Hello Trevor,
> I have a few questions but as they may be a bit basic I won't send this
> to the ASREML list.
I disagree. Others are at the same point as you and are looking to
the list for ideas so your experience is valuable. I'll reply to the list
Hope you can forgive me but we are all still learning.
> (1) If I first use REML,in GENSTAT for WINDOWS, to analyse the data
> as a lattice and then compare it to ASREML most things are similar (i.e.
> variance components, variety means etc) but the REML deviance is 1647
> with 122df while ASREML gives LogL=-730.979 with 125df. Clearly -2 times
> the LogL is 1461.958 which is not equal to 1647. I assume this is
> reasonable given the different algorithms used and also it is the change
> when fitting different random models that is more of interest. However
> the difference in the df confuses me. Is 122 preferable to 125 (given
> 1df for mu, 24df for variety and 3 random components i.e. 150-1-24-3=122
> ).
The diffenece in likelihood is not casued by the difference between
Fisher Scoring (GEML/GENSTAT) and Average Information (ASREML).
It has three possible sources. First, the true LogL contains some constants which,
being constants, are sometimes omitted from the evaluation
of the likelihood. Second, changing the scale of the data will change
the LogL. Thirdly, changes to the way the fixed effects model is
parameterized will change the LogL.
So LogL values are only comparable on the same data, fitting the same fixed model
and in the same program.
You are right that difference in LogL (caused by changing the model for
ranbdom terms in the model) is what is of interest.
Concerning degrees of freedom, dropping the extra degrees of freedom is
good to the Lattice model but in general, I am not sure we can drop a
degree of freedom for every variance parameter. Therefore, ASREML does not
delete degrees of freedom for variance components. In more complex
models, it is not always obvious were the degrees of freedom should come from.
More may happen on this fron in the future.
> Second REML gives a Wald test of 347.3 on 24df while ASREML (if
> I'm correct) gives a Wald test of 14.47 on 24df. Is the former to be
> compared to a chi-square on 24df and the latter to a F on 24 and
> 125(122?)df????
Yes (more or less). 14.47 = 347.3/24
The Chi-square test assumes infinite error degrees of freedom so I prefer an
F test (also I can relate easier to an F test without referring to tables).
However, the proper error degrees of freedom is unknown and is probably
less than 125. If you did a fixed effects ANOVA
Reps 5
Treatments 24
Cols in Reps 24
Rows in Reps 24
Residual 72
you would probably test Treatments against Residual on 24 degrees of freedom
but if Row and Column effects were significant, it would be unclear exactly
what is the best test. The Wald F statistic is giving us a valid F value
for testing but we still do not know its correct error df (it will be somewhere
between 24 and 120).
>
> (2) While I've been able to fit the sequence of models and reproduce
> the numbers shown (within a reasonable no. of significant figures -
> including the change and the variance parameters) in your tables 7 and 8
> I can't get the same LogL.
> i.e. yours mine
> lattice -671.7 -731.0
> Model 1 -670.4 -729.7
> 2 -664.5 -723.3
> 3 -659.7 -718.5
> 4 -657.4 -716.2
> 5 -657.0 -715.8
> Ari says there was some late change in the final proof. Are my
> calculated values correct??
The values in the paper were actually from Brians Splus routine rather than
ASREML. The difference I think has two components. ASREML includes 0.5
times the error df in its evaluation of the LogL (125/2 = 62.5 which is the
bulk of the difference(58.8)). The rest is probably a difference from
constaining method (e.g. using mu con(var) rather than mu var
as the fixed model.
> (3) One thing which concerns me is if I include c(var) instead of var
> in the fixed model this changes the LogL(eg M1 -732.9 compared to
> -729.7). Everything else is identical although I do get confused by the
> 8 mu line at the bottom of the ?.asr output and have difficulty showing
> variety means are identical. Am I right to assume that the LogL should
> be unchanged when using c(var) instead of var in the model ??
>
OK the diffenece between c(var) and var is 3.2
( I predicted 3.7 above but maybe one of your numbers is slightly wrong).
You are wrong to assume the LogL does not change. Any change to the fixed
model will alter the LogL.
With the mu var model,
mu line is actually the predicted mean for variaty 1
and the 'var' lines are diffeneces from variety 1.
The var_1 effect is not printed in the .asr file but is printed in the .sln
file (with a value of zero)
> (4) Finally having become "comfortable" with ASREML I can produce
> output files (?.asr, ?.sln etc) and using -s2 in DOS I can produce a
> variogram(but lacking axis identification and scaling). However I believe
> there is S-PLUS code which converts these basic ASREML output files to
> the nice graphics and summary statistics used in your paper. Ari gave me
> a copy of Brian's "lhicode.q" but I'm finding great difficulty doing
> anything useful with it on my PC. This is perhaps understandable given
> I'm a S-PLUS novice. Ari was going to help me with this but
> unfortunately he never seems to find time given his new responsibilities.
> Is there any documentation available which could help me??
>
You should have a file asr.s which you can
source('asr.s') into Splus. You may need to modify the 'ASREML' command
in asr.s depending on where ASREML.EXE (ASREMLWN.EXE) is on your PC.
See some information in chapter 7 of the manual).
It does not do all the graphics that are in the paper but does do the
same graphics as ASREML but in a form that you can print from Splus.
I quess I'll have to put some axes on the ASREML graphs and enable them to be
printed. - Another job to do.
> Thanks for your efforts in making ASREML available and I hope the above
> does not take too much of your time, while giving you some feedback of
> matters which can cause concern.
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
Arthur Gilmour PhD email: gilmoua@agric.nsw.gov.au
Senior Research Scientist (Biometrics) fax: <61> 2 6391 3899
NSW Agriculture telephone work: <61> 2 6391 3815
Orange Agricultural Institute home: <61> 2 6362 0046
Forest Rd, ORANGE, 2800, AUSTRALIA
ASREML is currently free by anonymous ftp from pub/aar on ftp.res.bbsrc.ac.uk
Point your web browser at ftp://ftp.res.bbsrc.ac.uk/pub/aar/ **NOTE CHANGE**
in the IACR-Rothamsted information system http://www.res.bbsrc.ac.uk/
To join the asreml discussion list, send the message
subscribe
to asreml-request@chiswick.anprod.CSIRO.au
The address for messages to the list is asreml@chiswick.anprod.CSIRO.au
Where is He who has been born King of the Jews? For we have seen His star
in the East and have come to worship Him. Matthew 2:2
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
************************************************************************
TREVOR W HANCOCK, Biometry, Dept. of Plant Science, Univ. of Adelaide.
Waite Institute, PMB 1 Glen Osmond, South Aust.,AUSTRALIA. 5064
Tel: (08) 8303 7288 International: 61 8 8303 7288
Fax: (08) 8303 7109 International: 61 8 8303 7109
email: thancock@waite.adelaide.edu.au
************************************************************************
----- End Included Message -----