Chunk #42 — Reasons to be Concerned about the Published cGxE Literature — Problems with the Recipe: Statistical Concerns in cGxE Research — The selection of model
As with choice of transformation, choice of how to model interactions can profoundly affect evidence for them. In particular, there has long been debate in epidemiology regarding the relative utility of risk differences versus risk ratios. That is, if the rate of illness is 10 and 50 per 10,000 in groups unexposed and exposed to some risk factor, are we more interested in the risk difference (40 new cases per 10,000) or the risk ratio of 5? From a practical perspective (e.g., public heath impact, focus for possible prevention, advice to patients), the risk difference approach has much to recommend it. However, the risk ratio approach is the more dominant in large part because it is easily implemented statistically in logistic regression. This distinction is critical because it defines the baseline model from which we assess interactions. In a risk difference framework, an interaction reflects a deviation from a model in which risk factors add together. In a risk ratio framework, an interaction reflects a deviation from a model in which risk effects multiply. Accordingly, the usage of logistic regression
