R help - Jun 2005 - Warning messages in lmer function (package lme4)

Hi:

I'm using function lmer from package lme4, and I get this message:



" There were 12 warnings (use warnings() to see them)"

So I checked them:

Warnings 1 to 11 said:

1: optim returned message ERROR: ABNORMAL_TERMINATION_IN_LNSRCH

 in: "LMEoptimize<-"(`*tmp*`, value = structure(list(maxIter = 50, 
...



and Warning 12 said:

12: IRLS iterations for glmm did not converge in: lmer(Enfermo ~ Bloque +
Trat * Var * dia + (1 | IDfruto), data = desinf,   ...



There was no error message in the call of lmer function, so:

What do these warnings mean?

Do these errors occur at earlier iterations but finally the model converges?

Can the model I got be trust?



Thanks, Gabriela.



PD: I'm using:

Windows XP ,

R 2.1.0,

Package lme4, version 0.95-6

And the fitted model was:

modelo1 <- lmer(Enfermo ~ Bloque+Trat*Var*dia+(1|IDfruto),

                 data=desinf,
subset=!desinf$dia==1&!desinf$Trat=="Sin_inoculo",

                 family=binomial)

On 6/13/05, CENDOYA, Mar??a Gabriela <gcendoya at balcarce.inta.gov.ar>
wrote:
> Hi:
> 
> I'm using function lmer from package lme4, and I get this message:
> 
> " There were 12 warnings (use warnings() to see them)"
> 
> So I checked them:
> 
> Warnings 1 to 11 said:
> 
> 1: optim returned message ERROR: ABNORMAL_TERMINATION_IN_LNSRCH
> 
>  in: "LMEoptimize<-"(`*tmp*`, value = structure(list(maxIter =
50,   ...
> 
> 
> 
> and Warning 12 said:
> 
> 12: IRLS iterations for glmm did not converge in: lmer(Enfermo ~ Bloque +
> Trat * Var * dia + (1 | IDfruto), data = desinf,   ...
> There was no error message in the call of lmer function, so:
> 
> What do these warnings mean?
> 
> Do these errors occur at earlier iterations but finally the model
converges?
Yes. The PQL and Laplace methods for fitting generalized linear mixed
models perform optimization of a criterion that is itself determined
by an optimization problem.  These warnings come from earlier
iterations and are warnings.
 
> Can the model I got be trusted?
Well George Box said that "All models are wrong" so in that sense you
can't trust any fitted model.  This type of model is difficult to fit
and you should, as with all model fits, check the parameter values to
see if they make sense to you.

You may want to try the same fit in a test build of R-devel, which
will become R-2.2.0.  In R-2.1.0 part of the optimization problem is
performed by optim() and I have had some peculiar results from optim
using method="L-BFGS-B".  In constrained minimization I have had
optim() declare convergence to a value on the boundary where the
objective function was larger than at the starting point.  I have
added the function nlminb() to the development version of R-2.2.0 and
use that for the constrained optimization problem in lmer().  I have
found the results to be more reliable than those from optim().