[R-sig-ME] lmer option nAGQ=0
Ben Pelzer
b.pelzer at maw.ru.nl
Fri Mar 21 16:14:55 CET 2014
Hi Ben,
The pdf file of lme4 speaks of "the penalized iteratively reweighted
least squares step" and I was so naive as to think thought this would be
the same as PQL. Of course, if this would have been so, it would have
been explicitly mentioned there.
Though I don't know what "conditional modes" are (and when these could
be useful) it's good to know that it's NOT the same as pql! Sorry for
the confusion, and thanks for solving this,
Ben.
On 21-3-2014 14:12, Ben Bolker wrote:
> On 14-03-21 07:13 AM, Ben Pelzer wrote:
>> Dear list,
>>
>> In R version 3.0.3, I recently udated the lme4 package and for a small
>> dataset of N=108 pigsties, I ran logistic regression with a random
>> intercept across the pigsties. The dependent variable is assumed to be
>> binomially distributed, and represents the number of pigs in the sty
>> that have a roundworm infection. There is a dichotomous predictor,
>> denoting two different types of sty. Using the option nAGQ=0 produces
>> PQL estimates. These estimates are, however, quite different from those
>> obtained using sas (glimmix), spss (genlinmixed) and glmmPQL in R: these
>> three routines produce very similar estimates of the two fixed and the
>> one random effect. Now I'm wondering what the reason for the differences
>> compared with lmer nAGQ=0 may be. The option to run PQL with lmer may be
>> attractive if one has many (complex) models and large datasets, hence my
>> questioning. Thanks for any help!!
>>
>> Ben.
>
> It's not true that AGQ=0 produces PQL estimates (so it's not at all
> surprising that the results don't match the results of PQL estimates).
> Rather, it produces 'conditional estimates' -- it estimates the
> conditional modes, but doesn't apply a Laplace approximation (or any
> other). Don't have time to say more now, sorry.
>
> Ben Bolker
>
>
>
>
More information about the R-sig-mixed-models
mailing list