Re: Variance estimates
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Variance estimates



Dear Steve,

There are a few issues to consider.

First, the R structure is defined the wrong way around.
You say the data was sorted and presumably the variance
for the first third should be less than for the middle third etc.

The structure as you have it implies data is order as 2547 lots
of 3.  You actually wanted 3 lots of 2547.

This can be specified either as I 
1 2 3 !STEP 0.1
 3 0 DIAG 9804.57 9761.96 9214.61 !GUUU !S2==1
2547 



or 
3 1 3 !STEP 0.1

2547      !S2=9805
2547      !S2=9761
2547      !S2=9215


Since your job is a univariate analysis (with 1 error structure), its
default parameterization is a scaling variance and variance ratios (Gammas).
So in the LogL lines is a variance (S2 is nearly 30).  This is the
relationship between GAMMAS and Variance components.
This is overcome by inserting !S2==1 in the R structure definition (see above).
The !S2==1  fixes the variance scale factor at 1.

What has happened is that ASREML has become overparameterised for the
variance and fixed the variance scale parameter at an initial value
becasue it was singular (in the information matrix).


I am not completely clear on your intentions.  I presume you have
several observations on most animals.  So you are fitting some sort
of nested model.  units wi animals wi sire wi herds

Presumably bdate is related to weight.
You have ignored bdate [which would need to be converted to age
to be useful]

It is sometimes useful to build up to the final model in stages
makinf sure you believe each stage . 


 X-Authentication-Warning: lamb.chiswick.anprod.csiro.au: petidomo set sender to 
asreml-owner@ram.chiswick.anprod.csiro.au using -f
> From: Steve Miller <millers@wright.aps.uoguelph.ca>
> Subject: Variance estimates
> To: asreml@ram.chiswick.anprod.csiro.au
> Date: Wed, 5 May 1999 17:15:42 -0500 (EDT)
> Mime-Version: 1.0
> 
> I am trying to interpret some output from a variance component
> estimation attempt with asreml.  I have run a random regression model
> for fat depth in beef cattle and I am modelling fat depth on weight.
> 
> The .as file is
> 
> Random regression of rib fat dpeth on weight
>  animal 3220 !I #Animal
>  bdate 731 !A # in format 22/07/69 etc
>  sire  199 !A #example 'NBBK090' with the quotes
>  breed   4 !A #!DHHHH !DMGMG !DSSSS #4 of these
>  herd   23 !A #
>  nut     8 !A #nutritional treatment imposed during grow out
>  fin     2 !A #finish, pasture or feedlot
>  mk1     3 !A #market, Domestic, Korean or Japanese
>  mk2     2 !A #market, Domestic or Export
>  cohort 12 !A #Taurus, steer, 1993, 1st intake TS931
>  dtwt   #date of weight eg   34491
>  wt !-422.5 !/237.5  #weight in kg standardized to -1 to 1 
>  dtp8  !M-1 #date of p8 fat measure
>  p8 !M-1 !*100 #p8 fat measure (mm)
>  dtrib !M-1 #date of rib measure
>  rib !M-1 !*100 #rib fat 12/13 measure (mm)
>  dtema !M-1 #date of eye-muscle area measure (sqcm)
>  ema  !M-1 #eye muscle area measure (sqcm)
>  bdate2 !M25406  #birth date in days like dtrib etc above
>  cg    41 !A #cohort.fin.mk2
>  n2    28 !A #cohort.nut
> 
> ../../../Data/data4.dat  !maxit 1000 # reading data file 
> rib ~ mu wt wt.wt cg cg.wt cg.wt.wt n2 n2.wt n2.wt.wt,
>   br br.wt br.wt.wt !r he he.wt s s.wt an an.wt 
> 
> 1 2 3 !STEP 0.1
> 2547
> 3 0 DIAG 9804.57 9761.96 9214.61 !GUUU
> 
> he 2
> 2 0 US 4702.17 4214.69 5946.26 !GUUU
> he
>  
> s 2
> 2 0 US 4702.17 4214.69 5946.26 !GUUU
> s
> 
> an 2
> 2 0 US 17854 18000 46610 !GUUU
> an
> 
> 
> The output for my variance components is
> 
>    Source       Model  terms     Gamma     Component    Compnt/StndErr
>    Residual   DIAG=Dia     1   328.423       9805.47         25.01
>    Residual   DIAG=Dia     2   326.861       9758.85         25.16
>    Residual   DIAG=Dia     3   308.629       9214.49         24.81
>   sire        US=UnStr     1   157.896       4714.19          7.31
>   sire        US=UnStr     1   142.277       4247.86          5.86
>   sire        US=UnStr     2   201.654       6020.64          5.98
>   animal      US=UnStr     1   583.459       17419.9         29.40
>   animal      US=UnStr     1   771.300       23028.1         29.85
>   animal      US=UnStr     2   867.785       25908.8         19.29
> 
> 
> 'Component' are the variance components I am interested in.  What are
> the 'Gamma' components and how sould I be using these?  I have three
> residuals.. the data is sorted by weight.  So there will be three
> residuals, one for each of three weight classes.
> 
> Also I have a problem that if I expand out the variances over the
> range of weights, in this case from -1 to 1 for the standardized
> weights some are negative... or sometimes the correlation between
> measures at different weights could be greater than 1.  How can I
> constrain the acceptable solutions to consider the parameter space
> where variances are not neagtive for example.

I do not follow what you mean here.

High covariances between the intercept and slope might be reduced by
centering the wt covariable.

Instead of using   2 0 US 4702 4215 5946
you can use        2 0 CORR .5 !GP !+2
                       4702 5946
                       
     Which will not allow the correlation to exceed 0.999
                         
> Take the component for animal above ... the variance at weight -1
> would be 17419.9 - 2 * 23028.1 + 25908.8 = -2727.5.  This is a
> negative variance... the problem exists in this case at the low
> weights.  I saw in the manual that placing some constraints on a
> function of the variance components is possible.  However, I can't see
> how to make it work for my example.  Hopefully someone out there can
> be of some help.
> 

The problem may be fixed by fixing up the error structure
because for light aniamls, you probably have an inflated error
variance so the animal component is trying to correct the inflation.

c 
> The problem appears to be for my animal or permanent environmental
> effects.  I have three measures per animal.
> 
> Steve Miller
> 
>
I hope this clarifies some things.
Arthur 
> 
> --
> Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml


<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
Arthur Gilmour PhD                    mailto: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 still free by anonymous ftp from pub/aar on ftp.res.bbsrc.ac.uk
    or point your web browser at ftp://ftp.res.bbsrc.ac.uk/pub/aar/ 

To join the asreml discussion list, send the message  
     subscribe
mailto:asreml-request@chiswick.anprod.CSIRO.au

To send messages to the list, mailto:asreml@chiswick.anprod.CSIRO.au

Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml

                        <> <> <> <> <> <> <>
"The Holy Scriptures are able to make you wise for salvation
   through faith which is in Christ Jesus.
 All Scripture is given by inspiration of God, and is profitable for
   doctrine, for reproof, for correction, for instruction in righteousness,
 That the man of God may be complete,
   thoroughly equipped for every good work."   II Timothy 3:15-17.
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>


--
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml