Re: convergence question

Dear Annette,

You can look in the .res file to see a little clearer what might be
happening
as the sequence of parameter values should be reported there, and you
might
get some extra decimal places on the LogL.

You might well conclude that the analysis has effectively converged
given the large amount of data, even though it has not passed the test in
ASReml.

In the next version, big LogL values are reported with an offset so that
extra precision is visible.

Best wishes in the Name of Christ Jesus,
the Saviour of all who believe.

Arthur Gilmour PhD
Principal Research Scientist (Biometrics)
NSW Department of Primary Industries
Orange Agricultural Institute, Forest Rd, ORANGE, 2800, AUSTRALIA

mailto:Arthur.Gilmour_at_agric.nsw.gov.au
fax: 02 6391 3899; 02 6391 3922 Australia +61
telephone work: 02 6391 3815; home: 02 6364 3288;
mobile: 04 2764 3288 (no reception at home)
http://www.agric.nsw.gov.au/reader/17263

ASREML website: http://www.asreml.co.uk/
The ASReml discussion group ASREML-L_at_agric.nsw.gov.au
has reactivated. To join, mailto:arthur.gilmour_at_agric.nsw.gov.au

Cookbook: http://uncronopio.org/luis/asreml_cookbook.html
                        <> <> <> <> <> <> <>

<><><><><><><><><><><><><><><><><><><><><><><><><>

Annette Winkelman <awinkelman_at_LIC.CO.NZ>
Sent by: ASReml users discussion group <ASREML-L_at_AGRIC.NSW.GOV.AU>
09/12/2004 09:58 AM
Please respond to ASReml users discussion group

        To: ASREML-L_at_AGRIC.NSW.GOV.AU
        cc:
        Subject: convergence question

ASReml Users,

I am running RR analyses of test day records and have found that the
log-likelihood is not converging. The data set contains 170,000+ cows and
650,000+ records. The manual indicates that convergence is presumed when
the log-likelihood changes less than 0.002* current iteration number and
the individual variance parameter estimates change less than 1%.

After 10 iterations the LogL did not converge so the analysis was
continued
using the 'C' option on the command line. After ten more iterations the
output again contained the message: "Warning: LogL not converged".
However, the output showed no change in the LogL after these 10 additional
iterations:

 Restarting iteration from previous solution
   1 LogL=-738973. S2= 1.6643 636602 df
   2 LogL=-738973. S2= 1.6643 636602 df
   3 LogL=-738973. S2= 1.6643 636602 df
   4 LogL=-738973. S2= 1.6643 636602 df
   5 LogL=-738973. S2= 1.6643 636602 df
   6 LogL=-738973. S2= 1.6643 636602 df
   7 LogL=-738973. S2= 1.6643 636602 df
   8 LogL=-738973. S2= 1.6643 636602 df
   9 LogL=-738973. S2= 1.6643 636602 df
  10 LogL=-738973. S2= 1.6643 636602 df

Additionally, the % change in parameters at the last iteration was 0 for
all parameter estimates.

Is a convergence criterion of "0.002* current iteration number"
appropriate
given the size of the likelihoods shown above?

Anne

___________________________________________________________________________
This message, including attachments, is confidential. If you are not the
intended recipient, please contact us as soon as possible and then destroy
the message. Do not copy, disclose or use the contents in any way.

The recipient should check this email and any attachments for viruses and
other defects. Livestock Improvement Corporation Limited and any of its
subsidiaries and associates are not responsible for the consequences of
any
virus, data corruption, interception or unauthorised amendments to this
email.

Because of the many uncertainties of email transmission we cannot
guarantee
that a reply to this email will be received even if correctly sent. Unless
specifically stated to the contrary, this email does not designate an
information system for the purposes of section 11(a) of the New Zealand
Electronic Transactions Act 2002.

This message is intended for the addressee named and may contain
confidential information. If you are not the intended recipient or
received it in error, please delete the message and notify sender. Views
expressed are those of the individual sender and are not necessarily the
views of their organisation.
Received on Thu Dec 09 2004 - 11:03:01 EST