[R-sig-ME] negative Hessian eigenvalues

[Ben Bolker]

On 14-05-23 11:34 AM, Tobias Heed wrote:
> Hello,
> 
> I have read the long thread about the convergence warnings in lmer v.
> 1.1.6. However, I am still unclear as to how to work with them. If I
> understand correctly, I should probably ignore the max | grad |
> warnings, at least if they are close to the tol value. However, what
> about the negative Hessian eigenvalues? Can I ignore those as well?
> Is the number of negative Hessian eigenvalues relevant at all? In
> older versions, models also did not always converge. Does any
> combination of the new warnings map to the cases in which earlier
> versions would have told me that they did not achieve convergence?

  max |grad| warnings are (mostly) ignorable: the convergence warning
thread gives a recipe for computing the _scaled_ gradients if you want
to double-check to see if they would still give a convergence warning
with 1.1-7.

  Negative Hessian eigenvalues are most likely to indicate problems with
uncentered/unscaled variables.

  I'm not sure precisely which version/warnings you have in mind, but I
suspect there's not a simple mapping.

  cheers
    Ben Bolker


> Thanks Tobias
> 
> 
> -- 
> -------------------------------------------------------------------------------------------
>
> 
Dr. Tobias Heed
> Biological Psychology and Neuropsychology Faculty of Psychology and
> Human Movement Science | University of Hamburg Von-Melle-Park 11,
> Room 206 | D-20146 Hamburg, Germany Phone: (49) 40 - 42838 5831 |
> Fax: (49) 40 - 42838 6591 tobias.heed at uni-hamburg.de 
> -------------------------------------------------------------------------------------------
>
> 
[[alternative HTML version deleted]]
> 
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>