Background: While its importance is well recognized, it remains challenging to test genetic association in the presence of gene-gene (or gene-environment) interactions. A major technical difficulty lies in the fact that a general model of gene-gene interactions calls for the use of often a large number of parameters, leading to possibly reduced statistical power. An emerging theme of some recent work is to reduce the number of such parameters through dimension reduction. Wang et al.  proposed such an approach based on the partial least squares (PLS) for dimension reduction. They compared their method with several others using simulated data, establishing that their PLS test performed best. Unfortunately, Wang et al. did not include in their evaluations several powerful tests just recently discovered for analyzing multiple SNPs in a candidate gene or region. Methods: In this paper, we first extend these tests to the current context to detect gene-gene interactions in the presence of nuisance parameters, then compare these tests with the PLS test using the simulated data of Wang et al. . Results: It is confirmed that some other tests can be more powerful than the PLS test, though there is no uniform winner. Some interesting, albeit not new, observations are also made: some of the new tests are more robust to the large number of parameters in a model and may thus perform well; on the other hand, even for a purely epistatic genetic model, some of the tests applied to a logistic main-effects model without any interaction terms may be superior to that based on a full model that explicitly accounts for gene-gene interactions. Conclusion: The proposed statistical tests are potentially useful in practice.
- Genome-wide association study
- Logistic regression
- Main-effects model
- Score test
- Sum of squared score tests