A robustness study to investigate the performance of parametric and non-parametric tests used in Model-Based Multifactor Dimensionality Reduction Epistasis Detection.

Model-Based Multifactor Dimensionality Reduction (MB-MDR) is data mining technique to identify gene-gene interactions among 1000nds of SNPs in a fast way, without making assumptions about the mode of genetic interactions. By construction, one of the implementations of MB-MDR involves testing one multi-locus genotype cell versus the remaining cells, hereby creating two imbalanced groups for trait distribution comparison. To date, for continuous traits, we have adopted a standard F-test to compare these groups. When normality assumption or homoscedasticity no longer hold, highly inflated results are to be expected. The power and type I error control of MB-MDR under these assumptions has been thoroughly investigated in Mahachie John et al [1].

The aim of this study is to assess, through simulations, the effects of ANOVA model violations on the performance of Model-Based Multifactor Dimensionality Reduction (MB-MDR). We quantify their effect on MB-MDR using default options, but at the same time introduce alternative options with increased performance. The better handling of imbalanced data using robust approaches [2] within a MB-MDR context is exemplified on real data for asthma-related phenotypes.

1. EJHG (2011), Early view
2. David Freedman, Statistical Models: Theory and Practice, Cambridge University Press (2000), ISBN 978-0521671057