A SIMPLE AND FLEXIBLE TEST OF SAMPLE EXCHANGEABILITY WITH APPLICATIONS TO STATISTICAL GENOMICS

Area of Science:

• Statistical genomics
• Multivariate data analysis
• Bioinformatics

Background:

• Assessing sample exchangeability and feature independence is fundamental in multivariate data analysis, particularly in statistical genomics for tasks like demographic inference and polygenic risk score construction.
• Existing methods may not adequately address both sample exchangeability and feature independence simultaneously or efficiently.

Purpose of the Study:

• To introduce a novel non-parametric method, the V test, for assessing sample exchangeability given feature dependency and feature independence given sample exchangeability.
• To provide a computationally efficient and flexible tool for analyzing complex genetic data.

Main Methods:

• Developed a non-parametric V test utilizing large-sample asymptotics to handle high-dimensional data.
• Conducted extensive simulations to evaluate the test's performance in controlling Type I error and comparing it with existing methods like random matrix theory-based approaches.
• Applied the V test to real genetic data from the 1000 Genomes Project.

Main Results:

• The V test effectively controls Type I error across various realistic scenarios.
• The V test demonstrates favorable performance compared to unsupervised stratification tests.
• Removing rare variants significantly increased the V test statistic's p-value in exchangeability assessments.
• The V test identified different optimal linkage disequilibrium (LD) splits compared to methods not based on hypothesis testing.

Conclusions:

• The V test is a conceptually simple, fast, and flexible method for addressing key questions in multivariate data analysis and statistical genomics.
• The V test provides valuable insights for assessing genetic sample exchangeability and optimizing LD splits for downstream analyses.
• Available software in R (flintyR) and Python (flintyPy) facilitates the application of these methods.