R help - Apr 2006 - lmer{lme4}, poisson family and residuals

Hello,

 

I’m trying to fit the following model:

Dependent variable: MAXDEPTH (the maximum depth reached by a penguin during
a given dive)

Fixed effects: SUCCESSMN (an index of the “individual quality” of a bird),
STUDYDAY (the day of the study, from -5 to 20, with 0=Dec 20), and the
interaction SUCCESSMN*STUDYDAY

Random effect: BIRD (the bird id, as each bird is performing several dives)

 

Using lmer with a poisson family, I got a huge amount of warning messages:

 
> reml.res <- lmer(MAXDEPTH~SUCCESSMN*STUDYDAY+(1|BIRD), family=poisson,
method=”Laplace”)

There were 50 or more warnings (use warnings() to see the first 50)
> warnings()
Warning messages:

1: non-integer x = 18.500000

2: non-integer x = 20.500000

3: non-integer x = 16.500000

…

 

Whereas it seems to be fine using the quasipoisson family:

 
> reml.res <- lmer(MAXDEPTH~SUCCESSMN*STUDYDAY+(1|BIRD),
family=quasipoisson(link="log"), method="Laplace")
> summary(reml.res)
Generalized linear mixed model fit using Laplace 

Formula: MAXDEPTH ~ SUCCESSMN * STUDYDAY + (1 | BIRD) 

 Family: quasipoisson(log link)

      AIC      BIC    logLik deviance

 39581.16 39616.86 -19785.58 39571.16

Random effects:

 Groups   Name        Variance Std.Dev.

 BIRD     (Intercept) 0.12383  0.3519  

 Residual             4.05206  2.0130  

number of obs: 9324, groups: BIRD, 12

 

Fixed effects:

                    Estimate Std. Error t value

(Intercept)        3.6725244  0.1063978  34.517

SUCCESSMN          0.2978806  0.6824748   0.436

STUDYDAY           0.0069852  0.0021121   3.307

SUCCESSMN:STUDYDAY 0.0571982  0.0123831   4.619

 

Correlation of Fixed Effects:

            (Intr) SUCCESSMN STUDYD

SUCCESSMN   -0.276                 

STUDYDAY    -0.099 -0.017          

SUCCESSMN:S -0.075 -0.035     0.823

 

My first question is therefore: can someone tell me why I can’t fit the
model with family=poisson?

 

Second, I would like to check the residuals but I really don’t know how,
even after several hours searching the archives and other helps. I tried:

 
> residuals(reml.res)
Error: 'residuals' is not implemented yet
> resid(reml.res)
Error: 'resid' is not implemented yet
> names(reml.res)
NULL
> mcmcsamp(reml.res)
Error in .local(object, n, verbose, ...) : 

        mcmcsamp for GLMMs not yet implemented in supernodal representation

 

I’m supposed to give a talk tomorrow and would be very grateful if someone
can help me…

 

I’m working with the following versions:

Package:              Matrix

Version:              0.995-8

Date:                 2006-03-21

Package:       lme4

Version:       0.995-2

Date:          2006-01-17

 

Thank you!

 

Amélie

 


	[[alternative HTML version deleted]]

Amelie LESCROEL <lescroel_cebc <at> no-log.org> writes:
> 
> Hello,
> 
> I?m trying to fit the following model:
> 
> Dependent variable: MAXDEPTH (the maximum depth reached by a penguin during
> a given dive)
> 
> Fixed effects: SUCCESSMN (an index of the ?individual quality? of a bird),
> STUDYDAY (the day of the study, from -5 to 20, with 0=Dec 20), and the
> interaction SUCCESSMN*STUDYDAY
> 
> Random effect: BIRD (the bird id, as each bird is performing several dives)
> 
  It isn't immediately clear to me why the maximum depth should
be Poisson-distributed ... as R is trying to tell you, Poisson
distributions only make sense for integer data.  Your best bets
are probably (1) reconsider the error distribution, (2) perhaps
trying using nlme instead of lmer -- it is more polished, and you
can use Pinheiro and Bates "Mixed-Effects Models in S and S-PLUS"
as a reference.

  Ben Bolker