Re: LogL Converged, Parameter Not Converged

On 10/14/2009 09:11 AM, KRISHNAMURTHY, SUBASH [AG/1000] wrote:
> Dear Dr Gilmour,
>
> I am still trying to understand how this !CONTINUE works and I apologize, if in the process, I am asking the same questions over and over.
>
> It seems to me that for the qualifier !CONTINUE to work, the .rsv file must be sitting there. Does this mean that I have to run the program twice each time I see this message of Parameter Not Converged?
There is no blanket recommendation because is really depends both your
model and data. As Victor said, you can try using more iterations by
default but that may just delay the message. Really if the number of
iterations gets very large it indicates some issue that you need to
address such as bad starting values, badly scaled data, or a poor model
(like having variance components with zero estimates).

As in your case, that 'B' means something! The problem appears to be
that you lack the necessary data to estimate your model as your variance
estimate for REP is zero. Really you should fit a more suitable model
for your data as you need to think about what your model actually
implies (I have no idea about meaning of REP and REP.GROUP1 in your
model). I do not know how any program is going to separate between REP
and REP.GROUP1 as defined as REP.GROUP1 appears to have no structure.
Rather you probably should just fit REP.GROUP1 alone (which is what the
output implies) as it may mean the same thing for your desired model.

> If that is the case, is there a way to tell ASREML to do this automatically.
>
> Also, is it safe to use the outputs like the Least Squared Means (BLUPS), standared erros of the LSM (BLUPS) etc., with this warning message?
>
Technically no because you have not converged and thus do not meet the
theoretical requirements. Also, you appear to have set a variance
component on the boundary condition so your original model is not the
same the one that was estimated. You may not find a difference in the
fixed effects but you should find a difference in the random effects.
But you may find similar values if you drop REP from the model.

