Yes but in very large datasets, isn't it the case that you end up finding correlations for many things at very low p values? That is (one of the reasons) why effect size matters. Still, the "odds ratios are below 3" requirement seems excessive.