On 13 Jun 2003 04:24:44 -0700, [EMAIL PROTECTED] (dave martin)
wrote:

> [EMAIL PROTECTED] (akhan) wrote in message news:<[EMAIL PROTECTED]>...
> > Is there any statistical metod which can be applied to test whether a
> > non-linear model fit a dataset well or significantly?
> 
> Fisher's F-test can be used to quantitatively compare two models of
> data.
> 
> The F-test can answer the question "Are these two models significantly
> different at the X% level?".

This bothers me - that's not the way that I would describe
the question.  Down below:  Clearly the two models are *different*,
anyway, by a factor of 1/T.    Does one have a lower residual?

You can solve by least-squares, comparing predicted to 
observed, if the errors of prediction are of similar size 
across the range.  But that's the non-assumption, for 
non-linear regression,  isn't it?

> 
> One model can be constant; i.e. assume that data scatter is entirely
> random about the dependent variable mean & not dependent on the
> independent variable(s) at all.
> 
> Or one can see if adding another parameter to a fitting function makes
> a significant difference.
> 

Nested models; assessed by reduction of residuals
of least squares, say, or by increase of Likelihood.

> Or one can find out which of two arbitrary models best fits the data;
> this is very useful for comparing two theories.

AIC  and BIC  are keywords for looking up comparisons
of non-nested models.

> 
> As an example, alternate theories for chemical diffusivity D are:
> 
> (1) D=(K)*exp(-Q/T)
> (2) D=(K/T)*exp(-Q/T) 
> where K and Q are experimentally determined constants and T is
> temperature.
> 
> Given a set of (D,T) data the F-test can be used to see if there is a
> significant difference between these two models.
> 

The F-test is used, by theory and by custom, to test
models that are *nested*, using the difference in d.f.  as 
the numerator degrees of freedom.  Here, that d.f.  seems
to be zero....  
I'll try to find something in this library book I have on the topic.


This does seem to be a curious example.  
I think I would show folks  the F-test, assuming one 
d.f.,  as a 'demonstration' of the size of the difference.  

But these models are surely  *different*  in a way
that seems pretty strong.   I guess, I am accustomed 
to worrying more about whether a variable is *in*  a
model at all, instead of worrying about what form it takes.

Biomedical data with subjective reports is usually not 
so definitive, not so well-measured as to select between
models like that. 


-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization."  Justice Holmes.
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to