Conditions (1)–(3) are reassuring because they mean that when a marker-exposure interaction exists, the marker must be associated with a causal variant. Furthermore, if the causal variant is independent of the exposure, then (4) shows that a marker-exposure interaction implies a causal variant-exposure interaction, at least under linear or log-linear regression (notably, this does not apply to logistic regression, although for rare outcomes βMX will approach 0). However, if there is dependence between the causal variant and the exposure, then a marker-exposure interaction does not imply a causal variant-exposure interaction. Therefore, tests of marker-exposure interaction are not valid tests of interaction between causal variant and exposure.