[R-sig-ME] Maximum nAGQ=25?

[David Duffy]

On Mon, 30 Sep 2013, Rafael Sauter wrote:

> This is still about the nAGQ maximum.
>
> Thanks for your replies.
> I am aware that 25 is already a large number of quadrature points for a
> scalar random effect.
> I had one particular example in mind when asking you about this 25
> quadrature points. It is the toenail data [...]

> The fixed effect intercept differs by 0.139 when using nAGQ=25 or
> nAGQ=50. For the model with a random intercept only the differences
> reduce to the third digit. Comparable differences can be found when
> using NLMMIXED in SAS.
>
> I don't know if this is a large difference or not but the point is that
> I assume from this example that with even more complicated random
> effects the upper limit of 25 might be too low. Or is it just generally
> a bad idea to fit GLMM models which still have varying estimates with
> nAGQ=25?

The toenail data are quite pathological, very sensitive to key single 
observations - search back through the r-s-m-m archives and you will find 
some analyses IIRC.

Cheers, David Duffy.