Investigations of the representation of covariate changes in NONMEM

Aims: A widely exploited feature of the mixed effect modelling software NONMEM (Icon, Ireland) is the ability to model the influence of covariates that change with time within a subject.  It is documented that NONMEM represents the covariate value between TIME1 and TIME2 as the value at COV2 rather than the more intuitive value of COV1.  For example, in the following dataset:


40       70

50       72

NONMEM interprets the value of WT at TIME 40 as 70, 40>TIME<50 as 72 and at TIME 50 as 72 (next observation carried backward (NOCB) for the interval between observations).  The alternative (last observation carried forward (LOCF) for the interval) would have the value of WT at TIME 40 as 70, 40>TIME<50 as 70 and at TIME 50 as 72.  Hence the database coding shown above would institute the change in WT at TIME>40 in a standard NONMEM model.

Methods: We investigated by simulation some consequences of representing covariate changes as NOCB by comparing the output of NONMEM 7.2 with reference software which coded covariate changes as LOCF (Scientist 2.1, Micromath).  The investigation used simulations of a 1 compartment i.v infusion model (ADVAN1 TRANS2) with a simultaneous change in two covariates (WT and ALB) at TIME=50 with the covariate effects coded as follows:



The size of the covariate changes were exaggerated to cause pronounced changes in the simulated concentration time-course that were visible when log-concentration was plotted against time.  This allowed the influence of the covariate in the interval between the observations to be inferred by the characteristic changes in the concentration time-course (half-life change for ALB, step change with no change in half-life for WT).

Results: The simulations using Scientist produced changes in the concentration time-course at TIME=50.  However, the simulations using NONMEM were dependent on the TIME value of the row immediately prior to TIME=50.  Coding a double covariate change in NONMEM on two different rows with the same time value produced different results depend on the order of the rows.  NONMEM could be forced to emulate LOCF behaviour by inserting a initial “dummy” row in the database with the same the TIME as the required covariate change.

Conclusions: The default NOCB behaviour of NONMEM is counter intuitive, and may not reflect general understanding of how changes in covariate values should be interpreted.  Further investigations of the implications of this behaviour are warranted.

Richard Upton

  • University of South Australia