Unlike traditional P-value interpretations, it provides an error rate that depends on the precise value of the observed P-value, but also does not require any Bayesian assumptions

The approach is to compute the chance of incorrectly finding further evidence for the alternative hypothesis when the null hypothesis is true, so I call it the 'False Evidence Rate'

Final note: the method is generic and works for any sort of big data effect estimate panel (not just GWAS), provided that the sampling error is roughly normal

However, we think that effect-size based methods are likely to produce better prioritisations, as lower-frequency genetic signals at higher effect size may be excessively down-weighted by focussing on signifiance. This will especially be the case as GWAS samples grow to even larger sizes

This is not to be too negative about the use of P-value (or FDR) significance for priorising GWAS signals. We found that 80% of variance in our effect sizes rankings were explained by P-values, which shows that they do a decent job.

We provide results showing that ranking by effect size prioritises signals that belong to the most important biological pathway in T1D therapy, the IL2 pathway. We also show that it enriches for genes that are implicated as causal by Mendelian Randomisation

Our Bayesian method is, in a sense, 'prior free' in that it only models the likelihood of the data as a flexible curve, without requiring the form of prior to be specified. Bayesian results are derived entirely from the liklihood fitting to the GWAS summary stats

It is being increasingly appreciated that, at some stage, most regions of the genome will show evidence for non-zero effects on a given trait, as sample sizes increase. It is therefore important to be able to prioritise by biological importance, for which effect size is likely to be a useful proxy