[R-sig-Geo] likfit problem in latest geoR and latest R

[Paulo Justiniano Ribeiro Jr]

Hi Ruben

I run a Linux machine with:

R : Copyright 2005, The R Foundation for Statistical Computing
Version 2.0.1 Patched (2005-01-25), ISBN 3-900051-07-0

from which I get
> ml<-likfit(mc96,cov.model="gaussian",ini.cov.pars=c(1.1,30000),nugget=0.9,lambda=0,fix.lambda=TRUE,hessian=TRUE)

> ml
likfit: estimated model parameters:
       beta       tausq     sigmasq         phi
"    6.706" "    1.072" "    1.355" "30000.000"

likfit: maximised log-likelihood = -2434

So... it runs but ... the algorithm is not moving away from the initial
value for $\phi$.

So dividing the coordinates by 1000 we get something more resonable for
this data:

> mc96$coords <- mc96$coords / 1000
> ml<-likfit(mc96,cov.model="gaussian",ini.cov.pars=c(1.1,30),nugget=0.9,lambda=0,fix.lambda=TRUE,hessian=TRUE)
> ml
likfit: estimated model parameters:
    beta    tausq  sigmasq      phi
"6.6014" "0.7072" "1.2788" "6.3682"

likfit: maximised log-likelihood = -2422

This shows there is a numerical problems in finding the maximum due to the
vary diferent scales of the parameters.
There are 2 possible solutions:
1. re-scale the coordinates (as above)
2. pass arguments to optim() using the "..." mechanism

You shoud inspect thins more carefully trying different initial values to
be sure about convergence

Cheers
P.J,



On Thu, 27 Jan 2005, Ruben Roa wrote:

> > -----Original Message-----
> > From: Paulo Justiniano Ribeiro Jr [mailto:paulojus at est.ufpr.br]
> > Sent: 26 January 2005 23:07
> > To: Ruben Roa
> > Cc: r-sig-geo at stat.math.ethz.ch
> > Subject: Re: [R-sig-Geo] likfit problem in latest geoR and latest R
> >
> >
> > Ahhhh
> > it is just a typo in you original call
> >
> > the argument is: fix.lambda
> > and not        : fixed.lambda
>
> Hi Paulo,
> No, i made the typo in the mail message sent to the R-sig-Geo list but
> not in R, so the problem with likfit in R 2.0.1 Patched is still there,
> at least for my Windows XP machine.
> I installed R 2.0.1 not the patched version and then i run the command
> again with the same data in the same computer with geoR 1.5-4 and it run
> perfectly well.
> Cheers
> Ruben
>
>

Paulo Justiniano Ribeiro Jr
LEG (Laboratório de Estatística e Geoinformação)
Departamento de Estatística
Universidade Federal do Paraná
Caixa Postal 19.081
CEP 81.531-990
Curitiba, PR  -  Brasil
Tel: (+55) 41 361 3573
Fax: (+55) 41 361 3141
e-mail: paulojus at est.ufpr.br
http://www.est.ufpr.br/~paulojus