Maximizing association statistics over genetic models

The assessment of the association between a candidate locus and a disease may require the assumption of an inheritance model. Most researchers select the additive model and test the association with the Cochran-Armitage trend test. This test assumes a dose-response effect with regard to the number of copies of the variant allele. However, if there is reason to expect dominance or recessiveness in the effect of the variant allele, the heterozygous genotype may be grouped with one of the two homozygous, depending on the inheritance model, and a simple test on the 2 x 2 table can be used to assess independence. When the underlying genetic model is unknown, association may be assessed using the max-statistic, which selects the largest test statistic from the dominant, recessive and additive models. The statistical significance of the max-statistic has been previously addressed using permutation or Monte Carlo simulation approaches. We aimed to provide simpler alternatives to the max-test to make it feasible in large-scale association studies. Our simulations show that this procedure has an effective number of tests of 2.2, which can be used to correct the significance level or P-values. We also derive the asymptotic distribution of max-statistic, which leads to a simple way to calculate the significance level and allows the derivation of a formula for power calculations in the design of studies that plan to use the max-statistic. A simulation study shows that the use of the max-statistic is a powerful approach that provides safeguard against model uncertainty.