Re: fitting random regressions
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: fitting random regressions



Hi,
I would be inclined to look at the response profiles and be sure that 
the profiles are approximately polynomial. Cubic splines are on 
alternative, of course, and perhaps could be quite useful if these 
profiles suggested that there was some nonlinearity

usage is straighforward

y ~ mu age !r spl(age) tree tree.age tree.spl(age)

etc

see Verbyla, Cullis, Kenward and Welham for details




On Fri, 13 Mar 1998, 
Luis Apiolaza Z. wrote:

> Hi,
> 
> First thanks to everyone for your help. In summary:
> - Using either the !REPEAT command or a separate pedigree file avoided
> the error when reading the pedigree.
> - the block.ide(age) didn't work, but using another factor coding for
> the ages worked fine.
> - Arthur suggested several steps for modelling random regressions. As I
> understood Arthur's code, the within individual errors were assumed to
> be independent, i.e. with a diagonal error matrix (in this case a US
> with zero off diagonals), and unstructured covariance matrix for both
> the tree.pol(age,2) effect and the block.age effect.
> 
> Block was not significant, so I took it out from the model.  After
> several attempts the program converged, but the variance components for
> the error matrix were not significant, while all the terms of the US
> matrix for the polynomial were significant.  Because sometimes Eucalypts
> data sets give funny results I tried with a different data set. This was
> a study of the behaviour of stem diameter and wood density with age,
> comprising 16 years (equally spaced) of measurements. This is just one
> family with 188 trees.
> 
> I tried just:
> 1- diam~mu age !r tree
> 
> 2- diam~mu age !r tree !f mv
> 1 2 !STEP .1
> 188
> 16 0 DIAG 195 200 ...
> 
> 3- diam~ pol(age,2) !r tree.pol(age,2) !f mv
> 1 2 !STEP .1
> 188
> 16 0 DIAG 195 200 ...
> 
> tree.pol(age,2) 2
> tree 0
> pol(age,2) 0 US ...
> 
> And again, the error components were non significant while the
> polynomial was highly significant. So I think that maybe the assumption
> about the error is wrong. Maybe it's worth to try with an AR(1)
> structure or alternative structures that reflect the autocorrelation of
> the measures within individuals.
> 
> If you think of a different option just send me a line. Thanks again,
> 
> 
> Luis
> 
> &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> Luis Apiolaza
> Institute of Veterinary, Animal and Biomedical Sciences
> Massey University
> Palmerston North
> New Zealand
> L.A.Apiolaza@massey.ac.nz
> 
> "Quote me as saying I was misquoted " - Groucho Marx
> &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> 
> 
> 
> 
> 
> 

...............................................................................
      Brian Cullis                       Tel: 02 6938 1855
      NSW Agriculture                    Fax: 02 6938 1809
      Wagga Agricultural Institute     email: brian.cullis@agric.nsw.gov.au
      Pine Gully Rd
      WAGGA WAGGA  NSW 2650