[R-sig-ME] glmmPQL simplification

[Ben Bolker]

On 11-01-23 04:43 PM, Iker Vaquero Alba wrote:

  [snip]

>    Nevertheless, I guess I will have to manage to find out which of the
> interactions are really biologically important to reply to my question,
> and remove the rest of them, so that the df problem with the parameter
> estimation disappears. 

  Yes. Note that it's not necessarily an estimation problem, it's more
of a "does it even make sense to try to estimate this much from this
size data set". Frank Harrell's regression modeling strategies book is a
great reference for this stuff.

>    So, if I got it right, and correct me if I'm wrong, I can use the
> Wald test p-values encoded in summary() from the glmmPQL function, but
> being *very* careful about the fact that that's not the very best
> possible test for that type of data, although it can still be a good
> approximation. And of course, after trying to write a simpler model to
> avoid the issue with the "df-observations ratio". 

  Yes.

>    I wouldn't like to impose on you at all, but there's another issue I
> would like to find an answer to, if it was possible. When trying to fit
> the same model with lmer, I get a warning like this:
> 
> //"Warning message: In mer_finalize(ans) : false convergence (8)"
> //
> 
>    and when calling "summary()", this error message:
> 
>    "in asMethod(object) : *matrix* is *not* *symmetric* [1,2]"
> 
>    I have posted it before and no one seemed to know the reason very
> well. So, sorry for the insistence, but actually, after trying all the
> possible options in terms of removing interactions, I realized the term
> which is causing the problem is "briventral:tlength", this is, the
> interaction between the two possible sexual signals I'm testing, tail
> streamer length and ventral colouration, an important interaction for
> testing my hypothesis. When not including it in the model, I don't get
> the warning or the error message at all, and I cannot include it in the
> model until I only have 6 terms left. I can have an idea of the reason
> for the warning thanks to this post:
> https://stat.ethz.ch/pipermail/r-sig-mixed-models/2010q1/003306.html
>    but I still don't know what that error message about the matrix not
> being symmetric means.

  It basically means that something has gone wrong with the
optimization.  If you were willing to send me your data I could try to
dig around and see what's going on (and possibly improve the error
message in a future version of lme4).

  The hard part about these things is that once stuff starts to go wrong
(typically because you don't have quite enough data, or your model is
such that it is very sensitive to the data), it's very hard to diagnose,
and very hard to fix the fitting problems.  Mostly this is because these
things are just hard and usually require a fairly deep knowledge of the
algorithms involved.  Some of the standard tricks are to reduce the size
of the problem (as you have tried to do), or to center and scale the
predictor variables.  Or to try another optimization method, but that's
usually not an option.

  good luck
    Ben Bolker