Model selection for inferential models with high dimensional data: synthesis and graphical representation of multiple...

Area of Science:

• Statistics
• Computational Biology
• Bioinformatics

Background:

• Variable selection in high-dimensional data is crucial for inferential research.
• Identifying true causal factors is often hindered by the choice of statistical method.
• Existing methods lack a formal approach to triangulate results using both model stability and coefficient estimates.

Purpose of the Study:

• To develop a flexible and straightforward method for triangulating variable selection results.
• To combine model stability and coefficient estimates for robust causal factor identification.
• To provide a formal approach for integrating multiple statistical methods in high-dimensional data analysis.

Main Methods:

• Evaluated six variable selection methods using simulated datasets with known relationships.
• Employed a bootstrap methodology to aggregate stability matrices across methods.
• Developed novel graphical approaches for visualizing and comparing method-specific and aggregated results.

Main Results:

• The proposed aggregated method successfully triangulated results across multiple variable selection techniques.
• Combined results incorporated uncertainty from between-method variability, offering a more comprehensive view.
• In simulated datasets, the aggregated method demonstrated comparable or superior performance to individual methods, with lower error rates and clearer identification of true causal variables.

Conclusions:

• The developed adaptable method provides a formal and flexible route for triangulating variable selection results.
• This approach enhances the reliability of identifying causal factors in high-dimensional data.
• The combined method offers a robust solution for navigating the complexities of statistical method selection in inferential research.