Modifying the false discovery rate procedure based on the information theory under arbitrary correlation structure...

Area of Science:

• Statistics
• Bioinformatics
• Genomics

Background:

• Controlling the False Discovery Rate (FDR) is crucial in multiple comparison procedures (MCPs) across scientific fields.
• Test statistic correlation can increase FDR variance and bias, necessitating robust control methods.
• Information theory offers a novel approach to modify MCPs for correlation structures.

Purpose of the Study:

• To develop and evaluate modified MCPs that account for arbitrary correlation structures using information theory.
• To compare the performance of proposed procedures (M1, M2, M3) against established methods like Benjamini-Hochberg (BH) and Benjamini-Yekutieli (BY).
• To assess the utility of these procedures in analyzing high-dimensional gene expression data.

Main Methods:

• Proposed three modified procedures (M1, M2, M3) based on conditional Fisher Information.
• Employed simulation studies with multivariate Gaussian features at varying correlation levels.
• Applied procedures to real high-dimensional colorectal cancer gene expression data.
• Utilized Efficient Bayesian Logistic Regression (EBLR) for predictive modeling.

Main Results:

• Proposed procedures performed similarly to BH when no correlation existed.
• BY was conservative, BH was liberal in low to medium correlations; proposed methods showed reduced feature screening with increased correlation.
• In highly correlated scenarios, proposed methods approached Bonferroni procedure performance.
• EBLR models using features screened by M1 and M2 exhibited minimum entropy and higher efficiency.

Conclusions:

• Modified procedures offer greater flexibility than BH and BY in handling test statistic correlations.
• These information-theory-based methods effectively reduce screening of non-informative features as correlations increase.
• The proposed procedures enhance the efficiency of predictive modeling in high-dimensional data analysis.