[R-sig-ME] simulation

Ben Bolker bbolker at gmail.com
Tue Dec 27 00:30:14 CET 2011


Scott Raynaud <scott.raynaud at ...> writes:

> 
> Ran some more simulations and never could get lmer to converge.  I dumped
> one of my simulation into SAS and received a message that the covariance 
> matrix for the random effects was not positive definite.  Now this is pretty 
> informative to me: it says that the variance of the random effects is zero 
> (or negative).  Does that imply I can omit the random effects? 
>  I'm still a bit 
> suspicious since things run fine when I omit temperature.  Am I missing 
> anything here?

   A non-positive definite variance-covariance matrix for the random
effects does not imply that *all* the RE variances are zero (which is
what you would need to be comfortable dropping the random effects.
I haven't looked at your problem in detail; part of the problem/reason
you're not getting much help here is that it's hard for people (including
me) to get themselves geared up to read moderately complex code to see
what's going on -- if it looks more than a 30 min investment just to
get up to speed, most people won't have time unless the problem is something
that is particularly interesting to them.
  It is indeed likely that your model is underdetermined/unidentifiable
in some way ... you could try dropping *selected* random effects and see
if that fixes the problem (in lmer or SAS).

   Ben Bolker




More information about the R-sig-mixed-models mailing list