Related Experiment Videos
Bias in the estimation of false discovery rate in microarray studies
Yudi Pawitan1, Karuturi R Krishna Murthy, Stefan Michiels
1Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Stockholm, Sweden. yudi.pawitan@meb.ki.se
Bioinformatics (Oxford, England)
|August 18, 2005
Summary
This study introduces a new mixture model to accurately estimate the proportion of non-differentially expressed genes (pi(0)) in microarray analysis. This improved estimation reduces bias and increases statistical power in identifying significant gene expression changes.
Area of Science:
- Genomics
- Statistical Bioinformatics
Background:
- The false discovery rate (FDR) is crucial for microarray studies, but its accuracy depends on estimating pi(0), the proportion of non-differentially expressed genes.
- Current methods often overestimate pi(0) and FDR, leading to reduced statistical power in identifying significant genes.
Purpose of the Study:
- To develop an improved method for estimating pi(0) in two-sample microarray comparisons.
- To address the bias in standard pi(0) estimation, particularly when pi(0) is far from 1 or test statistic non-centrality parameters are small.
Main Methods:
- Derived a natural mixture model for the test statistic in two-sample comparisons.
- Developed an explicit bias formula for standard pi(0) estimation.
- Proposed a practical, likelihood-based procedure for improved pi(0) estimation using the mixture model.
Main Results:
- Identified significant bias in standard pi(0) estimation under specific conditions.
- Demonstrated that mixture-model estimates of pi(0) are less biased than standard estimates through simulations.
- Explained discrepancies between non-parametric and model-based pi(0) estimates.
Conclusions:
- The proposed mixture model and likelihood-based procedure offer a more accurate estimation of pi(0) for microarray data.
- This improved estimation enhances the reliability of FDR calculations and increases statistical power.
- An R-package, OCplus, is available for implementing these methods.