[R-sig-ME] Help! What are the typical diagnosis that one can do when, facing
jwiley.psych at gmail.com
Thu May 10 20:05:50 CEST 2012
On Thu, May 10, 2012 at 10:46 AM, Ben Bolker <bbolker at gmail.com> wrote:
> Highland Statistics Ltd <highstat at ...> writes:
>> > Help! What are the typical diagnosis that one can do when facing "fail to
>> > converge"?
>> > Could anybody please shed some lights on us, besides increasing the number
>> > of iteration limit?
>> 1. Simplify your model
>> 2. Center or standardise your covariates
>> 3. Increase the number of iterations
>> 4. Lower the convergence criteria
>> 5. Provide starting values.
> Very nice.
> #6 try an alternative package/optimization platform, if one
> is available (in this case, if you are using a complex model
> for heteroscedasticity and want to stay within R, probably not ...)
You could specify a different optimization method via the control
argument. I think depending on the sort of complex model, OpenMx is a
viable alternative. It does not handle random effects yet to my
knowledge (I know the developers are hard at work to incorporate
that), but if you just have repeated measures, you fit a latent growth
style model, and create pretty much any pattern of associations you
can dream up a matrix for. If you are not used to working with
specifying matrices for SEM style models, that could be a substantial
learning curve but it puts you in a very powerful and flexible
framework for fitting many types of models.
>> Most likely you will need to do 1.
>> 2 may work....
>> 3-5 is for if you are stubborn and don't want to do 1.
>> For a more specific answer you will need to provide more details.
> R-sig-mixed-models at r-project.org mailing list
Ph.D. Student, Health Psychology
Programmer Analyst II, Statistical Consulting Group
University of California, Los Angeles
More information about the R-sig-mixed-models