cGLMM

cGLMM implements a privacy-preserving Expectation-Maximization procedure to construct generalized linear mixed models (GLMMs) on horizontally partitioned biomedical data for collaborative analyses such as genome-wide association studies (GWASs).

Key Features:

• GLMM construction: Builds generalized linear mixed models that incorporate random effects to model data with multiple sources of random variation.
• Privacy-preserving EM algorithm: Uses a collaborative Expectation-Maximization algorithm that enables parameter estimation without transferring raw data to a central server.
• Horizontal data partitioning: Operates on horizontally partitioned datasets where each participant holds a different subset of records but all records share the same set of known fixed and random effects.
• Local data control: Ensures each participating party retains control over its observational values of fixed effect variables and corresponding outcomes.
• Mathematical equivalence: The collaborative EM algorithm is mathematically equivalent to traditional EM algorithms used in GLMM construction.
• Implementation: Implemented in R to perform computational GLMM construction and estimation.
• Secure communication: Utilizes the rsocket package to facilitate secure inter-party data communication during collaborative computation.
• Validation: Has been tested on simulated and real human genomic datasets to demonstrate performance on relevant biomedical data.

Scientific Applications:

• Genome-wide association studies (GWASs): Enables detection of genetic variants associated with phenotypes such as human diseases using GLMMs on distributed cohorts.
• Collaborative multi-institution analyses: Supports joint analysis of large patient cohorts from multiple institutions while preserving individual-level privacy.
• Modeling hierarchical biomedical data: Applies GLMMs to biomedical datasets with multiple sources of random variation and correlated observations.

Methodology:

cGLMM applies a privacy-preserving Expectation-Maximization algorithm equivalent to traditional EM for GLMMs on horizontally partitioned data, with participants retaining local records; it is implemented in R and uses the rsocket package for secure inter-party communication.