Normalisation weight in a covariate function affects the relative standard error of clearance: an example with paediatric data
Swantje Völler (1) ; Sebastiaan Goulooze (1); Pyry Välitalo (1); Elisa Calvier (1); Elke Krekels (1); Catherijne A.J. Knibbe (1,2)
(1) Division of Pharmacology, LACDR, Leiden University, Leiden, The Netherlands, (2) Department of Clinical Pharmacy, St. Antonius Hospital, Nieuwegein, The Netherlands
Objectives: While in population models most covariates are normalized to the median value, bodyweight (BW) normalization to 70 or 1 kg is often applied. Using a phenobarbital dataset in neonates we investigated the impact of normalization weight on the relative standard error (RSE) and bootstrap values of clearance (CL) estimates. In addition, bootstrap confidence intervals (95%CIs) of typical CL predictions between 1 to 100 kg were generated. Finally, using simulations we studied this impact for different weight distributions.
Methods: For this analysis, a pharmacokinetic model for phenobarbital in 53 (pre)term neonates (0.45 - 4.4 kg) in which BW was incorporated as covariate on CL using a power function with estimated exponent was used. Normalization weight was set to 1 kg, 2.7 kg (median BW) and 70 kg to perform model estimation and bootstraps (1000 runs) using NONMEM. In addition, for all three normalizations, 1000 CL functions were calculated over a weight range of 1-100 kg using the bootstrap estimates of CL and exponent. 95% CI of these functions were compared across normalizations. Additionally, simulations with different median BW and BW ranges were performed.
Results: The RSE of the estimated normalized CL value was the lowest with a normalization to median BW (8.1%), compared to the 1 kg (10.5%) and 70 kg (48.8%), and was comparable for the bootstrap and the NONMEM covariance matrix. The exponent and the RSE of the exponent remained unchanged for all runs. The bootstrap 95%CI of CL over a weight range of 1-100 kg was independent of the normalization weight used, implying that a normalization weight further away from the median of the data does not impact the precision of the CL estimation in the weight range of the studied population. Simulation studies showed that the increase in RSE with 70 kg normalization was highest for paediatric studies with a narrow weight range and with a median BW away from 70 kg.
Conclusions: Using a normalization weight outside the observed covariate range can result in a high RSE of the corresponding population estimate, as the obtained RSE corresponds to the RSE for an extrapolated population parameter value. As this RSE might not be informative about the precision of the CL estimate in the studied BW range, normalizing BW on 70 kg in the paediatric population should be applied with caution as this might lead to wrong decision making in model building.
