High-dimensional randomization-based inference capitalizing on classical design and modern computing

Area of Science:

• Genomics
• Biostatistics
• Environmental Health

Background:

• High-dimensional data analysis often involves repeated hypothesis testing, leading to complications.
• Small sample sizes can exacerbate issues with reliance on asymptotic p-values.

Purpose of the Study:

• To propose and illustrate a statistically valid method for analyzing high-dimensional data.
• To address complications arising from repeated hypothesis tests and asymptotic p-values.
• To introduce a method using a scientifically motivated scalar summary statistic.

Main Methods:

• Utilized a crossover study design with seventeen participants.
• Examined the effect of ozone exposure versus clean air on the DNA methylome.
• Applied a proposed test yielding a single null randomization distribution and a Fisher-exact p-value.
• The multivariate outcome comprised 484,531 genomic locations.

Main Results:

• The proposed test provides a single, statistically valid p-value, irrespective of data structure.
• Demonstrated the method's application in a real-world environmental exposure study.
• Highlighted the importance of a priori selection of a relevant test statistic for power.

Conclusions:

• The proposed method offers a statistically robust alternative to common practices in high-dimensional data analysis.
• Careful selection of the summary statistic is crucial for the relevance and power of the test.
• Advocates for moving beyond asymptotic p-values and arbitrary significance thresholds.