[R-sig-ME] Using nAGQ = 0 in glmer or lmer

Phillip Alday me @end|ng |rom ph||||p@|d@y@com
Wed Mar 2 06:21:41 CET 2022

If my understanding of lme4 is correct, then nAGQ=0 and nAGQ=1 both use
the Laplace approximation, but differ in whether the fixed effects are
optimized as part of the PIRLS step (=0) or via the nonlinear optimizer
(=1). Generally the latter is more accurate, but in casual testing, I
haven't seen much of a difference.

I tend to use nAGQ=0 for "good enough" things like simulation-based
power analysis / model selection and then use nAGQ=1 for the final model
used for interpretation.

There are a few pathological cases where the fit will converge for one
of those options but not the other (I've only seen this with Poisson
models), in which case the decision is made for you.

On 15/2/22 2:39 pm, Hedyeh Ahmadi wrote:
> Hello All,
> I have used the option nAGQ=0 in many of lme models and I have read about it, my understanding is that the difference between the default, which nAGQ=1 vs nAGQ=0 is the estimation process in the background.
> My question is that, in general if we use nAGQ=0, is our estimation still trustworthy?
> Does it make it more accurate/trustworthy if our sample size is large?
> Are there scenarios that we should absolutely not use nAGQ=0?
> Thank you in advcance.
> Best,
> Hedyeh Ahmadi
> 	[[alternative HTML version deleted]]
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

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