[R-sig-ME] nAGQ = 0

Rolf Turner r.turner at auckland.ac.nz
Sun Sep 3 23:48:52 CEST 2017


On 04/09/17 03:48, Jonathan Judge wrote:
> Rolf:
> 
> I have not studied this extensively with smaller datasets, but with
> larger datasets --- five-figure and especially six-figure n --- I have
> found that it often makes no difference.

> When uncertain, I have used a likelihood ratio test to see if the
> differences are likely to be material.

> My overall suggestion would be that if the dataset is small enough
> for  this choice to matter, it is probably also small enough to solve the
> model through MCMC, in which case I would recommending using that,
> because the incorporated uncertainty often gives you better parameter
> estimates than any increased level of quadrature.


Thanks Jonathan.

(a) How small is "small"?  I have 3 figure n's. I am currently mucking 
about with two data sets.  One has 952 observations (with 22 treatment 
groups, 3 random effect reps per group).  The other has 142 observations 
(with 6 treatment groups and again 3 reps per group).  Would you call 
the latter data set small?

(b) I've never had the courage to try the MCMC approaches to mixed 
models; have just used lme4.  I guess it's time that I bit the bullet.
Psigh.  This is going to take me a while.  As an old dog I *can* learn 
new tricks, but I learn them *slowly*. :-)

(c) In respect of the likelihood ratio test that you suggest --- sorry 
to be a thicko, but I don't get it.  It seems to me that one is fitting 
the *same model* in both instances, so the "degrees of freedom" for such 
a test would be zero.  What am I missing?

Thanks again.

cheers,

Rolf

-- 
Technical Editor ANZJS
Department of Statistics
University of Auckland
Phone: +64-9-373-7599 ext. 88276



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