Anne,
The 'floating invalid' problem may occur when G matrices are non-
positive definite which lead to negative residual sum of squares. Don't
know the direct solution for this problem but trying the following may
help: i) Identify variance components close to being negative (or
singular) and drop them. ii) Check covariances trying to go out of
parameter space - change model if so iii) Use !continue option on the
data file to stabilize the program, even with correct starting values ?
Other comments on causes and fixes for 'Floating Invalid' problem in
general ??
Haja
From: "Anne Winkelman" <awinkelman@lic.co.nz>
To: asreml@chiswick.anprod.csiro.au
Date sent: Thu, 2 Nov 2000 09:16:39 +1200
Subject: ASREML crashes
>
>
> Hi,
>
> I am using asreml to run random regressions using Legendre polynomials. All is
> well with the test examples and I am able to get the same solutions using asreml
> and dxmrr. Problems arise when I attempt to analyse actual data. Because of
> the computational demands associated with running rr models, I have started by
> analysing the data by season (i.e. year). I have 7 years of data. Each year has
> been analysed using the same model. There are two random effects in the model -
> one for the genetic effect and one for the permanent environmental effect. The
> second degree polynomial is included in the model as animal.phi and perm.phi
> for the genetic and permanent environmental effects, respectively (see the
> program below). The problem is that for some years the program crashes and for
> others the program runs properly and produces reasonable results. The program
> gets as far as this before it crashes :
>
> Thu Nov 2 08:57:37 2000 512.00 Mbyte Unix rrprogram
> PEDIGREE [rrped97es.txt ] has 40156 identities, 119809 Non zero elements
> Univariate analysis of yobs
> WARNING - Fewer levels found in category than specified
> WARNING - Fewer levels found in sire than specified
> WARNING - Fewer levels found in dam than specified
> WARNING - Fewer levels found in tdgroup than specified
> 11 19 0.000000000000000E+000 0
> Data record order assumed correct: not reordered
>
> In other words, it doesn't get to the point where it tells you how many
> equations have been formed and how many of those are in the 'dense' section.
>
> The error message I get when the program crashes is:
>
> forrtl: error (65): floating invalid
> /usr/bin/sh: 19072 Abort - core dumped
>
> The program crashes regardless of the starting values used for the coefficients
> of the polynomials. However, if I change the model so that the random
> regression is included for only the animal effect or only the permanent
> environmental effect (i.e. include animal.phi perm OR animal perm.phi) , then
> the program does not crash.
>
> Thanks in advance for any suggestions on how to fix this problem,
>
> Anne Winkelman
> Livestock Improvement
> Hamilton
> New Zealand
>
> The command file I am using is:
>
> rrprogram.as Random regression: Legendre polynomials
> category 6
> animal !P
> sire 1000 !I
> dam 11992 !I
> tdgroup 2000 !I
> age 5 !I
> covar1
> covar2
> covar3
> covar4
> covar5
> yobs
> dim
> phi 3 !F
> rest 3 !F
> perm 11992 !I
> rrped97es.txt !MAKE # pedigree file: animal, sire, dam
> rrphi97s.txt !MAXIT 40 # sorted by category, animal
> yobs ~ age covar1 covar2 covar3 covar4 covar5 phi,
> !r animal.phi perm.phi !f tdgroup !f mv
> 3 1 2 # R header line <sites> <dimensions> <G-structures>
> 16691 0 0 !s2=0.06499 # R str <levels> <0 if sorted> <structure>
> 17898 0 0 !s2=0.04141
> 10324 0 0 !s2=0.03222
> animal.phi 2 # G structure <model_term> <dimensions>
> animal 0 AINV
> 3 0 US !+6 # <order> 0 (sortc) <model> <init parm> <!qualifier>
> 0.02737 # v11
> 0.006671 0.005080 # c21 v22
> -0.002988 -0.0006959 0.0009269 # c31 c32 v33
> perm.phi 2 # perm environment <model_term> <dimensions>
> perm 0 I
> 3 0 US !+6 # <order> 0 (sortc) <model> <init parm> <!qualifier>
> 0.1276 # v11
> 0.006377 0.02221 # c21 v22
> -0.01276 -0.007385 0.01171 # c31 c32 v33
>
>
>
>
> ___________________________________________________________________________
> THE FOLLOWING WARNINGS ATTACH TO THIS COMMUNICATION
> This material is intended for the named recipient only and has been created
> solely for the purpose and function of Livestock Improvement Corporation Limited
> and may contain information that is subject to the Privacy Act / Confidentiality
> / Privilege / Copyright. Any person other than the named recipient is prohibited
> from retaining / disclosing / copying / distributing / using all or part of this
> material.
>
>
> --
> Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml
--
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml