The lasso--a novel method for predictive covariate model building in nonlinear mixed effects models.

Area of Science:

• Pharmacometrics
• Statistical Modeling
• Computational Pharmacology

Background:

• Stepwise covariate modeling (SCM) is standard for population pharmacokinetic/pharmacodynamic (PK/PD) models.
• SCM can lead to selection bias and poor prediction in small datasets.

Purpose of the Study:

• Implement and evaluate the lasso method for covariate selection in NONMEM.
• Compare lasso's performance against SCM regarding predictive ability, validation, and runtime.

Main Methods:

• Standardized covariates to zero mean and unit standard deviation for lasso.
• Fitted models with a restriction on the sum of absolute covariate coefficients.
• Implemented lasso as an automated tool using PsN.
• Compared lasso and SCM across 16 scenarios with varying dataset sizes and covariates, using 100 replicate datasets.

Main Results:

• Lasso consistently outperformed SCM in predicting external data across all scenarios.
• Lasso's cross-validation provided accurate estimates of prediction error.
• Lasso demonstrated faster runtimes than SCM, especially when run in parallel.

Conclusions:

• Lasso is superior to SCM for developing predictive covariate models, particularly for small datasets or subgroups.
• Lasso offers computational efficiency and integrated model validation via cross-validation.
• Lasso eliminates the need for user-defined P-values for covariate selection.