[R] non linear modelling with nls: starting values

[Gabor Grothendieck]

You could try fitting several cases:

- fit the reduced model with only one exp term, i.e. one of the alphas is 0
- fit the model with both alpha's constrained to be positive &
sufficiently away from 0
- fit the model with both alpha's negative & sufficiently away from 0
- fit the model with one positive and one negative and sufficiently away from 0

On 9/18/06, Sebastian P. Luque <spluque at gmail.com> wrote:
> On 18 Sep 2006 17:55:24 +0200,
> Peter Dalgaard <p.dalgaard at biostat.ku.dk> wrote:
>
> > "Gabor Grothendieck" <ggrothendieck at gmail.com> writes:
> >> Here are some approaches:
>
> >> - we only have 4 parameters so just use grid search to get starting
> >> values as in:
>
> >> https://stat.ethz.ch/pipermail/r-help/2005-September/079617.html
>
> >> - there are singularities near beta_1 = beta_2 and near alpha_1 = 0 and
> >> near alpha_2 = 0 so reparameterize and use the upper and lower bounds
> >> to avoid those regions.  You could try a separate reduced model for
> >> those.
>
>
> > Or just use SSbiexp and reparametrize (it is exactly the same model)
>
> Thank you Gabor and Peter.  With your help, I was able to fit the model
> using SSbiexp as Peter suggested.  However, the dependent variable spans
> both negative and positive values, and the negative values cause the
> fitting procedure to fail with a NA/NaN/Inf error.  This might be related
> to the singularities that Gabor mentioned.  If the dependent variable is
> shifted upwards, so all values are positive, the model is fit without
> problems.  Any ideas on how to get around this issue without shifting the
> data, so that the parameters are expressed in the original scale?
> Hopefully it's possible to do it using SSbiexp.  Thanks again.
>
>
> Cheers,
>
> --
> Seb
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>