Current sample sizes are clearly adequate for testing association of a polygenic score in a replication sample, as long as full size samples are used for both training and testing. This is already apparent from the extraordinary significance levels reported in the seminal studies [3], [6], but here it is shown that those results are compatible with realistic genetic models and are not necessarily explained by analytic biases that accumulate across SNPs. This had been previously shown by the ISC study, which simulated plausible genetic models and showed that they led to similar results to those observed in the data [3]; here, the result is shown directly for the common quantitative model, without recourse to simulations. Studies that split a single sample into cross-validation subsets have been less successful [8], [14], but here it is shown that this could be explained by their limited sample sizes, and more encouraging results for the same traits could be obtained with modestly increased samples.
