[R-sig-ME] random effects models with small n

[Felix Bach]

Yes, there are multiple measurements. However the literature says that the sample size of subjects need to have a certain size to make reliable inferences for fixed effects and an n of 8 per group is rather small. I would assume that the estimates of the covariance matrices are very unstable and even if the estimate is unbiased you would get a result which could not be generalised to the underlying population 
Felix


-------- Original-Nachricht --------
> Datum: Sat, 18 Jun 2011 09:44:00 -0400
> Von: Thomas Levine <tkl22 at cornell.edu>
> An: Felix Bach <Felix_Bach at gmx.de>
> CC: r-sig-mixed-models at r-project.org
> Betreff: Re: [R-sig-ME] random effects models with small n

> Not knowing more about the data, I have to guess that multiple
> measurements were taken on each of the 16 animals, making the sample
> size larger than 16.
> 
> http://en.wikipedia.org/wiki/Statistical_unit
> 
> Tom
> 
> On Fri, Jun 17, 2011 at 7:08 AM, Felix Bach <Felix_Bach at gmx.de> wrote:
> > Dear all,
> >
> > I recently saw an article in which the authorsperformed a random effects
> analysis with lme4  which used a group (2 levels)x condition (3 levels) x
> effect (4 levels) and ratio (10 levels) as fixed factors and subject as
> random factor. Sample size was 16 animals. They used MCMC estimation and
> obtained p-values of 0.0001 (lets ignore the use of p values in a Bayesian
> framework). How reliable are those estimates? I can't imagine that you can
> reliable are p-values (or credibility intervals) with such a complex models and
> a sample size of 16? If the model assumption of an exchangable covariance
> structure is valid, could you trust the results?
> > Best wishes, Felix
> >
> > --
> >
> > _______________________________________________
> > R-sig-mixed-models at r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >

--