The spurious interactions we have described are not a serious problem when the aim is to construct epidemiological models of risk, perhaps for disease prediction, in which case model fit may be improved by interaction terms. The real difficulty is with inference of biological interaction from statistical models, and our observations add to established concerns over the interpretation of statistical interactions that are model dependent.3 We believe that additional caution is required in the interpretation of gene-environment interactions, to allow for the possibilities of gene-environment dependence and imperfect LD between marker and causal variant. We suggest that sensitivity analyses such as that described above ought to be routinely performed to reduce the possibility of false positive reports of interaction.