RE: nonmem question regarding condition number

From: Herbert Struemper Date: March 24, 2016 technical Source: mail-archive.com
Dear Laureen, I recommend that you determine with exploratory plotting how strong the evidence for the 3-compartment behavior is (eg. in semi-log plots: is 3 cpt behavior present in most subjects? are inflection points reasonably collocated? how do slope changes relate to measurement noise? etc.). If you do see strong evidence for 3 cpt behavior and have a biological rationale for a third compartment, I wouldn't just drop the 3 cpt model because of a high condition number in the model you tested, especially since you said that precision did not deteriorate. The fact that the 2 cpt model was borderline ill-conditioned (998 ~ 1000) with serially sampled IV data suggests that the 2 cpt model was already over-parameterized. My recommendation is to simplify the 2 cpt model (eg. by simplifying random structure, looking for large entries in correlation matrix or RSEs, ...) and then run the comparison again. Herbert
Quoted reply history
From: [email protected] [mailto:[email protected]] On Behalf Of Denney, William S. Sent: Thursday, March 24, 2016 6:53 AM To: L.A. ten Berg - Lammers; '[email protected]' Subject: [NMusers] RE: nonmem question regarding condition number EXTERNAL Hi Laureen, Large condition numbers (typically interpreted as >1000) indicate that two or more parameters in the model are highly correlated in their covariance and that the model parameters are difficult to identify. Given your description below, I would not suggest using the 3-compartment model (even with its 180 point reduction in OFV) because two of your parameters are so highly correlated that the condition number is reaching the level of computational precision (2E16 is very large for a condition number). I'd also look at the NONMEM output to see if the 3-compartment model had any errors around the convergence. Thanks, Bill From: [email protected]<mailto:[email protected]> [mailto:[email protected]] On Behalf Of L.A. ten Berg - Lammers Sent: Thursday, March 24, 2016 6:10 AM To: '[email protected]' Subject: [NMusers] nonmem question regarding condition number Dear nmusers, I am currently analyzing PK profiles after a single iv adminstration. Volunteers were sampled frequently: 12 samples per curve. For the structural model 2- and a 3-compartment models are compared. The CWRES vs time plot of the 2-compartment shows the typical wave, indicative for a 3-compartment model. Introduction of a third compartment eliminated the wave, OFV decreased with 180 points, precision of the estimates did change notably, but the condition number increased from 998 to 2.0*E16. Does anyone know how to interpret this large increase in condition number? Best wishes, Laureen ten Berg [Beschrijving: cid:[email protected]] Laureen ten Berg-Lammers Hospital Pharmacist Academisch Medisch Centrum Universiteit van Amsterdam Meibergdreef 9 | 1105 AZ Amsterdam | Room: M01-224 Tel: +31(0)20-5662726 | Pager *62726 | [email protected]<mailto:[email protected]> ________________________________ AMC Disclaimer : https://www.amc.nl/disclaimer ________________________________
Mar 24, 2016 L.a. Ten Berg - Lammers nonmem question regarding condition number
Mar 24, 2016 Bill Denney RE: nonmem question regarding condition number
Mar 24, 2016 Carolina Consuelo Llanos Paez RE: nonmem question regarding condition number
Mar 24, 2016 Sven Mensing Re: nonmem question regarding condition number
Mar 24, 2016 Ayyappa Chaturvedula Re: nonmem question regarding condition number
Mar 24, 2016 DJ Eleveld-Ufkes RE: nonmem question regarding condition number
Mar 24, 2016 Herbert Struemper RE: nonmem question regarding condition number