[R-sig-ME] LMER False Convergence (8)

[Ben Bolker]

Daniel Anderson <daniela at ...> writes:

> Dear all, 
 
> I am trying to fit a cross-classified growth model with the lme4
> package. The model is cross-classified because students moved
> between schools during the three years of the study (for a similar
> application, see Luo & Kwok, 2012). I have estimated the following
> model with both lme4 and the HLM software.
 
> m.1<-lmer(MthScale ~ Clock + (Clock | mastid) + (Clock | School), 
> data=dta2, REML = F, verbose = T)
> 
> where, 
> MthScale is a mathematics achievement test, 
> Clock is a time variable coded 0, 1, 2,
> mastid = id for students,
> School = id for schools.


>  When I fit the model I get the following error message "In
> mer_finalize(ans) : false convergence (8)". Yet, my results are
> quite similar to the results I obtained from the HLM software, which
> gave me no such error (see attached).
 
> I have seen that some people have had success using David Hughes
> approach
> (http://davidhughjones.blogspot.com/2009/11/lme-false-convergence.html),
> but that did not work for me. I also read that it could be a result
> of the optimizer not reaching its predefined minimum, but that the
> result is likely to be a minimum
> (https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q1/015743.html).
>  Finally, I tried

> fitting the following much simpler model, and still received the
>  same error message:
 
> m.2<-lmer(MthScale ~ Clock + (Clock | mastid),
>  data=dta2, REML = F, verbose = T)
> 
> So, my questions are (a) are there any other ideas for
>   getting the model to converge without the warning
> message, and/or (b) is it safe to continue with model 
> building despite the warning?
 
> Below is the output from the verbose = T argument. I've also
> attached a word document comparing the estimates between lme4 and
> the HLM software.

The Word document didn't make it to the list, I think (probably
got stripped by the mailing list software).  Any chance of
attaching as a text file?

  If the estimates from HLM and lme4 are close (i.e. the differences
are much smaller than the estimated standard errors), then I can't
see any reason to worry very much ...

  Are you willing to give the development version of lme4
a try?  It's slightly more robust and eminently more tunable.
I think you should be able to install it via

install.packages("lme4",repos=c("http://lme4.r-forge.r-project.org/repos",
         getOption("repos")))