Dept. of Drug Metabolism and Kinetics, N.V. Organon, Oss, The Netherlands.
Objectives:
The pharmacokinetics of Drug A were determined in 39 healthy women during a treatment period of three to six months with two different formulations. Results of this analysis will be presented elsewhere. The aim of the current presentation is to describe a data analytic issue encountered during the above mentioned analysis: the effect of sampling scheme on parameter estimation.
Design:
Study I. 10 healthy female volunteers (age 24 to 34) were treated with Drug A during three months. Fixed-time samples were taken on day 7,14 and 21 of month 1 and 3 (+/- 60 samples per subject, ''full sampling'').Study II. 29 healthy female volunteers (age 20 to 35) were treated with two different formulations of Drug A during 6 months. Non-fixed-time samples were taken on day 3,14,17 and 21 of month 1,3 and 6 (+/- 10 samples per subject, ''sparse sampling'').
Materials & Methods:
The ''full'' and ''sparse'' data were pooled and analysed using double precision NONMEM (version IV level 1.0) and double precision PREDPP (version III level 1.0) with S-Plus as a statistical tool.
Results:
The pooled data were fitted to a two compartment model with first-order absorption. An initial screening for covariate effects using a generalized additive model resolved that the distribution rate constant K23 was significantly influenced by the type of study (I, II) and a number of other covariates. Incorporating a study-effect in the NONMEM model resulted in a significant improvement of fit and in a disappearance of the influence of the other covariates on K23. To see whether this effect could be explained by the difference in sampling between Study I and II (''full'' and ''sparse'', resp.), the ''full'' data were reduced to ''sparse'' data by randomly removing all but one sample per day per subject. No study-effect nor other covariate effects were found for K23 with the reduced dataset.
Discussion:
The finding that type of study did not influence K23 using the reduced dataset implies that the study-effect was caused by the difference in sampling. The influence of sampling schedules on parameter estimation is a well-known feature. Questions arise concerning identifiability and reliability of covariate effects. Was the finding of significant covariate effects an artefact caused by a study-effect, which was in turn caused by the difference in sampling scheme? Or does sparse sampling conceal covariate effects which would be revealed using a full sampling schedule? In this case a clear answer could not be given.
Conclusion:
Questions arose concerning the interaction between the effect of sampling schedule, parameter estimation and elucidation of covariate effects.