across different variances for the random statement with glmer
(NOT with lmer that seems to work differently)?
Sorry for asking two different questions, but an answer to only one would
already help me.
Presently, I keep running out of memory (8GB) when trying to fit a set of
binary GLMMs with binomial error distribution using glmmadmb that I cannot
specify in glmer (which doesn't have any memory problems). Maybe
specifying starting values in glmmadmb would help?
I can fit all of a set of models with PQL (asreml-r package), but only
some with Laplace approximation (what I want) using glmer (lme4 package).
Hence, it is possible to fit the models generally. But I have no clue how
to specify a certain covariance structure in glmer (common covariance +
diagonal, equal variance).
I tried the following models with both glmer and glmmadmb; both work fine:
a) ~ 1 + Fixed.Effects + (1|Group) + (1|Fixed.Effect:Group) # simplest
model; common covariance (fitted by the Group term that is crossed with
Fixed.Effect) & common variance between / for all levels of Fixed.Effect
by Group
b) ~ 1 + Fixed.Effects + (0+Fixed.Effect|Group) # maximal model; different
covariance & different variance between / for all levels of Fixed.Effect
by Group levels
However, an intermediate model specification results in different
covariance structures in glmmadmb (where I run out of memory; I tested its
correct behaviour with a subset of data) vs. glmer:
c) ~ 1 + Fixed.Effects + (1|Group) + (0+Fixed.Effect|Group) # I try to
fit: a common covariance among levels of Fixed.Effect by Group, different
Group variances for each level of Fixed.Effect
The asreml-r syntax for the model I try to specify in c) is:
fixed = ~ 1 + Fixed.Effects,
random = ~ Group + idh(Fixed.Effects):Group
or alternatively (I'm not particularly interested in the amount of
variance for the crossed Group term, just in inferences about fixed
effects):
random = ~ corh(Fixed.Effects):Group
For c), glmmadmb and I understand each other - no covariances are fit
among Fixed.Effect by Group levels in addition to that implied by the
Group random term - whereas proper glmer syntax must be different as it
fits BOTH individual-pair covariances (printed as correlations with
'summary') AND the random term variance (I'm far away from wanting to fit
this).
How do I specify model c) appropriately in glmer or how do I specify
starting values in glmmadmb (to get rid of my memory-limitation problem
instead of learning new glmer syntax)?
Many thanks in advance,
Paul
--
Paul Debes
DFG Research Fellow
University of Turku
Department of Biology
Itäinen Pitkäkatu 4
20520 Turku
Finland
More information about the R-sig-mixed-models
mailing list