Dear Uilson,
Let me attempt to explain your results.
Assuming the rest of the job is correct, 'TRUE BLUP' is obtained
my setting !MAXIT=1 [see description of MAXIT in section 3.2.7 of the manual]
Setting !MAXIT=0 will be ignored and the default value of !MAXIT=10
would be used.
Also, since ASREML was designed promarily to estimate variances rather than
just do BLUP, it is not always obvious how to do a particular analysis
in a particular way.
The confusion possabily arises because ASREML parameterizes multivariate
analysis in terms of variance components but univariate analysis in terms
of variance ratios. Special use of the !S2== option is required to
override the default and set the error variance.
I have modified the jobs below to do what I think you want.
Arthur
> I was trying to obtain "TRUE" BLUP with asreml and I got some
> results I did not expect. I tried three types of analysis using as priors
> estimates of variance components obtained from another software: a)
> maxit=0 (analysis 1, below); b) maxit=1000, but using the option !GF
> (analysis 2); and c) I used as priors my variance components divided by
> the error variance component (5.387632/6.155146=0.8753078 and
> 6.155146/6.155146=1.0) and the option !GF.
> In analysis 1, asreml fails to converge, does not print the
> "estimated" variance components and the solutions for animals are
> different from those obtained with MTDFREML for true BLUP.
Because the setup is wrong [for what you are trying to do], the starting values
are not appropriate which is probably why this job failed.
I am suprised you say it does not print the 'estimated variance components'
but does print [albiet wrong] solutions for animal effects.
I suspect the .sln file contains solutions from a previous run.
>In analysis 2,
> my priors are different from the estimates for variance components, but
> are equal to the gammas. Also, the solutions for animals in analysis 2 are
> different from those obtained by MTDFREML and those obtained in analysis
> 1.
> In Analysis 3, the estimated variance components are close to
> those obtained with the other software. Also, here the solutions for
> animals are the same as those obtained with MTDFREML.
> After these results, I tried a fourth type of analysis (analysis
> 4, below). I used 5.387632 and 6.155146 as priors (assuming these are the
> best priors I could have) and the option maxit=1000. Now, asreml does not
> converge and does not print the variance components estimates; but prints
> the solutions for animals.
>
> Questions:
> 1. Which analysis should be used to obtain blup without estimating
> variance components ("true BLUP")?
> 2. What are priors for areml: the gammas or the variance components?
> 3. What is the meaning of the Sources of variation: "Variance"and
> "Residual"? Note here that when I used(results not shown):
> x1 ~ mu test test/blk !r anim 5.387632
> 1 1 0
> 0 0 IDEN 0
> the source of variation "Residual"does not appear in my results in the
> file FN.asr
>
> Thanks. Sorry the long message.
> Uilson Lopes
>
>
> ANALYSIS 1
> FN.as ----------------------
> Simulated data
> test 4 #
> blk 15 #
> anim 915 !P
> male
> female
> x1
> /home/uvl/HUBER/ped.dat
> /home/uvl/HUBER/dial.dat !maxit 0
> x1 ~ mu test test/blk !r anim 5.387632 # !GF
> 1 1 0
> 0 0 IDEN 6.155146 # !GF
By default, ASREML does univariate analysis [with one section of data]
using a variance ratio; the error variance is added to the model as a scaling
factor. In this analysis, the error variance is
fitted twice [as error variance and as the explicit IDEN 6.155146 term] hence
ASREML will find a singularity in the Average Information as well as
having poor starting values.
