Federated learning algorithms for generalized mixed-effects model (GLMM) on horizontally partitioned data from...

Area of Science:

• Statistical modeling
• Biostatistics
• Distributed computing

Background:

• Generalized linear mixed effect models (GLMMs) are crucial for analyzing hierarchical and non-independent data.
• Federated learning enables distributed data analysis without centralizing sensitive information.
• Challenges exist in applying GLMMs in federated settings due to computational complexity and data distribution.

Purpose of the Study:

• To develop federated solutions for generalized linear mixed effect models (GLMMs).
• To address numerical errors and singularity issues inherent in federated GLMM computations.
• To evaluate the performance of federated GLMMs against centralized approaches.

Main Methods:

• Utilized Laplace approximation (LA) and Gaussian Hermite approximation (GH) for log-likelihood function approximation.
• Implemented federated decomposition to bring computation closer to distributed data sources.
• Employed log-sum-exponential trick and adaptive regularization to overcome numerical instability.

Main Results:

• The proposed federated GLMM methods effectively handle hierarchical data in a distributed manner.
• Demonstrated comparable performance with Laplace approximation and superior performance with Gaussian Hermite approximation.
• Successfully revealed parameter significance in distributed datasets.

Conclusions:

• Federated GLMMs with different approximations can analyze versatile biomedical data with mixed effects.
• The methods address non-independence arising from hierarchical structures (e.g., institutes, regions).
• Supports researchers in privacy-preserving analysis of complex biomedical datasets.