Replication using independent data is the gold standard. For this, we used the ACS data that had the heavy-smoking/light-smoking phenotype as proxy for the FTND derived nicotine dependence/non-dependence phenotype from COGEND. Our criteria for deciding that an RPM result is recapitulated in the independent dataset involve two factors. First, the joint RPM model should account for substantially more of the trait variance than the sum of the univariate RPM models in the independent dataset (we used an arbitrary threshold of 15% increase or greater). The second condition we require for replication is that the RPM model produced in the replication data should be consistent with the RPM model in the initial dataset. By this we mean that, in general, genotypes that are of high risk in the initial model should be high risk in the replication, and the same for low risk genotypes. We note that although this second replication condition is heuristic, (1) it is a higher dimensional analog for the routine procedure of checking univariate results to ensure that allelic effect is in the same direction in the
