Adaptive Multivariate Global Testing.

Area of Science:

• Statistics
• Neuroimaging Analysis
• Multivariate Data Analysis

Background:

• Challenges exist in testing global hypotheses with multivariate observations, particularly when sample size (n) is small compared to data dimension (K).
• Emerging applications in neuroimaging frequently encounter these small n, large K scenarios.

Purpose of the Study:

• To develop a robust methodology for hypothesis testing in high-dimensional settings with limited sample sizes.
• To introduce adaptive test designs that optimize predictive power while controlling Type I error.

Main Methods:

• Employing adaptive designs for sequential modification of test statistics based on accumulated data.
• Reducing the high-dimensional design space to a low-dimensional projection space for simplified analysis.
• Utilizing a general optimality result applicable to adaptive design settings.

Main Results:

• The proposed adaptive tests demonstrate substantial efficiency improvements over standard tests, especially when n is close to K.
• The dimensionality reduction enables simpler power analysis and comparisons with alternative methods.
• Empirical studies using simulated and EEG data confirm the effectiveness of the proposed methods.

Conclusions:

• The developed methodology offers a powerful and efficient approach for hypothesis testing in challenging high-dimensional, low-sample-size scenarios.
• Adaptive designs and dimensionality reduction are key to improving statistical power in fields like neuroimaging.
• Incorporating prior knowledge can further enhance test power, as shown in the EEG study.