From: Erik Olofsen <E.Olofsen@lumc.nl>

Subject: Large components in gradient vector

Date: Wed, 4 Apr 2001 17:15:08 +0200 (CEST)

With a data set - model combination which leads to a bit problematic convergence but good fits in the end I'm observing sudden jumps in values of components of the gradient vector, typically to values of the order 1E+08, which seems dependent on the number of required significant digits, and on the precision of calculations in the model. Is there another explanation for/solution to this problem?

Leiden University Medical Center

From: "Bachman, William" <bachmanw@globomax.com>

Subject: RE: Large components in gradient vector

Date: Wed, 4 Apr 2001 11:25:37 -0400

Sounds like the parameters for which the gradient is escalating are poorly determined. The estimation is then "floundering about" in that region of the response surface and generating the large gradient values. You might eliminate these parameters if not essential to the model or try fixing them if you have a reasonable guess for them.