Fig. 1. Simulation study results.

Method comparison, parameter estimation results and the behaviour of the false discovery rate (FDR). Models were estimated on a data set of M = 50,000 uncorrelated markers and N = 5000 individuals in 25 replicate simulations of 5 chains with 3000 iterations. Phenotypes were created from Generalised gamma distributions (see Supplementary note) using p = 500 causals markers and retaining heritability of h² = 0.5; independent data set had the same markers with N = 1000 other individuals. a Prediction accuracy of four methods when predicting to an independent data set given different generalised gamma distributions. The plot centres indicate the mean and error bars indicate the standard deviation of the correlations across simulations; (b) Mean precision values for each level of recall for four methods using Weibull phenotype (theta = 1); (c) Regression slope (true effect size ~ estimated effect size) when estimating non-zero marker effects given different theta values estimated with BayesW at each iteration across all simulations; (d) BayesW SNP heritability estimates given different generalised gamma distributions and different used mixtures at each iteration across all simulations; (e) relationship between the posterior inclusion probability (PIP) and false discovery rate (FDR) given different generalised gamma distributions for BayesW, for each PIP we present mean FDR values; (f) relationship between the PIP and FDR for a different number of mixture distributions used using Weibull phenotype and BayesW, for each PIP we present mean FDR values. In panels c–d, the bounds of the box show the interquartile range, centre shows the median and minimum and maximum indicate the 95% credibility interval.