Nick Holford
University of Auckland
Objectives: To evaluate NONMEM and Monolix in terms of parameter estimation bias and uncertainty coverage bias using a parametric bootstrap procedure.
Methods: Default SAEM estimation options (NONMEM AUTO, Monolix algorithms options created when no algorithms file supplied as input) were used. All calculations were performed using the NeSI PAN cluster. Four problems of increasing difficulty were tested: warfarin pharmacokinetics, simple and complex tumour growth inhibition, viral load kinetics. Uncertainty bias was based on the bootstrap standard deviation relative to the standard error which described 95% coverage of the bootstrap distribution.
Results and Conclusions: Some problems were better described using FOCE while others were better described using SAEM (see Table 1 for tumour growth inhibition [1] example). Both methods had biased uncertainty relative to bootstrap coverage (Table 2).
Table 1 Estimation Bias for Model Parameters
|
Tham TGI |
NONMEM |
NONMEM |
Monolix |
|
|
Method |
FOCE |
SAEM |
SAEM |
|
|
Option |
INTER |
AUTO |
No algo |
|
|
Parameter |
TRUE |
MDL |
MDL |
MDL |
|
POP_SIZE0 |
6.39 |
1.1% |
3.7% |
10.8% |
|
POP_TOVER |
33.9 |
-37% |
22% |
230% |
|
POP_AE50 |
6324 |
60% |
1.7% |
-76% |
|
POP_TEQ |
4.407 |
111% |
24% |
-34% |
|
PPV_SIZE0 |
0.6 |
-7% |
-6% |
-33% |
|
PPV_TOVER |
0.420 |
-72% |
38% |
67% |
|
PPV_AE50 |
1.200 |
-4% |
-24% |
20% |
|
PPV_TEQ |
0.200 |
-64% |
59% |
324% |
|
RUV_CV |
0.199 |
-3.9% |
-5% |
58% |
TRUE=Parameter used for simulation MDL=Model parameter bias
Table 2 Uncertainty Bias for Model Parameters
|
Tham TGI |
NONMEM |
NONMEM |
NONMEM |
Monolix |
|
Method |
FOCE |
FOCE |
SAEM |
SAEM |
|
Option |
INTER |
AUTO |
No algo |
|
|
Parameter |
95 SE |
AsymRSE |
AsymRSE |
AsymRSE |
|
POP_SIZE0 |
7% |
-3% |
26% |
-6% |
|
POP_TOVER |
65% |
-16% |
4% |
-17% |
|
POP_AE50 |
61% |
45% |
42% |
125% |
|
POP_TEQ |
64% |
-38% |
10% |
11% |
|
PPV_SIZE0 |
10% |
-10% |
13% |
42% |
|
PPV_TOVER |
105% |
164% |
289% |
-41% |
|
PPV_AE50 |
41% |
-47% |
140% |
123% |
|
PPV_TEQ |
120% |
14% |
65% |
-35% |
|
RUV_CV |
5% |
30% |
22% |
22% |
95SE= standard error which described 95% coverage of the bootstrap distribution
AsymRSE=bias of asymptotic relative standard error relative to 95SE
References:
[1] Tham LS, Wang L, Soo RA, Lee SC, Lee HS, Yong WP, et al. A pharmacodynamic model for the time course of tumor shrinkage by gemcitabine + carboplatin in non-small cell lung cancer patients. Clin Cancer Res. 2008;14(13):4213-8.
Reference: PAGE 23 () Abstr 3143 [www.page-meeting.org/?abstract=3143]
Poster: Methodology - Estimation Methods