Senior Director Bristol Myers Squibb Lower Gwynedd, Pennsylvania, United States
Disclosure(s):
Chuanpu Hu, PhD: No financial relationships to disclose
Objectives: The full model approach has been increasingly used to assess covariate effects in population pharmacokinetics [1]. In practice, the construction is often based on univariate regression, and the conclusions are based in conjunction with a final model obtained by stepwise reduction. This usage deviates from the originally designed full model construction based on the principle of pre-specification [2], resulting the loss of desired statistical properties such as unbiasedness and p-value interpretation. In contrast, prespecifying the full model based on adequate sample size and subsequently reducing the model in one step by removing those covariates with negligible effect sizes have been used [3,4]. This investigation aims to illustrate the theoretical and practical differences between the two approaches.
Methods: Data from a recent regulatory submission analysis were used as illustration. The population pharmacokinetics of luspatercept, a first in class erythroid maturation agent, were characterized in patients with myelodysplastic syndrome in two phase 2 and two phase 3 studies. The dataset consisted of 470 patients, with 3504 serum concentration measurements. Results of the two full model approaches were compared.
Results: A one-compartment model with first-order absorption was developed as the structural model. The submission full covariate model, based on univariate regression, had body weight and albumin on clearance and volume, and 5 additional covariates on clearance. 2 of these 5 additional covariates were subsequently eliminated by a stepwise reduction, leaving a total of 5 covariates on clearance in the submission final model. In contrast, the full model based on the principle of pre-specification was able to evaluate a total of 12 covariates on clearance, with body weight as the only covariate on volume. The subsequent effect-size based reduced covariate model on clearance was similar to the submission final model, with only 1 covariate being different. Thus, the full model based on the principle of pre-specification was able to evaluate more covariate effects by showing that they do not affect luspatercept clearance. While the final model results between the stepwise-reduction and the effect-size based reduction were similar, the latter was less time consuming.
Conclusions: The pre-specified model allowed more covariate effects to be evaluated, and had improved statistical properties, including unbiasedness and p-value interpretations. Reducing the full model based on effect size reached similar results as stepwise reduction but required less effort.
Citations: [1] Gastonguay MR (2004). A full model estimation approach for covariate effects: inference based on clinical importance and estimation precision. AAPS J; 6:S1. [2] Harrell F (2001) Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis. Springer, New York. [3] Hu C et al. (2014) Latent variable indirect response joint modeling of a continuous and a categorical clinical endpoint. J Pharmacokinet Pharmacodyn 38 (2):237-260. [4] Yao Z et al. (2018) Population pharmacokinetic modeling of guselkumab, a human IgG1λ monoclonal antibody targeting IL-23, in patients with moderate-to-severe plaque psoriasis, J Clin Pharm, 58(5):613-627.
Keywords: Population pharmacokinetics, model selection, confirmatory analysis
