[R-sig-ME] Variance estimate at 0 with lmer - not with lme - issue or rounding ?

[Simon Chamaillé-Jammes]

Very much appreciated. Excellent (and quick) support as usual.

Jehol. 

<(By the way, is that really the output from the model fit
< because the name "name" in the last line should be "indiv".)
I had indeed converting the variables' names to make my message more
general. I forgot to change this one, but it's the right model output.


 

On Fri, 2012-05-11 at 14:14 -0500, Douglas Bates wrote:
> On Fri, May 11, 2012 at 2:16 AM, nospamjehol-lists at yahoo.fr
> <nospamjehol-lists at yahoo.fr> wrote:
> > Hello all,
> >
> > I'm getting puzzled by the following:
> >
> > I'm fitting a simple hierarchical model with individual identity as random factor. Fitted with lmer (lme4 package) the result look like this:
> >
> > Formula: response ~ some fixed effects + (1 | indiv)
> > Random effects:
> >  Groups   Name        Variance Std.Dev.
> >  indiv     (Intercept) 0.0000   0.0000
> >  Residual             1.4802   1.2166
> > Number of obs: 116, groups: name, 69
> >
> > The variance estimate for individuals is a perfect 0 (which is always suspicious...and maybe an indication of problem in data distribution across effects)
> 
> No, an estimate of zero is not suspicious.  It is simply an indication
> that the variability between individuals is not significantly larger
> than what one would expect from the random variability in the
> response.  (By the way, is that really the output from the model fit
> because the name "name" in the last line should be "indiv".)
> 
> > I then fitted the model with lme (in nlme package) and got:
> > Formula: lme(fixed=response ~ some fixed effects,random= ~ 1 | indiv)
> > Random effects:
> >  Formula: ~1 | indiv
> >          (Intercept) Residual
> > StdDev: 9.796623e-05 1.2166
> > which I agree is a (very) small variance estimate but is less suspicious.
> 
> One of the internal differences between lme and lmer is that the
> variance component estimates in lmer can go to zero gracefully whereas
> in lme they can't because lme works on the scale of the logarithm of
> the relative standard deviation.  It took a long time to figure out
> how to solve the "going to zero gracefully" problem.
> 
> > I'm hitting this "issue" regularly with this dataset, as under some combination of fixed effect variables the variance estimate for individuals suddenly drops to a perfect 0.
> >
> > So to help me figure out if I should look for a problem in the distribution of data across variables, my question is:
> >
> > Is the difference in variance estimated by lmer and lme function such as the one observed here likely just caused by a rounding issue (and thus I should assume that with lmer the estimate is correct and simply 0), or is it sign of a real "problem" in my dataset?
> 
> Just a rounding issue.
> 
> > As ever, thanks for your help (here and in other posts that provide answers so we don't have to ask)
> >
> > Jehol
> >        [[alternative HTML version deleted]]
> >
> >
> > _______________________________________________
> > R-sig-mixed-models at r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >