Flexible modeling via a hybrid estimation scheme in generalized mixed models for longitudinal data

Area of Science:

• Statistics
• Computational Statistics
• Biostatistics

Background:

• Generalized mixed models (GMMs) are computationally intensive for likelihood inference.
• Laplace's approximation in GMMs can be inadequate for individuals with sparse data.
• Existing methods struggle with the computational complexity and data sparsity in GMMs.

Purpose of the Study:

• To propose a hybrid estimation scheme for GMMs that addresses the inadequacy of Laplace's approximation with sparse observations.
• To enhance the computational tractability and statistical accuracy of likelihood inference in GMMs.
• To enable flexible modeling of covariate effects and facilitate model selection.

Main Methods:

• A hybrid estimation scheme combining Monte Carlo approximations (importance sampling) for sparse data and Laplace's approximation for non-sparse data.
• Diagnostic checks to assess the adequacy of Laplace's approximation.
• Use of regression splines for flexible covariate effect modeling and model selection procedures for knot and variable selection.

Main Results:

• The proposed hybrid method improves the accuracy of likelihood inference for GMMs, particularly with sparse data.
• Demonstrated computational and statistical advantages through simulation studies.
• Successfully applied to longitudinal fecundity data, modeling overdispersion using a double exponential family.

Conclusions:

• The hybrid estimation scheme offers a computationally tractable and statistically sound approach for GMMs with sparse observations.
• This method allows for flexible modeling and improved inference in complex statistical analyses.
• The approach is effective for analyzing longitudinal data with overdispersion.