>Percent variance accounted for is an exactly analogous calculation, but
>based on mean squares. Still, there is a problem in that the expected
>value of MS(Total) is influenced not only by the extent to which the model
>explains variation in the response variable, but also by the design of the
>experiment, for instance the number of replications of each level of a
>treatment factor. I think the Genstat people would say that for this
>reason it is better to compare variance components. Which is all very well
>for random effects, but what do you do with fixed effects...?
Wouldn't the smallest var(e) not be the better criterion for best model?
(definitely much better than R2, for reasons just described by Nick)
A R2 type criterion (in the lines of what Nick described) could be
1-(var(e)/var(E)) where var(E) is the residual variance
under a model with just fitting mu and var(e) the resid. variance under the
This criterion is just as valid for fixed as for mixed models, and can be
easily calculated with asreml.
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml