Christian Hollensen
Quantitative Clinical Pharmacology, AstraZeneca
Objectives: To explore the implication of estimating the variance of random effects as a random effects or fixed components in NONMEM.
Methods: A PK data set consisting of 100 subjects was simulated using a 2 compartment model with first order absorption and repeated 100 times [1]. 3 different estimation methods scenarios were explored: (1) Changing the variance estimation of IIV random effects from random to fixed components, (2) changing the variance estimation of random errors from a random to a fixed component and (3 changing the variance estimation of random errors from a random to a fixed component with data points below level of quantification using the M3 method with lower limit of quantification (LLOQ) at 4 four levels (15 %, 25 %, 35 % and 50 % LLOQ of data set). All simulations and estimations were performed in NONMEM V7.3.0 using similar initial estimates. The estimated post hoc parameters were compared to the simulated parameters to quantify the absolute relative deviation of the two estimation method in each of the 3 scenarios for every subject.
Results: NONMEM post hoc estimations were found for all data sets. In scenario (1) the mean deviation was smaller using fixed effects to estimate the variance of IIV random effects except for clearance. The difference ranged from 0.008 % for clearance to 3.4 % for central volume. In scenario (2) the mean deviation was smaller using fixed effects to estimate the variance of random errors. The difference ranged from 0.005 % for clearance to 7.3 % for the central volume. In scenario (3) the mean deviation increased with increasing proportion of LLOQ data points. The mean deviation was smaller using fixed effects to estimate the variance of the random errors.
Conclusions: NONMEM seems to have irregular behavior when estimating the variance of random effects as random or fixed components even though likelihood function should be the same. This work suggests that random effects in general should be estimated as fixed component. Further analysis on simpler and more complex models with differing initial estimates should be performed in order to further illuminate the degree of irregularity and to guide future modelling approaches in NONMEM.
References:
[1] Ahn JE, Karlsson MO, Dunne A, Ludden TM. Likelihood based approaches to handling data below the quantification limit using NONMEM VI. J Pharmacokinet Pharmacodyn. 2010 Jun;37(3):305-8.
Reference: PAGE 25 (2016) Abstr 5971 [www.page-meeting.org/?abstract=5971]
Poster: Methodology - Estimation Methods