The error is rather in _your_ log-likelihood: you are going to have to try
much harder to avoid underflow/overflow.
Avoiding log(exp(a1*x1+a2*x2)) would be a start. You might begin to
appreciate why R's d*** and p*** functions have a `log.p' argument.
Better starting values would help, probably a lot. Also read about
scaling problems in ?optim.
On Wed, 1 Sep 2004, Olympio T. Neto wrote:
> Friends
>
> I'm trying fit a survival model by maximum likelihood estimation
> using this function:
>
> flver=function(a1,a2,b1,b2)
> {
>
>
lver=-(sum(st*log(exp(a1*x1+a2*x2)))+sum(st*log(hheft(exp(b1*x1+b2*x2)*t,f.heft)))
>
-(exp(a1*x1+a2*x2)/exp(b1*x1-b2*x2))*sum(-log(1-pheft(exp(b1*x1+b2*x2)*t,f.heft))))
> }
> emv=mle(flver,start=list(a1=0,a2=0,b1=0,b2=0))
>
> where hheft and pheft are functions defined in polspline package.
> My variables are:
> t: time
> st: censor
> x1 and x2: Covariates
>
> I don't find any estimation because an error occurs:
>
> Non finite/finite difference value [0].
>
> Can I treat this error? I try using "SANN" method and this error
occurs
> too. There is another alternative?
(Unlikely, given where in the code this occurs. You haven't even copied
it correctly!)
--
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595