Regression Analysis ...

ingber at umiacs.umd.edu ingber at umiacs.umd.edu
Sat May 25 09:04:41 EST 1991


This was a reply to a 'sci' posting, and I thought it might be of
interest to some members of this group as well.
==================
In many systems, the issue may simply be that the "residuals"
have functional dependence on the underlying variables, i.e.,
they vary over regions of the data.  Then, it may be worth while
trying to develop a better understanding of this "noise,"
or even to phenomenologically include such dependence in the
fitting process.  This is preferable to "hand-fitting."

This usually means that you cannot use simple canned regression
techniques, but are forced to use some other fitting techniques to
simultaneously account for nonlinearity and stochasticity.

I've opted for this approach, and can offer two short papers I wrote
which do not require any expert knowledge about a specific system,
e.g., nuclear physics or neuroscience, the latter being my present
application of these techniques.

On request, I'd be glad to email uuencoded-compressed PostScript drafts
of these papers ready to run off ('lpr') on a PostScript laserprinter.
I don't wish to start making lots of copies and run up a huge
postage bill.

%A L. Ingber
%T Very fast simulated re-annealing
%J Mathl. Comput. Modelling
%V 12
%P 967-973
%D 1989

%A L. Ingber
%T Statistical mechanical aids to calculating term structure models
%J Phys. Rev. A
%V 42
%D 1990
%P 7057-7064

> Article 2260 of sci.math.num-analysis:
> >From: mcdonald at aries.scs.uiuc.edu (Doug McDonald)
> Newsgroups: sci.math,sci.math.num-analysis,sci.physics,sci.chem
> Subject: Re: Regression analysis for an exponential
> References: <callahan.674938484 at newton.cs.jhu.edu>
> Xref: mimsy sci.math:18102 sci.math.num-analysis:2260 sci.physics:19836 sci.chem:4044
> 
> In article <callahan.674938484 at newton.cs.jhu.edu> callahan at cs.jhu.edu (Paul Callahan) writes:
> >In article <1991May22.154003.7805 at ux1.cso.uiuc.edu> mcdonald at aries.scs.uiuc.edu (Doug McDonald) writes:
> >>We got, by hand a pretty good fit, within about 2%. It was clear to me that
> >>the errors were systematic - our function was just not complicated 
> >>enough and I was loathe to add some arbitrary terms. But this referee insisted
> >>that a "hand" fit was NEVER acceptable - we HAD to do a least squares.
> 
> BUT if you don't get such good residuals, then you are FORCED to make some
> decision about where to get a good fit, and where a bad. 
> 
> 
> The point is, to me clear: why bother covering up your personal decisions
> with some hoity-toity mathematical fitting?? 
> 
> Doug McDonald

Lester Ingber
ingber at umiacs.umd.edu
--
Domain: curtiss at umiacs.umd.edu		     Phillip Curtiss
  UUCP:	uunet!mimsy!curtiss		UMIACS - Univ. of Maryland
 Phone:	+1-301-405-6710			  College Park, Md 20742



More information about the Comp-bio mailing list