[R] standard errors for predict.nls?

[Prof Brian Ripley]

On Mon, 3 Nov 2008, Ben Bolker wrote:

> Prof Brian Ripley wrote:
>>> Christoph Scherber <Christoph.Scherber <at> agr.uni-goettingen.de>
>>> writes:
>>>
>>>>
>>>> Dear all,
>>>>
>>>> Is there a way to retrieve standard errors from nls models?
>>>> The help page tells me that arguments
>>>> such as se.fit are ignored...
>>>>
>>>> Many thanks and best wishes
>>>> Christoph
>
>> In general using the delta method (which is I guess what you mean, local
>> linearization via derivatives) is nowhere near accurate enough to be
>> useful.  That's why it has not been done on several occasions in the past.
>> If you think it might be, see ?delta.method in package alr3.
>>
>> I would suggest using simulation/bootsrapping to explore the uncertainty.
>> There is an example in MASS of doing so (and that illustrates some of
>> the pitfalls).
>
>  Hmmm.  By an example, do you mean an example of using bootstrapping to
> explore uncertainty in general, or of using bootstrapping to get
> standard errors of predictions from nonlinear regressions?  I looked
> through my copy of MASS (4th ed.) and found only section 5.7
> (bootstrapping in general) and chapter 8 (nonlinear and smooth
> regression, esp. p. 225 "bootstrapping" for getting bootstrap c.i.'s on
> parameter estimates).  I didn't find anything *specifically* covering
> s.e./c.i. for nls predictions, but maybe that's not what you meant.

I meant the example on p.225 on bootstrapping a nls fit (and that you 
needed to bootstrap residuals in some cases).  You can use almost 
identical code to set s.e./c.i. for nls predictions.

>  And yes, I meant "delta method" rather than "delta function" in my
> original post.  Oops.
>
>  I might add something quick/dirty/naive to the wiki giving
> some examples of delta method/bootstrap approaches ...
>
>  If there is no intention to add confidence interval calculation
> to predict.se in the foreseeable future might I suggest that the details
> under "Value" as to what "se.fit" will do when it is implemented be
> removed? (And perhaps even a statement to the effect [as you say
> above] that delta method is considered unreliable?)  As written it's a
> bit of a tease ...

I didn't write that ... and its author might have other opinions.

>  cheers
>    Ben Bolker
>

-- 
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