Multivariate association and dimension reduction: a generalization of canonical correlation analysis

Area of Science:

• Statistics
• Data Analysis
• Environmental Science

Background:

• Identifying complex relationships between multiple variables is crucial in data analysis.
• Existing methods may not fully capture nonlinear dependencies or handle multiple sets of vectors effectively.

Purpose of the Study:

• To propose a novel generalized index for uncovering relationships between sets of random vectors.
• To extend the methodology for multiple sets and groups of sets, including significance testing.
• To demonstrate the method's utility with simulations and real-world environmental data.

Main Methods:

• Developing a generalized index based on minimizing L(2) distance between projected vectors and unknown functions.
• Employing the Nadaraya-Watson smoother for estimating unknown functions.
• Utilizing a bootstrap procedure for detecting the number of significant relationships.
• Applying the multiple-set methodology to environmental data (mortality, weather, pollutants).

Main Results:

• The proposed index effectively recovers relationships between vector sets.
• The method successfully identifies both linear and nonlinear relationships.
• Application to environmental data revealed significant links between mortality, weather, and pollutants.
• Dimension-reduced vectors were used to build nonlinear time-series regression models for mortality.

Conclusions:

• The generalized index provides a powerful tool for exploring complex multivariate relationships.
• The methodology is versatile and applicable to diverse fields, including environmental health.
• The findings support the use of this approach for building predictive models from high-dimensional data.