If the last 4 lines are changed to
> /home/uvl/HUBER/dial.dat !maxit 1 ### <<< MAXIT changed
> x1 ~ mu test test/blk !r anim 5.387632 # !GF
> 1 1 0
> 0 0 IDEN 6.155146 !S2==1 # !GF ### <<< !S2==1 added
the job should work as you intended. The last line could also be
written
0 !S2==6.155146
>
> FN.asr-------
>
> Logliklihood decreased to -.46E+04: Trying again with reduced updates
> Fault 0 Convergence Failed
> Final parameter values 59.26395 1.00000**********
> Finished: Sun May 17 17:23:38 1998 Convergence Failed
>
> FN.sln -----------------------
> anim 10 2.851 1.114
> anim 11 -2.006 1.118
> anim 12 -1.223 1.107
>
>
> ANALYSIS 2
> FN.as -------------------------
> /home/uvl/HUBER/dial.dat !maxit 1000
> x1 ~ mu test test/blk !r anim 5.387632 !GF
> 1 1 0
> 0 0 IDEN 6.155146 !GF
>
This analysis also suffers from the confusion because the model is
overspecified
It should work if the last line is either
0 0 IDEN 6.155146 !GF !S2==1
or
0 !S2==6.155146
> FN.asr ---------
> Source Model terms Gamma Component Compnt/StndErr
> anim 915 915 5.38763 12.7216 20.49
> Variance 900 840 1.00000 2.36126 20.49
> Residual identity 900 6.15515 14.5339 20.49
> Finished: Fri May 15 10:42:52 1998 LogL Converged
>
> FN.sln --------------
> anim 10 2.906 1.663
> anim 11 -2.049 1.666
> anim 12 -1.243 1.658
>
>
>
>
>
> ANALYSIS 3
> FN.as -------------------------
> /home/uvl/HUBER/dial.dat !maxit 1000
> x1 ~ mu test test/blk !r anim 0.8753078 !GF
> 1 1 0
> 0 0 IDEN 1.0 !GF
>
This job would work if you dropped the last 2 lines completely
[assuming 6.15515 is the correct solution for residual variance]
Alternatively, the last line should read
0 0 IDEN
>
> FN.asr ---------
> Source Model terms Gamma Component Compnt/StndErr
> anim 915 915 0.875308 5.32896 20.49
> Variance 900 840 1.00000 6.08810 20.49
> Residual identity 900 1.00000 6.08810 20.49
> Finished: Sun May 17 18:10:37 1998 LogL Converged
>
>
> FN.sln --------------
> anim 10 2.850 1.109
> anim 11 -2.005 1.113
> anim 12 -1.223 1.102
>
> ANALYSIS 4
> FN.as -----------------
> /home/uvl/HUBER/dial.dat !maxit 1000
> x1 ~ mu test test/blk !r anim 5.387632 # !GF
> 1 1 0
> 0 0 IDEN 6.155146 # !GF
> FN.asr ---------------------
>
> Logliklihood decreased to -.46E+04: Trying again with reduced updates
> Finished: Fri May 15 11:06:29 1998 Convergence Failed
>
> FN.sln --------------
> anim 10 2.906 1.663
> anim 11 -2.049 1.666
> anim 12 -1.243 1.658
This run should do the same thing as ANALYSIS 1 did. The blups
you report seem to be those from run 2. They are not reported
if convergence fails.
So assuming 5.38 and 6.155 are the correct REML solutions,
you should get these solutions from a run that just ends
/home/uvl/HUBER/dial.dat
x1 ~ mu test test/blk !r anim
If you wanted BLUPS using a particular heritability,
the job should end
/home/uvl/HUBER/dial.dat !maxit 1
x1 ~ mu test test/blk !r anim .87531
If you want BLUPS using a particular Genetic Variance and Error variance
the job should end
/home/uvl/HUBER/dial.dat !maxit 1
x1 ~ mu test test/blk !r anim 5.387632
1 1
0 0 !S2==6.15515
I hope this helps.
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
Arthur Gilmour PhD email: Arthur.Gilmour@agric.nsw.gov.au
Senior Research Scientist (Biometrics) fax: <61> 2 6391 3899
NSW Agriculture <61> 2 6391 3922
Orange Agricultural Institute telephone work: <61> 2 6391 3815
Forest Rd, ORANGE, 2800, AUSTRALIA home: <61> 2 6362 0046
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/
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
<> <> <> <> <> <> <>
"Seek first the kingdom of God and His righteousness" Jesus; Matthew 6: 33
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>