Estimating the local false discovery rate via a bootstrap solution to the reference class problem.

Area of Science:

• Statistics
• Bioinformatics
• Medical Imaging

Background:

• Local false discovery rate (LFDR) estimation is crucial for analyzing large datasets like genomic and imaging data.
• Existing LFDR methods may not fully leverage contextual information available in biological data.
• Covariates can provide additional biological context to improve statistical testing.

Purpose of the Study:

• To introduce a novel LFDR estimation model incorporating covariates.
• To evaluate the performance of the proposed adaptive reference class (ARC) method.
• To demonstrate the method's utility in real-world biological and medical imaging datasets.

Main Methods:

• Developed an LFDR estimation model that includes a covariate for each hypothesis test.
• Employed a bootstrap approach to estimate the optimal tuning parameter by minimizing bias and variance.
• Assessed the ARC method's performance theoretically and through simulations under specific prior probability assumptions.

Main Results:

• The ARC method demonstrates asymptotically optimal mean squared error compared to methods using the entire hypothesis set.
• Simulation studies confirm the estimator's performance for finite hypothesis numbers.
• The method was successfully applied to coronary artery disease (CAD) genome-wide association (GWA) and diffusion tensor imaging (DTI) data.

Conclusions:

• The ARC method offers an improved approach to LFDR estimation by integrating covariate information.
• This method enhances statistical power and accuracy in complex biological and medical data analysis.
• The findings support the broader application of covariate-informed statistical methods in high-throughput studies.