Optimal tests shrinking both means and variances applicable to microarray data analysis.

Area of Science:

• Biostatistics
• Genomics
• Statistical genetics

Background:

• Microarray data analysis faces challenges due to the "large p small n" characteristic, leading to low statistical power in individual gene hypothesis tests.
• Existing methods like the FS test use James-Stein shrinkage for variance estimation to improve power, but further optimization is needed.

Purpose of the Study:

• To develop an optimal Bayes test (MAP test) for microarray data that maximizes average power by modeling means and variances.
• To derive an efficient empirical Bayes approximation (FSS test) to the MAP test for practical application.

Main Methods:

• Developed a theoretical framework to model key parameters with distributions for an optimal Maximum Average Power (MAP) test.
• Derived the FS test as an empirical Bayes approximation for variance modeling.
• Introduced the FSS test by modeling both means and variances, providing a computationally efficient approximation to the MAP test.

Main Results:

• The FSS test, which shrinks both means and variances, achieves numerically identical average power to the computationally intensive MAP test.
• Extensive numerical evidence demonstrates the FSS test's uniform superiority in average power over established methods like the F test, SAM, and moderated t-test.
• Theoretical analysis confirms the FSS test's optimality in power when controlling the false discovery rate (FDR).

Conclusions:

• The FSS test provides a significant advancement in microarray data analysis, offering optimal average power and improved performance.
• This method uniformly outperforms existing statistical tests for gene expression analysis, particularly when controlling the FDR.
• The FSS test represents a computationally efficient and statistically powerful approach for identifying significant genes in high-dimensional genomic studies.