Using Stochastic Simulation-Estimation and Automated Model Development to Assess Power and Accuracy for Covariate...

Area of Science:

• Pharmacometrics
• Clinical Trial Simulation
• Population Pharmacokinetics

Background:

• Evaluating study designs for covariate effect identification in population pharmacokinetic (PopPK) modeling often assumes the true model is known.
• This study compares this assumption with the realistic scenario where the PopPK model is built from the generated data.

Purpose of the Study:

• To compare three approaches for PopPK model building: stochastic simulation and re-estimation (SSE), automated model development with exploratory covariate search (AMD-exploratory), and automated model development with structural covariate forcing (AMD-structural).
• To assess type 1 error (T1E), covariate identification power, and covariate parameter accuracy under different design sparsities.

Main Methods:

• Simulated a covariate effect (pregnancy on clearance) to evaluate T1E and power.
• Compared SSE, AMD-exploratory, and AMD-structural approaches.
• Assessed performance across rich, medium, and sparse study designs.

Main Results:

• SSE showed inflated power estimates; AMD-structural had a 20% inflated T1E rate.
• Covariate identification power varied: SSE (99%, 100%, 79%), AMD-exploratory (74%, 72%, 41%), AMD-structural (92%, 93%, 80%) for rich, medium, sparse designs, respectively.
• Sparse designs highlighted differences, with AMD-exploratory often failing to identify correct covariates. Relative root mean squared error (rRMSE) for parameter estimates was lowest in SSE, followed by AMD-exploratory, then AMD-structural.

Conclusions:

• SSE provides optimistic power estimates due to its data-driven nature.
• AMD-based approaches incorporate model uncertainty, offering a more realistic reflection of real-world PopPK analysis conditions.
• The choice of modeling strategy significantly impacts covariate identification and parameter accuracy, particularly in sparse designs.