The likelihood and the deviance have are considerably more valuable than
the R-squared value. Even in simple models we know that the R-squared
value has major problems - especially with the number of parameters.
With modern methods, you can easily compute the likelihood to compare
models either directly or using Akaikes information citerion etc., or to
test fixed effects and variance components. There are issues about testing
effects that are associated with sample size and boundary conditions.
There has been some discussion about these on this mailing list.
Note that the residual variance is a bad criterion as the log(y) produces
a smaller residual than the untransformed value - it does not mean that
the model fits any better.
The appropriateness of a model relies on performing diagnostics. The
R-squared value never provides any information about this even in ANOVA
On Fri, 4 Aug 2000, Haja Kadarmideen wrote:
> Dear ASREML'ers,
> A very simple question and sorry if this has been clarified before.
> What is the equivalent of R-square (Multiple Correlation Coefficient) in
> ASREML-GLMM and how to get this ? -something that tells us about
> the adequacy of the model fitted, with effect of mean of y eliminated ?
>I am not sure if testing significance of 'fixed' or 'random' effect in the
> model would neccssarily lead to similar statistic, as they don't give us
> any idea about the appropriateness of the whole model itself ?
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml