Re: Amazing REML
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Amazing REML

Dear Luc and others,

Luc's example of REML (ASREML) getting an answer for a covariance
when there is no information is an instructive one.

His data consisted of weights of surviving animals, missing values
for non surviving animals, and he was doing a bivariate analysis
of weight with survival.

I report what ASREML produces for a range of starting values.

The DATA file is

14.53 1
10.12 1
16.30 1
15.54 1
13.36 1
12.77 1
14.06 1
12.23 1
* 0
* 0
* 0
* 0

The basic ASREML job is

Luc Janss example                # Header
 W S                             # data fields
ljanss.dat                       # data file name
 W S ~ Trait                     # bivariate linear model
1 2
12                               # 12 records
2 0 US 3 .2 .3                   # Error variance matrix (input)

For the two cases Luc reported, the LogL values from ASREML
were -8.476 and -9.01  which agree with his values except for a factor
of -0.5.

I ran ASREML to convergence with a range of covariances.  In every case
the LogL converged to the same value but the WT variance
and the Covariance estimate changed depending on the starting value.

The cases were

Initial       W         Fitted       S
Covariance   Variance   Covariance   Variance

-.1          3.8644      -.0808      0.24242
0.0          3.8375      0.0000      0.24242
0.05         3.8442      0.0404      0.24242
0.1          3.8644      0.0808      0.24242
0.2          3.9452      0.1616      0.24242
0.4          4.2685      0.3232      0.24242

All these final models had the same LogL and the determinant of the matrix
is the same as well (0.9303).

If ASREML detects a singularity in the Average Information matrix,
it will not update the corresponding parameter but will flag it
as having no information.  This typically occurs with variances
but for covariances, a singularity does not appear in
the information matrix so all parameters are updated.  

In this case, there is no maximum value to the likelihood; just
a ridge of maximum likelihood  and so all solutions have
equal likelihood.   I.e.  ASREML is combining the information
in the model (and initial values) with the data but once it
makes the variance of W compatible with the Covariance,
it has no information with which to further change the covariance.

In other words, if you supply a covariance and there is no
information in the data on the value of that covariance, REML
cannot contradict your initial value except so far as it
is inconsistent with the variance of W in this case.  

This agrees with Dan's conclusion (erratum)

I trust thus explanation is helpful.


Arthur Gilmour PhD                    email:
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
    or point your web browser at 

To join the asreml discussion list, send the message  

The address for messages to the list is

Asreml mailinglist archive:

                        <> <> <> <> <> <> <>
"For the wages of sin is death,
  but the gift of God is eternal life in Christ Jesus our Lord." Romans 6:23