A novel case-control subsampling approach for rapid model exploration of large clustered binary data.

Area of Science:

• Biostatistics
• Data Science
• Epidemiology

Background:

• Identifying factors associated with outcomes is crucial for inference and prediction.
• Big data analysis, especially with complex models like Generalized Linear Mixed Models (GLMMs), can be computationally intensive and time-consuming on standard hardware.
• Efficient model exploration is needed to overcome computational constraints in big data settings.

Purpose of the Study:

• To propose a novel subsampling scheme for rapid model exploration of clustered binary data.
• To enable the use of flexible and complex model setups, such as GLMMs with additive smoothing splines, on typical workstations.
• To approximate model estimates from full cohort analyses using a case-control-type design and sampling fractions.

Main Methods:

• A novel subsampling scheme is proposed for clustered binary data.
• Prospective cohort studies are reframed into a case-control-type design.
• Cluster-specific sampling fractions are derived to incorporate cluster variation.
• The method allows for the use of GLMMs with additive smoothing splines for flexible modeling.

Main Results:

• The proposed subsampling scheme enables rapid model exploration on standard workstations.
• Model estimates from the subsampling approach approximate those from full cohort analyses.
• Computationally prohibitive analyses can be conducted in a timely manner.
• The approach was applied to analyze risk factors for adverse reactions in blood donation.

Conclusions:

• The novel subsampling method significantly speeds up the analysis of clustered binary data.
• Complex statistical models can be efficiently explored and applied on standard computing resources.
• This approach facilitates timely identification of risk factors in large datasets, exemplified by blood donation adverse reactions.