Dear TOre,
The message
WARNING: 1 variance structures were modified to make them positive definate
ASREML has not converged to a maximum likelihood solution.
Suggestion: rerun with -C option
means that while running the model, ASREML produced a US structure
which was not positive definite. It modified the result
to make it positive definite using a BENDing strategy on
the underlying correlation matrix. If this happens repeatedly
ASREML will fix the matrix but in this instance it only
happened once and subsequent updates seem
'to have produced positive definite matrices so normal convergence
has occurred.
It is likely to occur when then covariance is increased (in magnitude)
and oversteps the PD criterion.
Arthur
> From: tore.ericsson@skogforsk.se
> To: gilmoua@apollo.agric.nsw.gov.au
> Subject: RE: WARNINGs
> Date: Thu, 27 Jan 2000 10:37:16 +0100
> MIME-Version: 1.0
>
> Thank you very much, very good comments which we will penetrate further! But
> what about my question concerning !DOPARTs 1--3? / Tore
>
> -----Original Message-----
> From: Arthur Gilmour [mailto:gilmoua@apollo.agric.nsw.gov.au]
> Sent: Thursday, January 27, 2000 1:11 AM
> To: tore.ericsson@skogforsk.se
> Subject: Re: WARNINGs
>
>
> Dear ASREMLers,
>
> Tore has sent me an bivariate analysis which quckly went near the solution
> but then took maybe 100 iterations to slow move to a better solution.
>
> What highlighted his attention was the 'Parmaters not converged'
> message
>
> You might find my comments of interest.
>
>
> Dear Tore,
>
> I have looked at your part 4 analysis and it definititely is
> converging slowly. I have no definitive answer but suspect it
> is a combination of factors:
>
> The traits are C97 with 7780 observations and
> H97 with 4996 observations
>
> First, both genetic and error correlations start off small and negative
> and this probably makes the liklihood surface very flat.
>
> Secondly, parameterizing as unstructured matrices makes for
> high correlation between variance and covariance.
>
> 3. The data structure probably has a lot to do with it too.
>
>
> Ignoring C97, The variances for H97 are 8.44777 and 139.193
> [i.e. assumomg correlations are 0)
>
> After many iterations in the bivariate model, these increase to
>
> 8.57 and 143.97 [with correlations of -.133 (G) and -.183 (E)
>
> Source Model terms Gamma Component Comp/SE % C
> Residual US=UnStr 1 0.586802 0.586802 62.09 0 P
> Residual US=UnStr 1 -1.68269 -1.68269 -9.12 0 P
> Residual US=UnStr 2 143.967 143.967 47.92 0 P
> Tr.Mot CORRelat 2 -0.133090 -0.133090 -0.79 0 P
> Tr.Mot CORRelat 2 0.691743E-02 0.691743E-02 3.75 0 P
> Tr.Mot CORRelat 2 8.57340 8.57340 4.52 0 P
> Covariance/Variance/Correlation Matrix US=UnStructu
> 0.5868 -0.1831
> -1.683 144.0
> Covariance/Variance/Correlation Matrix CORRelation
> 0.6917E-02-0.1327
> -0.3231E-01 8.572
>
> I could get there reasonably quickly fitting
> !PART 05
> !EXTRA 20 #!CONTINUE
> C97 H97 ~ Trait at(Tr,1).c(Bbl) at(Tr,2).c(Blk) !r Tr.Mot -Tr.Fat
> and(Tr.Fat)
> 1 2 1 # R header line: sites dimensions G-structures
> 0 0 0 # struct 1 (n 0 I)
> Trait 0 US .58679 -1.55843 139.40 !GP # struct 2
> Tr.Mot 2 # G header line 1
> Trait 0 CORUH -.2003 .006915 8.461 !GP # struct 1
> 0 0 0 # struct 2 (Mot 0 I
>
>
> I.e. by increasing the covariances manually.
>
> Given the error covariance still increased in magnitude from the initial
> -1.55
> this should be rerun starting the covariance at -1.75 say.
>
> Notice that the residual variance has increased (4%) suggesting
> that the missing observations are not at random but are associated
> predominantly with one tail. This is supported by the movement of the mean
> down from 70.1902 in the univariate analysis to 68.5361 in the
> bivariate analysis. Thus the third of the data missing for H97
> had higher values of C97. This is actually what has happened as the mean
> C97 over all data is 4.996 but the mean C97 for those records with H97
> was only 4.425.
>
>
> So the convergence is slow because ASREML is projecting bigger variances
> and covariance for H97 based on the data distribution. I.e. both
> are increasing (in magnitude) together.
>
> This has been an interesting situation which I have not seen often.
> I trust the discussion is helpful.
>
> Arthur
> <><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
> Arthur Gilmour PhD mailto:Arthur.Gilmour@agric.nsw.gov.au
> Principal 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 list archive: http://www.chiswick.anprod.csiro.au/lists/asreml
>
> <> <> <> <> <> <> <>
> "Christ Jesus came into the world to save sinners" I Timothy 1:15.
> <><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
>
>
>
>
>
>
>
>
>
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
Arthur Gilmour PhD mailto:Arthur.Gilmour@agric.nsw.gov.au
Principal 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 list archive: http://www.chiswick.anprod.csiro.au/lists/asreml
<> <> <> <> <> <> <>
"Christ Jesus came into the world to save sinners" I Timothy 1:15.
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
--
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml