We used a joint 2-df LRT to combine information from the genetic marginal effect (β) and the gene–treatment interaction effect (δ) as the primary test. Specifically, this test constrained both β ≡ 0 and δ ≡ 0 under the null hypothesis and was shown previously to provide good power across a wide range of underlying true causal models (80). Using this model, we first determined the most appropriate genetic model by testing both dominant and additive genetic models. The best fitting genetic model as determined by the model yielding the lowest P-value, was then used in all the subsequent analyses. Regression analyses were performed using the R Statistical Program (81).