Bruce

>
> Best regards,
>
> -Subash
> 314-694-6364
>
>
>
> -----Original Message-----
> From: ASReml users discussion group [mailto:ASREML-L_at_DPI.NSW.GOV.AU] On Behalf Of Arthur Gilmour
> Sent: Tuesday, October 13, 2009 12:15 AM
> To: ASREML-L_at_DPI.NSW.GOV.AU
> Subject: Re: LogL Converged, Parameter Not Converged
>
> Dear Subash,
>
> !CONTINUE means CONTINUE from the point were the last run terminated
> by retrieving the variance parameters held in the .rsv file and using
> them
> as initial values.
>
> !CONTINUE does not change anything else.
>
> If you had !MAXIT 300 and iteration stopped at some point before 300
> iterations, then I would expect it would stop at the same point if you
> had !MAXIT 500 if that was the ONLY change to the code.
>
> However, if you had !MAXIT 300 and it stopped after 300 iterations,
> I would suspect it stopped because it reached 300, in which case if you
> had !MAXIT 500, it would keep going.
>
> The default value of the stepsize parameter does depend slightly on the
> MAXIT parameter so there is a small chance that the iteration sequence
> might be slightly different.
>
> I am surprised that the example below would take more than 300
> iterations to converge. It seems to be 2 reps of 26 treatments
> divided into 2 groups, in which case it would be fairly orthogonal.
> And there is no other random term for REP.GROUP1 to oscillate with,
> especially when REP gets fixed!
>
> But it is a case where the variance component is small (Comp/SE is .32)
> so a small change can represent a largish %change.
>
>
> On Mon, 2009-10-12 at 09:55 -0500, KRISHNAMURTHY, SUBASH [AG/1000]
> wrote:
>
>> Dear Dr. Gilmour,
>>
>> This certainly helps.
>>
>> How long does ASREML continue to iterate with the !CONTINUE qualifier? The % change for one of the parameters was -1 as shown below;
>>
>> Source Model terms Gamma Component Comp/SE % C
>> REP 2 2 0.101193E-06 0.310276E-07 0.00 0 B
>> REP.GROUP1 4 4 0.485443E-01 0.148846E-01 0.32 -1 P
>> Variance 52 27 1.00000 0.306618 3.50 0 P
>>
>> For this dataset the model parameters converged with the !CONTINUE qualifier, as confirmed by Dr Cullis.
>>
>> I am still confused with the fact that the model parameters converged with 300 iterations and not with 500 iterations.
>>
>> Best Regards,
>> -Subash
>> 314-694-6364
>>
>>
>> -----Original Message-----
>> From: ASReml users discussion group [mailto:ASREML-L_at_DPI.NSW.GOV.AU] On Behalf Of Arthur Gilmour
>> Sent: Sunday, October 11, 2009 5:03 PM
>> To: ASREML-L_at_DPI.NSW.GOV.AU
>> Subject: Re: LogL Converged, Parameter Not Converged
>>
>> Dear Subash,
>> Further to Brian's comments:
>>
>> 1) Having run 100 iterations, if you increase MAXIT to 300 but
>> otherwise leave the job unchanged, the first 100 iterations of the new
>> run will usually be identical to the first run (except the STEPSIZE
>> might be slightly different).
>>
>> However, with !CONTINUE, ASReml reads back the .rsv file and so
>> starts the second run from where the first finished. Brian alluded to
>> the fact that for some XFA models with PSI going to the boundary,
>> CONTINUE can allow those PSI to be formally fixed at zero which may help
>> the rest of the iteration.
>>
>> 2) The message 'LogL Converged, Parameters Not converged' arises when
>> the likelihood value is changing very slightly (less than the
>> convergence criterion' but when ASReml looks at the % change of the
>> parameters, some have changed more than 1 %.
>>
>> So, When I see that message, I do two things:
>> a) I look for which parameter has not converged: it will have % value
>> greater than zero in the variance component table of the .asr file.
>> Typically it is a very small component which is effectively zero.
>> b) Occasionally it is a pair of terms and the parameter estimates are
>> oscillating. This can be recognised by looking at the parameter history
>> reported in the .res file.
>>
>> 3) Failure to converge quickly occurs sometimes with US variance
>> structures when they are constrained Positive Definite, but the maximum
>> of the REML likelihood function occurs at a point outside the
>> constrained parameter space. Then ASReml will use EM steps to approach
>> the constrained REML solution, but this is notoriously slow.
>>
>> 4) XFA2 and XFA3 models are also sometimes slow to converge.
>>
>> 5) I do not just increase MAXIT unless I can see the job is converging
>> and just needs some more iterations. Often I will decide that while it
>> has not converged according to the criterian written into the program,
>> it has converged sufficient for my purposes.
>>
>> 6) If a few more iterations are required, I would use !CONTINUE
>> rather than waiting for ASReml to repeat a whole lot of work it has
>> already done once.
>>
>> I trust this helps
>>
>> On Sun, 2009-10-11 at 20:43 +1100, brian.cullis_at_INDUSTRY.NSW.GOV.AU
>> wrote:
>>
>>> Dear subash
>>> I am trying to fit a linear mixed model for a Group Unbalanced Block
>>> design, and I get a warning message in the .ASR file as “LogL
>>> Converged, Parameters Not converged”. I increased the number of
>>> iterations from 30 to 100 and still got the same warning message.
>>> Then, I changed it to 300 and the model converged. So I thought it is
>>> safe to use a big number for !MAXIT and changed it to 500. I got the
>>> same warning message which, I thought, doesn’t make any sense for
>>> parameters to converge at 300 iterations but, not with 500 iterations.
>>> What could be the reason? The other question I had is, Should I use !
>>> CONTINUE option instead of increasing the number of iterations. How is
>>> it different from increasing the number of iterations using !MAXIT?
>>> What impact does it have on the run time?
>>>
>>> 1. This message prints out in some situations when a parameter may be
>>> small or poorly estimated - in the sense that the data may not support
>>> the model you are fitting. If you would like to sanitise the data for
>>> confidentiality reasons I would be most happy to try to run it for you
>>> to see what I can make of it
>>> 2. continue is often a more reasonable way to achieve convergence as
>>> some parameters may get fixed at boundary values etc
>>>
>>>
>>> warm regards
>>>
>>> Brian Cullis|Research Leader, Biometrics&
>>> Senior Principal Research Scientist |
>>> Industry& Investment NSW | Wagga Wagga Agricultural Institute | Pine
>>> Gully Road | Wagga Wagga NSW 2650 |
>>> PMB | Wagga Wagga NSW 2650
>>> T: 02 6938 1855 | M: 0439 448 591 | F: 02 6938 1809 | E:
>>> brian.cullis_at_industry.nsw.gov,au
>>> W: www.industry.nsw.gov.au |
>>>
>>> Visiting Professorial Fellow
>>> School of Mathematics and Applied Statistics
>>> Faculty of Informatics
>>> University of Wollongong
>>>
>>> Professor,
>>> Faculty of Agriculture, Food& Natural Resources
>>> The University of Sydney
>>>
>>> Adjunct Professor
>>> School of Computing and Mathematics
>>> Charles Sturt University
>>>
>>>
>>>
>>>
>>>
>>>
>>> This message is intended for the addressee named and may contain confidential information. If you are not the intended recipient, please delete it and notify the sender. Views expressed in this message are those of the individual sender, and are not necessarily the views of their organisation.
>>>
Received on Thu Oct 14 2009 - 11:17:01 EST