From: Brian M. Baker <baker@xtal200.harvard.edu>
  To  : rasmb@alpha.bbri.org <rasmb@alpha.bbri.org>
  Date: Wed, 26 Apr 2000 15:13:45 -0400

Re: NONLIN question

Plug plug plug...

I routinely use the "original NONLIN" (for lack of a better term) in
all its gory FORTRAN detail. It requires that you write the fitting
function in fortran and then compile it with the original source code.
This is its greatest strength, as well as weakness. Strong, because you
can code any model you can imagine, so for example you can make
anything you want global or local, you can analyze as many related (or
unrelated!) datasets as you want, you're not limited to analytical
solutions for variables (e.g. in complex binding polynomials, although
this wouldn't be a centrifuge use), etc, etc. Weak because anytime you
want to make a change you need to recompile - but you get used to it
quickly. Looking at residuals can be a pain, but you also get used to
this. It does spit out parameter correlation coefficients. I'm somewhat
surprised that the "centrifuge NONLIN" (as I call it) doesn't give you
these. True, some other commercial packages (e.g. Origin) allow you to
program fitting functions ala nonlin, but as Olin pointed out these
don't always provide correlation coefficients. 

Fortran NONLIN is available from Mike Johnson's web site.

Cheers,
-Brian

On Wed, 26 Apr 2000 13:02:49 -0500, H. Olin Spivey wrote:

>   I don't use Nonlin, so I don't know what the "multi-channel" option is,
>but the question gives me the opportunity to plead for an improvement in
>most least-squares programs.  Specifically, I am amazed that no
>least-squares programs that I have ever seen, other than ours, actually
>prints the parameter correlation coefficients  ........
============================================
Brian M. Baker
Department of Molecular and Cellular Biology
Harvard University
baker@crystal.harvard.edu
(617) 496-6074