Hi Jarrod,
I actually have 254 observations (152 individuals), and I left the
default prior
And indeed, the chain doesn't look very nice. But I can't get what
is the prpoblem....
Cheers
Celine
Hi Celine,
There is more variance than you expect (0.68/0.52 = 1.31X), but
this might be consistent with chance if sample size is small. For
example if n=30 you expect var(x)/mean(x) > 1.31 in about 10% of
cases if lambda=0.52. For n=30 I would expect values of zero for
the units variance to have some support in the posterior
(conditional on the prior of course). For sample sizes of around
100 I would expect the posterior to be well away from zero. How
many data do you have?
From a model perspective having a units variance of zero is not a
problem per se. From the perspective of MCMCglmm it will mean the
chain will not mix (if it is always exactly zero) or mix slowly (if
it is near zero).
Cheers,
Jarrod
Quoting C?line Teplitsky <teplitsky at mnhn.fr> on Fri, 11 Jul 2014
16:21:49 +0200:
Hi Jarrod,
many thanks for your answer. I've been trying to understand better
the idea behind the models before answering, but I'd like to be
sure I got this right.
In the data set I have
var(y)=0.68
mean(y)=0.52
and if I run a model with only intercept and residual, I get an
intercept of -0.81, so that the expected variance would be 0.44,
suggesting the data could be a bit overdispersed. But the residual
in this model is collapsing on 0.
In your latest version of the course notes, you mention p37" if
the residual was zero, then e would be a vector of zero and the
model would conform to the standard Poisson glm." So do I get this
right that no residual in a Poisson model is ok, just an indicator
of no overdispersion, but is not per se a problem?
Many thanks again for your help
Cheers
Celine
Le 23/06/2014 21:22, Jarrod Hadfield a ?crit :
Hi C?line,
Zero residual variance with (truncated) Poisson response would
imply that the data are under-dispersed with respect to the
(truncated) Poisson model. You could check this by comparing the
variance of the data with the expected variance given the
intercept.
Cheers,
Jarrod
Quoting C?line Teplitsky <teplitsky at mnhn.fr> on Fri, 20 Jun 2014
14:39:33 +0200:
Dear all,
I have recently bumped twice in the same issue running glmm in
MCMCglmm: the posterior distribution of residual collapses on 0.
While I have often seen it for other effects (e.g ID) and
interpreted it as evidence of non existence / non significance
of these effects, I can not get why residual variance would not
be well defined.
More specifically, with priors V=1, nu=0.02, I was trying to
estimate additive genetic variance in age at first breeding. I
first tried a Poisson distribution and the posterior
distribution of the residual looked more or less ok, although
not perfectly bell shaped. Then I thought as age at first
breeding could not be zero, that a zero truncated Poisson might
be better but then the posterior distribution of residual
variance totally collapses on zero. As I thought it could be due
to over parametrisation, I rerun the model with only intercept
but results were the same.
Is it a problem with the variables distributions not really
fitting the distribution I'm specifying? Any help would be
greatly appreciated!
Many thanks in advance
Celine
--
Celine Teplitsky
UMR 7204 - CESCO
D?partement Ecologie et Gestion de la Biodiversit?
CP 51
55 rue Buffon 75005 Paris
Webpage : http://www2.mnhn.fr/cersp/spip.php?rubrique96
Fax : (33-1)-4079-3835
Phone: (33-1)-4079-3443