[R-sig-ME] lme4 feature request: standardizing data and false convergence

[Adam D. I. Kramer]

Hi Jeffrey,

 	You can also avoid convergence errors in glmer by making sure your
expected coefficients are in the realm of 1e0 or 1e1. Standardizing works
because it mean-centers your data, so coefficients will probably be around
the mean...but you can fix things just as easily by dividing a variable with
a tiny coefficient by 100 or 1000.

 	...but...there's also a bit of an issue of HOW to standardized
in a multi-level model. For example, do you standardize a variable relative
to its grand mean?  Relative to the level-1 mean?  Or the highest level?  Or
relative to each mean in turn?  Etc.

 	Given that SAS has done this by default, as you suggest, which mean
do they standardize relative to?

--Adam


On Mon, 21 Jun 2010, Jeffrey Evans wrote:

> When using glmer in lme4 I frequently get false convergence errors. Many
> have posted to this list about this topic. The solution seems to be to
> standardize the X matrix first, which is effective at resolving the issue.
>
> Is there, (or could there be in future versions) a simple way to request
> that glmer optionally standardize the data internally and return
> back-transformed parameter estimates (on the original scale)?
>
> It seems that SAS does this by default in PROC GLIMMIX (version 9.2) and
> rarely had issues with regression conversion. It would be a huge help.
>
> Jeff Evans
> Dartmouth College
>
> 	[[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>