[R-sig-ME] R-sig-mixed-models Digest, Vol 28, Issue 4

Douglas Bates bates at stat.wisc.edu
Thu Apr 2 23:02:05 CEST 2009

On Thu, Apr 2, 2009 at 4:09 AM, Iasonas Lamprianou <lamprianou at yahoo.com> wrote:

> Dear all,
> I'll re-send this request since I got no reply to the first one.
> It is an issue which I face currently with lmer and MLWin and SPSS. This problem makes me feel very undomfortable. I have one standardized variable which represents the academic performance of children,  and I also have information about their school and their class.  I run the model with SPSS and lmer and I get the same result (both use REML). Then I use MLWin and I get different (but more reasonable results). MLWin uses IGLS and RIGLS and MCMC (all three methods agree when I use MLWin). I hereby present my numbers:

> I run the following model:

> Linear mixed model fit by REML
> Formula: Zmg_Arxiki ~ 1 + (1 | school)
>   Data: data
>   AIC   BIC logLik deviance REMLdev
> 21693 21714 -10844    21684   21687
> Random effects:
> Groups   Name        Variance Std.Dev.
> school   (Intercept) 0.37043  0.60863
> Residual             0.74465  0.86293
> Number of obs: 8448, groups: school, 47

> Fixed effects:
>            Estimate Std. Error t value
> (Intercept) -0.13515    0.08971  -1.506

> However, this is NOT reasonable because the variable is a standardized variable and the variance should be 1.0!! SPSS gives the same results.
> If I run the same model in MlWin, I correctly get

And you know there are the correct results because ...?

> S2 for the school level=0.077
> S2 (error) = 0.923
> and the total is 1.0 (correct)!

> Could anyone please let me know what happens? Any help is welcome.

I don't have access to neither the data nor MLWin so I don't have the
advantage of knowing what the correct results are.  :-)

Do you know if MLWin fits using maximum likelihood (ML) or REML?  If
it uses maximum likelihood then you might try fitting the model with
lmer setting REML = FALSE and see whether that can reproduce the
"correct" results.

By the way, why is it necessary that the sum of the variance estimates
in a mixed-effects model fit to a standardized variable must add to 1?
 Especially for imbalanced data I don't think that is required.

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