According to a recent paper in the American Journal of Human Genetics, most genome-scan studies looking for loci that influence complex human diseases or traits, overestimate how much the loci that are found contribute to such phenotypes. This might not come as news to those quantitative geneticists who map quantitative trait loci (QTL) in domesticated or experimental species, but, for many human geneticists, this finding might come as a rather rude awakening. Importantly, the knock-on effect of such overestimations is that follow-up studies can underestimate the power required to replicate the original study, which might be why so many replication studies come up empty-handed.

The primary goal of a genome scan is to find the genetic loci that influence a particular phenotypic trait. If evidence for a locus is found, it is also very useful to know the locus effect size—how much of the heritability of the phenotype is attributable to variation at that locus. And that's where the problem lies. The position of the locus is estimated by the LOD score, which measures the likelihood of linkage at any given location, but because the LOD score is positively correlated with the effect size of the locus, the location and the heritability of the locus cannot be estimated independently from the same data set.

To study this phenomenon—which has been well documented in experimental species—in the context of human genetics, Göring et al. simulated and mathematically analysed a fairly typical genome scanning experiment: the hypothetical population comprises 1,000 families, each with two offspring; the trait being modelled has a heritability of 0.5, with variable numbers of QTL; genotypic data are available from markers at 2-cM intervals; and there are no complicating factors such as epistatic or gene–environment interactions. But even with this well-behaved data set, the authors found that there is no way to reliably estimate the heritability effect of any QTL. Their mathematical analysis of the problem shows that the QTL effect is always overestimated, although this bias does fall away with increasing sample size.

So, if a published genome scan contains an estimate of the effect of a particular QTL, calculated from a single data set, it will almost certainly be inflated. When a follow-up study fails to replicate the first report of a QTL, it might be that such a locus exists, but that the follow-up study did not have enough power to detect it—so it's not all bad news. However, the way forward could be daunting. The authors suggest that the most realistic solution is to use independent data sets for the estimation of location and effect size, and that means big samples—very big samples indeed.