From: Erik Olofsen +31 71 5263344 / 5152719
Subject: Choosing between conditional estimation methods
Date: 21 Oct 1997 13:53:51 -0400

Dear NONMEM users,

I have analyzed a data set with the first-order (FO) method and the first-order conditional estimation method without (FOCE) and with eta-epsilon interaction (FOCEI).

All three methods yielded similar goodness-of-fit plots as described in NONMEM Guide VII. The values of the parameters obtained by the three methods are also similar but somewhat different, and I'm now trying to determine which set should be the final one.

From the guide I understand that the model parameters obtained with the FOCE method should be given greater credence, because the adequacy of the first-order approximation of the Hessian matrix is better when for the etas the individual conditional estimates are used.

Eta-epsilon interaction can arise when the intraindividual variance is heteroscedastic, or when the intraindividual error is not normally distributed. In this case, one would expect that the FOCEI method is the best one, but according to the guide this need not be true (p7). So how does one choose between the FOCE and the FOCEI methods?

Thanks in advance for any suggestions,

Erik Olofsen
Deparment of Anesthesiology
Leiden University Medical Center
The Netherlands