Performance in population models for count data, part II: a new SAEM algorithm.

Area of Science:

• Statistical modeling
• Pharmacometrics
• Clinical trial analysis

Background:

• Mixed-effects models are increasingly used for analyzing count data in clinical trials.
• Current estimation algorithms like LAPLACE and Gaussian Quadrature (GQ) have limitations, including parameter bias and long runtimes.
• The Stochastic Approximation Expectation Maximization (SAEM) algorithm is efficient for continuous data analysis.

Purpose of the Study:

• To implement and evaluate a novel SAEM algorithm for analyzing count data.
• To assess the performance of the SAEM algorithm in terms of accuracy and speed compared to existing methods.
• To extend the SAEM algorithm for parameter and Fisher information matrix estimation.

Main Methods:

• Implementation of a new SAEM algorithm in MATLAB.
• Stochastic Monte Carlo simulations with re-estimation based on established scenarios.
• Exploration of six probability distribution models, including those for over- or under-dispersion.
• Estimation of parameters and standard errors.

Main Results:

• The SAEM algorithm demonstrated low relative bias (<0.92% for fixed effects, <4.13% for random effects) across all studied models.
• Empirical and estimated relative standard errors showed high similarity (<1.7% difference).
• The algorithm achieved significantly faster computation times (max 95s for parameter estimation, 56s for SE estimation) compared to LAPLACE and GQ.

Conclusions:

• The SAEM algorithm is a powerful and efficient tool for the accurate analysis of count data in mixed-effects models.
• It provides reliable parameter and standard error estimates with reduced bias and improved speed.
• The algorithm is integrated into Monolix 3.1, offering a valuable alternative for clinical trial data analysis.