Comparison of models for baseline
Dansirikul, C. H.E. Silber, M.O. Karlsson
Division of Pharmacokinetics and Drug Therapy, Uppsala University, Uppsala, Sweden
Objectives: Baseline response can be modeled in different ways. Estimation of the typical value and interindividual variability (IIV) of baseline in the population (Model 1, M1) is considered the gold standard. Inclusion of the observed baseline response as a covariate, acknowledging the residual variability (M2) has been suggested as an alternative1. A more general version of this model (M3) also takes into account IIV in baseline. Furthermore, subtraction of baseline (M4) from observed responses, which has traditionally been used, was also explored. The aim of this study was to investigate the relative performance of various baseline modeling approaches.
Methods: PD responses over a single dosing interval were simulated from an indirect effect model in which a drug acts through stimulation or inhibition of the response according to an Emax model. Baseline response was simulated using M1. The performance of all models was investigated under 22 designs, each containing 100 datasets. NONMEM VI beta was used to estimate model parameters with the FO and the FOCE method. The mean error (ME, %) and root mean squared error (RMSE, %) were computed and used as an indicator of bias and imprecision, respectively. Absolute ME (aME) and RMSE from all methods were ranked within the same design, the lower the value the better. Average rank of each method from all designs was reported.
Results: For the estimates of the typical value and IIV of Emax, EC50, and kout, the use of M1, M2 and M3 with the FO method displayed similar bias. When the FOCE method was used, a smaller bias was noticed compared to the FO method and the average rank of aME for M1, M2, and M3 was 3.3, 3.8 and 2.9, respectively. M4 was found to give the largest bias (rank of 7.0 with FO, and 4.9 with FOCE). The difference in imprecision of these estimates between FO and FOCE methods was small. The smallest imprecision was noted with the use of M1 (rank 3.1 for both FO and FOCE) and increased, in order, with M3 (3.9 for both FO and FOCE), M2 (4.6-FO; 4.5-FOCE), and M4 (6.5-FO; 6.4-FOCE), respectively.
Conclusions: M3 performance was most similar to M1 and slightly better than M2. The FOCE method led to a smaller bias, but no marked reduction in imprecision, of parameter estimates compared to the FO method. The largest bias and imprecision of parameter estimates was noted with M4.
References
[1] Lewis Sheiner. NONMEM Tip#16 – April 2, 2003 – Modeling a "baseline" component and an additive "drug" component.