A robustness study of parametric and non-parametric tests in model-based multifactor dimensionality reduction for...

Area of Science:

• Statistical genetics
• Bioinformatics
• Computational biology

Background:

• Statistical methods require validating model assumptions for reliable results.
• Epistasis detection using Model-Based Multifactor Dimensionality Reduction (MB-MDR) can be affected by violations of normality and homoscedasticity.
• Understanding these effects is crucial for accurate genetic association studies.

Purpose of the Study:

• To assess the impact of non-normality and heteroscedasticity on MB-MDR performance for detecting 2-locus epistasis.
• To evaluate the effectiveness of data transformations in mitigating these effects.
• To compare different association testing strategies within the MB-MDR framework.

Main Methods:

• A simulation study was conducted using pure epistasis models for quantitative traits.
• Data distributions included normal, chi-square, and Student's t with constant or non-constant variances.
• MB-MDR was applied with both standard Student's t-tests and Welch's t-tests, with and without trait transformations (logarithmic, standardization, rank-based).

Main Results:

• MB-MDR maintained type I error control and low false positive rates across tested conditions and association tests.
• MB-MDR using Welch's t-tests showed generally lower power compared to Student's t-tests.
• Rank-based transformations significantly increased power compared to other transformations.

Conclusions:

• Rank-transforming quantitative traits prior to analysis is recommended for MB-MDR screening of gene-gene interactions.
• Using Student's t-tests within MB-MDR is effective for association testing after rank transformation.
• This approach enhances the reliability and power of epistasis detection in genetic studies.