Filtering, FDR and power.

Area of Science:

• Bioinformatics
• Statistical genomics
• High-dimensional data analysis

Background:

• Filtering methods like fold-change or variance filters are commonly used in high-dimensional data analysis, such as differential gene expression analysis.
• These filters aim to reduce the multiple testing penalty and enhance statistical power.
• However, filtering can introduce bias into the multiple testing correction, affecting results.

Purpose of the Study:

• To investigate the impact of filtering on multiple testing correction bias in high-dimensional data analysis.
• To develop a method for assessing and mitigating bias introduced by filtering techniques.

Main Methods:

• Utilized simulation studies and an experimental dataset to evaluate the bias in false discovery rate (FDR) estimation.
• Analyzed the probability of filtering non-differentially expressed probes across the p-value range.
• Proposed a statistical test to detect FDR bias introduced by specific filters.

Main Results:

• A biased multiple testing correction occurs when non-differentially expressed probes are not filtered uniformly across all p-values.
• Filters associated with the tested hypothesis, such as fold-change, significantly bias the FDR.
• Filters with minimal FDR bias offer limited additional power for detecting differentially expressed genes.

Conclusions:

• Filtering in high-dimensional data analysis requires careful consideration due to potential bias in multiple testing correction.
• A proposed statistical test can guide researchers in selecting appropriate filters and filtering levels to minimize FDR bias.
• This approach aids in more reliable differential gene expression analysis.