[R-sig-ME] valid estimates using lme4?

[Abhijit Dasgupta]

"Most accurate", in terms of least biased among the procedures tested, but not unbiased by any means. Even NLMIXED shows substantial bias in the Type I error, though perhaps less than the others. 

I think the basic point coming out of the paper is that estimation and inference in GLMMs is actually numerically a difficult problem in many cases, and so one should be careful when using them. 

I'm R addicted as well, so I'm not defending SAS by any means. I'm just saying that this paper doesn't support the statement that SAS is accurate and validated for this kind of data. I haven't seen anything similar done for Stata (the gllamm routines) or SPSS.

Abhijit
On Oct 29, 2011, at 2:19 PM, Alessandro Moscatelli wrote:

> The authors states:
> 
> "Judging from the results of our simulation study, we conclude that
> the SAS NLMIXED procedure
> provides the most accurate parameter estimates and inference (type I
> error) under correct model
> assumptions. Among the remaining procedures, estimates are generally
> biased, especially the type I
> errors, with the level of accuracy depending not so much on the number
> of quadrature points (for
> the procedures based on the integral approximation), but rather on the
> magnitude of the values of the
> parameters."
> 
> This is not my personal opinion (I don't use SAS at all...I am an
> R-addicted!) I am just citing the article...
> 
> 2011/10/29 Abhijit Dasgupta <adasgupta at araastat.com>:
>> Going over the paper quickly, it is just as much an indictment of SAS as it
>> is of R. So much for "validated statistical package".
>> 
>> On Sat, Oct 29, 2011 at 6:42 AM, Alessandro Moscatelli <moskante at gmail.com>
>> wrote:
>>> 
>>> The referee may have this article in mind:
>>> 
>>>  Zhang H, Lu N, Feng C, Thurston SW, Xia Y, Zhu L, Tu XM. On fitting
>>> generalized linear mixed-effects models for binary responses using
>>> different
>>> statistical packages. Stat Med. 2011 Jun 10. doi: 10.1002/sim.4265. [Epub
>>> ahead
>>> of print] PubMed PMID: 21671252; PubMed Central PMCID: PMC3175267.
>>> 
>>> when he ask for a "validated statistical package (SAS, STATA, SPSS) as
>>> lme4 is known not to be the best". Hope this can help you with the
>>> answer!
>>> 
>>> Alessandro Moscatelli
>>> Department of Cognitive Neuroscience
>>> Bielefeld Univerisy (DE)
>>> 
>>> _______________________________________________
>>> R-sig-mixed-models at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> 
>>