[R-sig-ME] Problem- Warning message

Ben Bolker bbolker at gmail.com
Wed Mar 4 03:15:16 CET 2015


Ken Beath <ken.beath at ...> writes:

> 
> There are two random effects, another one for obs which is modelling
> overdispersion.
> 
> It looks like the optimizer has decided to stop, but has stopped when the
> gradients are not zero, that is it is not a maxima. It could just be the
> numerical gradients are a problem, but it could be something else. I've
> been having similar trouble with a glmm, and it seems to be a problem with
> the way glmer does something.
> 
> I don't know how much data that you have, but it is possible that your
> model is just too complex for the data available. The first rule of these
> models is to start with something simple and then add to the model until
> something breaks, then you can think about why it broke.

  As regular readers of this list will probably know, the tests for
convergence that we added in 1.1-6 were too strenuous (improved for
1.1-7, but still subject to false positives).  We are still discussing
how we can increase specificity without hurting sensitivity too much.
In the meantime, as a rule of thumb, I would probably not worry about
convergence warnings reporting scaled gradients of <0.1, especially
with large data sets.  However, 1.02 looks pretty big.

I would examine your data graphically to look for anomalies,
and try the recommendations listed at
http://rpubs.com/bbolker/lme4trouble1

> 
> On 27 February 2015 at 04:14, Davide Bellone <bellonedavide1 <at> gmail.com>
> wrote:
> 
> > Goodevening,
> >
> > this is my first mail . From what a collegue told me I can ask to the
> > members of this mailing list to help me with my problem.
> > So, what I am trying to do is a glmer model. I have 1 random effect (Site)
> > and 6 fixed effects.
> >
> > Model <- read.table("c:\\TXT X R\\Tabella modello gruppi
> > completo.txt",header=T)
> > y <- cbind(Model$Rem, Model$Total-Model$Rem)
> > Model$wood <- as.factor(Model$wood)
> > Model$under <- as.factor(Model$under)
> > Model$Manage <- as.factor(Model$Manage)
> > Model$Size <- as.factor(Model$Size)
> > Model$Site <- as.factor(Model$Site)
> > obs <- 1:length(Model$Site)
> > obs
> > options(contrasts=c("contr.sum", "contr.poly"))
> > R1 <- glmer(y ~ Model$Manage + Model$age +  Model$veg + Model$Size +
> > Model$wood + Model$under + Model$age*Model$Manage + Model$veg*Model$Manage
> > + (1|Model$Site)+(1|obs) , family= binomial)

This would probably be better as

Model$obs <- 1:nrow(Model)
R1 <- glmer(cbind(Rem,Total-Rem) ~ Size + wood + under + (age+ veg)*Manage +
             (1|Site)+(1|obs) , family= binomial)


> >
> > The first output is:
> > Warning message:
> > In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
> >   Model failed to converge with max|grad| = 1.02046 (tol = 0.001, component
> > 3)
> >
> > Can someone explain if there is some problem in data and if there is a
> > solution?
> >



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