[R-sig-ME] False convergence lmer

Ben Bolker bbolker at gmail.com
Thu Nov 22 23:39:05 CET 2012

Brickhill, Daisy <r01db11 at ...> writes:

> I am modelling how the movements of birds effects their reproductive
> success. I have 7 groups of nests which have differing mean brood
> sizes (e.g. group A: brood size 3.5, group B: brood size 3.0 etc). A
> bird dispersing from A to B therefore would lose 0.5 of a chick.What
> I want to know is if the difference in brood size gained (or lost)
> by movements varies with nest availability in 'better' areas
> (i.e. all those groups that have higher brood size than the group of
> origin). I have constructed a mixed model using lme4 with a gaussian
> distribution:

Brood size difference between origin and  destination ~
  number of better nests available +number of nests available in the group of
origin + 
   (1|group of origin) + (1|group of destination).

> However I get the warning message:
> Warning message:
> In mer_finalize(ans) : false convergence (8)
> Looking at the mailing list I have tried standardising the predictor
variables, I have tried dividing them
> by 100. 

 (Dividing by 100 seems useless: why?)

> The only thing that works on the model is removing either the random effect of
 group of origin or
> group of destination (removing the fixed effects still gives the error
message). I wondered if the
> problem is that the same combinations of groups of origin and destination
always have the same difference
> between them?

  Offhand, it shouldn't.

> The other thing that occured to me is that birds are quite likely to
> remain in their group of origin and for some groups this is as high
> as 67% - might that be a cause?  If it's not likely to be this can
> anyone give me any pointers as to what to look for, to work out what
> it is in my data that is causing this problem?

  It's really hard to say precisely (unfortunately).  You might try the
development version of lme4 to see if it helps.

  Remember that this is a warning and not an error.  Do the answers look
reasonable?  Are some of the random effects variances set to zero?

  This is incomplete but I'm sending it before I forget about it.

  Ben Bolker

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