Bag of little bootstraps for massive and distributed longitudinal data

Area of Science:

• Statistics
• Computational Biology
• Data Science

Background:

• Linear mixed models (LMMs) are crucial for longitudinal data analysis.
• Traditional bootstrap methods for LMM variance components are computationally infeasible for massive datasets.
• Health systems and tech companies generate large-scale longitudinal data.

Purpose of the Study:

• To extend the scalable bag of little bootstraps (BLB) method to longitudinal data.
• To develop an efficient Julia package (MixedModelsBLB.jl) for LMM inference.
• To provide a computationally advantageous alternative to traditional bootstrap methods.

Main Methods:

• Extension of the bag of little bootstraps (BLB) algorithm for longitudinal data.
• Development of the MixedModelsBLB.jl package in Julia.
• Simulation studies and real-world data analysis for performance evaluation.

Main Results:

• The proposed BLB method demonstrates favorable statistical performance.
• Significant computational advantages over the traditional bootstrap method were observed.
• Achieved a 200x speedup for variance component inference on datasets with 1 million subjects.
• The MixedModelsBLB.jl package uniquely handles datasets exceeding 10 million subjects on desktop computers.

Conclusions:

• The BLB method provides a scalable and efficient solution for LMM inference on large longitudinal datasets.
• MixedModelsBLB.jl offers a practical tool for analyzing big data in health systems and technology.
• This advancement significantly improves the feasibility and speed of statistical inference for variance components.