Evaluation of the nonparametric estimation method in NONMEM VI: application to real data

Area of Science:

• Pharmacometrics
• Statistical Modeling
• Computational Biology

Background:

• Parametric methods like First-Order (FO) and First-Order Conditional Estimation (FOCE) are commonly used in NONMEM for model development.
• Evaluating the performance of nonparametric estimation methods against established parametric techniques is crucial for advancing modeling practices.

Purpose of the Study:

• To compare the performance of nonparametric estimation methods (FO-NONP, FOCE-NONP) against parametric methods (FO, FOCE) using real-world datasets.
• To assess the precision and bias of parameter and parameter distribution estimation using different methods.

Main Methods:

• Four estimation methods were evaluated: FO, FOCE, FO-NONP, and FOCE-NONP.
• 25 real-world models were re-analyzed, and numerical predictive checks were performed.
• 1000 datasets were simulated per model/method to generate prediction intervals and compute error metrics.

Main Results:

• Nonparametric methods generally exhibited less imprecision and bias compared to parametric methods.
• FOCE-NONP showed significantly lower imprecision and bias than FOCE for several outcomes.
• FO-NONP demonstrated even more pronounced improvements over FO.

Conclusions:

• Nonparametric estimation methods in NONMEM VI offer improved performance over traditional parametric methods (FO, FOCE) when applied to real datasets.
• These findings support the adoption of nonparametric approaches for more robust model parameter estimation.