dear all - below is a reply from arthur to a message sent the other day.
---------- Forwarded Message ----------
You can make a judgement yourself about whether you need to do a few more
iterations when 'LogL converge, Parameters not converged' appears.
In small analyses where there is not much information, the LogL may not
change much for quite larege changes in the parameters. It also depends
on whether it is a final analysis or an interim analysis.
You can use the !EXTRA 3 qualifier to make ASREML do
3 (in this case) more iterations after it has met the LogL convergence
Sometimes it meets the LogL criterion only becasue some parameters trying
to go negative have temporarily been held positve. Then it is appropriate
to do a few more
You should also check the parameters unless some have been fixed by the
program but aftyer other things have converged, the fexed parameters need to
relaxed and a better solution found.
So, if you frequently get the message 'parameters not converged',
I suggest you routinely include !EXTRA 2 and see if that overcomes the
especially in small jobs when it take longer to review the output than it
does to do a few more iterations.
The 'LogL converged' message indicates that two
successive LogL values were very close. It does not really mean
a maximum has been found.
'Paramerers not converged' indicates that some parameters changed more than
1 % in the final
iteration. This might be a very small change on a parameter close to zero
it hardly matters, or a bigger change on a larger parameter when it may well
Routine use of !EXTRA 2 will reduce the frequency of 'Parameters not
You may want to rerun the model even when 'LogL converged' if some of
the parameters have been fixed (B) by ASREML.
You always need to
make sure you understand the model. If not sure,
try thinking of an alternative parameterization which can confirm the
(There are usually several ways of fitting a particular model).
Finally, I'm unclear as to your final question. Normally
parameters converged will mean the variance components have converged.
The only exception I can think of is an overparameterised model where
there is no informationto estimate a particular parameter. Then the
parameters will depend on the starting values. E.G. your model includes
a correalation between two terms but there are no observations
upon which to estimate the correlation.
I trust this helps.
From: C. Balde
Sent: 5/13/00 11:49 PM
Subject: !Continue option
When running ASREML sometimes one get the message 'LogL conerged,
Parameters not converged'. I thinks the case is ASREML has found a
maximum loglikelihood value, but in the las iteration the variance
components varied over a prefixed percentage (0%, and in the % column in
asr file this value is 1 or more). But running again the model with the
'!continue' option the LogL value stays the same and the parameters
could converge (sometimes one need re-running the model more than 1
time). In addition to this when one re-run the model after the
parameters have converged, the variance components change (not much).
So the questions are:
1. When parameters don't converge at first time (but LokL do), must one
re-run the model with the !continue option still they do ?
2. When the model converge (Logl and parameters), must one re-run the
job still variance compobnentes don't change ?
Ingeniero Forestal (e)
Universidad de Chile
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml