larger sample size and is more genetically similar to the whole GIANT meta-analysis sample relative to the QIMR cohort, where ancestry is restricted to the British Isles14. We could also only consider results that were consistent using two independent cohorts as reference samples. However, this might be too conservative, as some real associations identified using one cohort with P values that just passed the arbitrary cutoff value of 5 × 10–8, might be eliminated from the analysis in another cohort due to random errors in the estimates of LD correlations. Our method is not limited to meta-analysis summary data but can also be applied to a single GWAS cohort with individual-level genotypes, in which case, the whole discovery sample is used as the reference sample, and our method then becomes equivalent to a multiple regression analysis (Online Methods and Supplementary Fig. 6). In this case, the automated stepwise selection procedure implemented in our software tool19, which has, to our knowledge, not been implemented in any other GWAS analysis tools in the public domain, would still be useful for data applications.
