Re: Question on Bias est. and the conditional method

From: ABoeckmann <alison@c255.ucsf.edu> Subject: Re: Question on Bias est. and the conditional method Date: Thu, 11 Jan 2001 11:15:36 -0800 (PST) Carl, I don't know why the first order method gives low estimates. However, it is not surprising that conditional estimation runs into difficulties. The models for K12 and K21 are: K12=THETA(3)+ETA(3) K21=THETA(4)+ETA(4) or K12=THETA(3)*EXP(ETA(3)) K21=THETA(4)*EXP(ETA(4)) The initial estimates for these thetas are 0.08 and 0.05. With additive eta, it is quite likely that a negative eta will give rise to negative K12 or K21. The mult. eta models are likely to work better. The overflow is probably due to a large value of eta, leading to a value of EXP(ETA(.)) that causes an overflow. (It may be that the data for INDIVIDUAL NO. 23 ID=0.10988500E+06 is especially problematic.) In the help item for abbreviated code, we describe the EXIT statement, that can be used to tell NONMEM to avoid certain vaues of ETA. I would put these statements in the $PK blocK; $PK IF (ABS(ETA(1)).GT.10) EXIT 1 100 IF (ABS(ETA(2)).GT.10) EXIT 1 200 IF (ABS(ETA(3)).GT.10) EXIT 1 300 IF (ABS(ETA(4)).GT.10) EXIT 1 400 With the NOABORT option of $EST, NONMEM will attempt to avoid these large etas and will try to continue. The PRDERR file will contain messages describing which condition(s) occured. E.g., if the user error code is 300, the trouble was with eta 3, and so on. Usually the exit conditions will not persist beyond the first few iterations. - Alison Boeckmann NONMEM Project Group