The Biometrics meeting in Adelaide was very succuessful
and one delegste referred to ASREML as 'the flavour of the month'
although I trust it has a longer life than that.
I reran your job and offer the following explanation.
> From: "Leigh Callinan" <email@example.com>
> Fellow ASREML users,
> I have a data set for a RBD with 5 levels of lime as treatment and 4
> blocks. When I analyse this using ANOVA or REML I get the same RMS
> and Log Likelihood whether I consider the treatments as discrete or
> continuous(ie use orthogonal polynomials( OPs))(as you'd expect)..
> When I fit the model using ASREML, the RMS is 6% lower than that
> from ANOVA or REML. LogL is reduced by 10% when treatment is fitted as
> discrete levels, and increased by 1% when treatment is fitted as OP's.
> I don't think I've made any programming mistakes.
> Can anybody see the explanation for this?
> Leigh Callinan
The variance component for Blocks in your job was negative.
In the ANOVA (Genstat and any other ANOVA procedure),
this would appear as a BLOCKS F ratio less than one.
In the GENSTAT REML, it would have reported a negative variance component
(gamma value -0.058) for Blocks.
The default action in the ASREML job was apparantly to bound the
variance component to small positive. This has the effect of reducing the residual
variance. The unbounded result can be obtained by using the !GU qualifier
after '!r block' in the model line. NB Genstat also allows the option
of forcing the block variance to be positive
The second part of your question relates to changes in the LogL.
Genstat REML and ASREML do different things with
some constants in the LogL. Therefore, when you change the
parameterization of the fixed effects in a model in ASREML,
the LogL changes even though the model is equivalent. This is just
a change in the Constant part of the LogL. It means that LogL
values cannot be compared when the fixed part of the model changes
[even to an equivalent model].
I hope this sets you straight.