[R] nlme and pnlsTol

[Dieter Menne]

Dear nlme-lovers,

I do a fit fo a 3-parameter fit to physiological data with nlme:

EmptD<-deriv(~(vol+slope*t)*exp(-t/tempt)...

The approach described in Pinheiro/Bates by using nlsList
as a first approximator was somewhat unstable (many NA), but
a direct fit converges quickly and fits look good:

contr=nlmeControl(pnlsTol=0.1)
v.nlme=nlme(v~EmptD(t,vol,slope,tempt),data=acg,
          verbose=F,control=contr,
          fixed=list(tempt+slope~treat+what,vol~1),
          random=vol+tempt+slope~1,
          start=c(150,-50,-50,-5,3,5,14,3,645))

However, I had to use a rather large value of pnlsTol of 0.1 go achieve
convergence.

Questions:

1) What is the price I have to pay when using so large pnlsTol? Could I miss
any more distant optima?

2) When using an even larger value of pnlsTol=0.3, AIC is lower by 2 and
standard errors are somewhat lower. Can I trust this result?

Dieter Menne