[R] Regarding nls()

[Karl Ove Hufthammer]

Spencer Graves:

> Bates' condemnation of R^2 has merit, but I would not go as far as
> he did in the comment cited below (dated 13 Aug 2000).  A standard
> definition of R^2 is as follows:
> 
> R^2 = (1 - var(prediction error) / var(obs)).
> 
> I can name several different ways of getting a negative R^2 in
> this case.  When that happens, it says the model is worse than useless,
> and you would be better off using the training set mean.
> 
> If I have an audience who wants an R^2 in an application where it
> is not clear what it even means, I try to briefly explain some of the
> difficulties while asking what question they are trying to solve using
> R^2.  Their answers will help me make a recommendation, which may
> include selecting which of the possible generalizations of R^2 to use.

I would like to recommend the following two articles on R²:

Model Comparisons and R²
    Richard Anderson-Sprecher 
    The American Statistician, Vol. 48, No. 2. (May, 1994), pp. 113-117. 
    http://www.jstor.org/stable/2684259

Cautionary Note about R²
    Tarald O. Kvalseth 
    The American Statistician, Vol. 39, No. 4, (Nov., 1985), pp. 279-285. 
    http://www.jstor.org/stable/2683704

-- 
Karl Ove Hufthammer