From: Ian Brooks <ianb@wolfram.com>
  To  : rasmb@alpha.bbri.org
  Date: Tue, 02 May 2000 10:55:27 -0500

RE: linear approximation for error estimates

One diagnostic that might be of use here is to check the intrinsic 
curvature and parameter effects curvature of the model at the best fit 
parameters.  While this is still no substitute for determining the 
confidence limits directly through profiling, bootstrapping or monte-carlo, 
it does tell you immediately whether there are likely to be serious 
deficiencies in the linear approximation intervals.

 From my work on measuring confidence intervals with Preston I know that 
for sed eq data the number of data points is important in determining the 
degree of non-linearity of the confidence intervals.  Models that are well 
behaved when given a full dataset can appear wildly non-linear when you 
only use every 3rd or 4th value.  (We used the profiling method to 
determine the confidence limits empirically).


	Ian

Ian Brooks Ph.D.
Applications Developer
Wolfram Research, Inc.
100 Trade Center Drive
Champaign, IL 61820
217-398-0700
217-398-0747 (fax)
ianb@wolfram.com