[R-sig-ME] Help! What are the typical diagnosis that one can do when facing "fail to converge"?

[Joshua Wiley]

On Thu, May 10, 2012 at 9:10 AM, Ben Bolker <bbolker at gmail.com> wrote:
> arun <smartpink111 at ...> writes:
>
>> I also had a similar warning message but with lmer ( (Warning
>> message:In mer_finalize(ans) : false convergence (8)).  I used
>> verbose=TRUE in the model statement.  It will print each iteration
>> estimates.  I also tried to increase the iterations, but it didn't
>> work.  Then, I found this blog
>> (http://davidhughjones.blogspot.com/2009/11/lme-false-convergence.html).
>> It says to look for betas with estimates very low and divide that
>> variable by 10 or 100.  This was the only solution that worked for
>> me.  But, the estimates of beta for the variable and its
>> interactions will be 10 fold higher than expected.
>
>  A more generic piece of advice would be to scale and center
> all continuous predictor variables ... it won't always
> help, but it's easy to try.
>
> orig_data <- data.frame(V1=factor(1:5),V2=1:5,V3=(1:5)*0.001,V4=LETTERS[1:5])
> scaled_data <- as.data.frame(lapply(orig_data,
>                   function(x) {
>                        if (class(x) %in% c("integer","numeric")) {
>                              scale(x) } else x
>                    }))
>
> In doing this, the scaling and centering factors seem to get lost,
> so it's not a perfect solution.
> A full-fledged auto-scaling solution *might* be built into some
> future version of lme4 ...

if so FYI is.numeric tests the mode and the mode of both integer and
numeric classes is numeric so you can simplify slightly to:
if (is.numeric(x)) {
  scale(x)
} else x

>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models



-- 
Joshua Wiley
Ph.D. Student, Health Psychology
Programmer Analyst II, Statistical Consulting Group
University of California, Los Angeles
https://joshuawiley.com/