Index:
[thread]
[date]
[subject]
[author]
From: HOLLADAYL@aol.com
To : rasmb@alpha.bbri.org
Date: Mon, 1 May 2000 21:50:39 EDT
I've not seen this with sed eq models but... linear approximations...
Hi all:
As Mr. Spock would say, "fascinating!". I've not seen this in any sed equil
model I've ever done, but in other models (e.g combining acid and Gdm-Cl
unfolding data), in which 2 parameters would be correlated not in anything
like a straight line, but a cresent shape, VERY curved. And in which case
the linear approximation of the uncertainty interval was way way too small.
So my caution is to take nothing for granted. There is no requirement that
the correlation between two parameters in a nonlinear model take on a linear
relationship. The actual contours in the residual sum of squares surface can
look like most anything you could imagine. I've seen some real doozseys from
unfolding models.
I favor doing bootstrap estimates for standard errors, and plotting the
bootstrapped estimates on scatter plots to see correlations. Generally when
there is a large difference between the linear approximation and bootstrap
there is a wierd (non straight line) correlation among at least two
parameters. In general the linear approximation will in my experience only
underestimate by at worst 2 fold or so the interval of uncertainty.
I've come to the theological conclusion that God may or may not exist, but
that Murphry certainly does rule in Heaven... so beware....
Les Holladay
Index:
[thread]
[date]
[subject]
[author]