Fig. 3

Collective inference performance. Ridge regression (RR) and CDA were compared using simulated data. We first sampled phenotype values of n individuals from the standard normal distribution. Restricting ourselves to the number of SNPs (m ≤ 20) allowing for the enumeration of all possible genotypes, we then assigned single-SNP and interaction parameters for m = 10 (a), m = 15 (b), and m = 20 SNPs (c) from normal distributions h _i ⁽⁰⁾ ~ N(0, 0.01), J _ij ⁽⁰⁾ ~ N(0, 0.01), h _i ⁽¹⁾ ~ N(0, 0.01), and J _ij ⁽¹⁾ ~ N(0.1, 0.01), under the dominant model. We next calculated the genotype distribution conditional on phenotypes for all possible genotypes, and chose a genotype for each individual based on this distribution. We repeated this sampling for 100 replicates. For each data set, we applied RR and CDA collective inference, using a single penalizer λ determined by optimizing R by cross-validation (right column). Power was defined as the proportion of replicates for which P < 0.05