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

[Ben Bolker]

Daniel Anderson <daniela at ...> writes:

> 
> Hey Ben, I would love to provide the dataset but 
> unfortunately I don't think it's possible. It's an entire
> state's educational accountability testing data and the state 
> that we got the it from has required we use
> all sorts of securities when accessing it. It's kind of silly 
> how many regulations they have on it
> actually, considering it is already completely de-identified,
>  but such is life. Thanks again for all
> your help though, I really appreciate it.
> 
> Daniel

  There are a few other things we can try remotely to determine
whether the correlation=1 is correct (and HLM and the old version
of lme4 are getting not-quite-the-right answer) or the smaller
correlation of around 0.87 is more correct -- it would just
have been easier to do the experimentation myself rather than
by remote control, as it were.

  Some of the steps (without going into detail) are

(1) compare deviances estimated from the different methods;
if they look like they're estimated on the same scale, see whether
how the devel-lme4 deviance compares to the others
(2) construct a likelihood profile and see if the correlation
parameter is really maximum at the edge
(3) try starting lme4 from the parameters estimated by the other
programs, and/or evaluating the deviance at those parameters
(this might require some reparameterization/hacking)

  I'm happy to continue the conversation offline.  (I have
a vested interest in making sure that the development version
of lme4 does _not_ do worse than the currently released version
under any examples that I can find out about ...)


> 
> > From: Ben Bolker <bbolker at ...>
> > Subject: Re: [R-sig-ME] LMER False Convergence (8)
> > Date: May 31, 2013 3:44:26 PM PDT
> > To: r-sig-mixed-models at ...
> > 
> > 
> > Daniel Anderson <daniela <at> ...> writes:
> > 
> >> 
> >> Thanks again Ben, I went ahead and used the zip file from the
> >> link you sent to install lme4-3. This time I did
> >> not get the convergence warning. However, the verbose=T 
> >> statement did not appear to work (i.e., it didn't
> >> print the iteration history). Probably not a big deal,
> >> but I thought it was odd and wondered if it might also
> >> not be printing the warning message despite having the same issue.
> > 
> >  Well, this is a development version, so there might be something
> > funny -- I'll check.
> > 
> >> 
> >> Below, I've placed my results from the same model for all 
> >> three programs: HLM, lme4-2, and lme4-3. I also
> >> added the correlation between the variance components into
> >> this table, because with lme4-3 I get an
> >> estimated correlation between the random student intercept/slope 
> >> of 1.0 (which was not the case with the
> >> other programs). 
> >> 
> >> So, at this point I think I'm going to run with lme4-2,
> >> and just ignore the convergence warning, unless you'd
> >> suggest otherwise?
> > 
> >  That seems perfectly reasonable.  Would it be possible for you
> > to send your data so we can have a look and see what's happening?
> > (Ideally you could send it in an anonymized form so that we could
> > use it for tests/examples in future versions of the package, but
> > for now just having it for private use among the lme4 developers
> > would be great.)
> > 
> >  thanks
> >    Ben
> > 
> > 
> > 
> > 
> > _______________________________________________
> > R-sig-mixed-models mailing list
> > R-sig-mixed-models at ...
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 
> Thanks,
> --
> Daniel Anderson
> Research Assistant
> Behavioral Research and Teaching
> University of Oregon
> 
> 	[[alternative HTML version deleted]]
> 
>