To determine a “best model” for nominally significant SNPs, we evaluated whether an added gender-by-genotype interaction term is itself significant and also whether modeling a log-linear trend effect of the genotype is sufficient. If the gender-by-genotype interaction term was significant at the p < 0.05 level, the logistic regression was repeated with that term, and the “best model” p-value reported for that two df test. To determine whether modeling a non-multiplicative genotype effect gives a better fit than the primary model, we added a single dichotomous variable coding for heterozygote status, allowing deviation from an ordinal model. When this variable was significant, we evaluated other modes of inheritance: recessive (1 df), dominant (1 df), or genotype coded with 2 degrees of freedom to allow an arbitrary model, and report the “best model” p-value for the appropriate alternative model.
