Fig. 4: Bias-variance tradeoff for family-based estimators.
From: Family-based genome-wide association study designs for increased power and robustness
a–d, The simulated datasets used in Fig. 2 are used for this demonstration: 2,000 independent sibling pairs and 18,000 singletons in each of two subpopulations with different levels of Fst (Methods): Fst = 0 (a), Fst = 0.001 (b), Fst = 0.01 (c) and Fst = 0.1 (d). The effective sample size (x-axis) is defined relative to that of the sib-difference estimator (Table 2) and should be equal to 1 (vertical dashed line) for the robust/sib-difference estimators—which are equivalent here—and higher than 1 for the other estimators. Bias (y-axis) is measured as the nonsampling variance (Methods) relative to that for standard GWAS with Fst = 0.001, and is expected to be above 0 (horizontal dashed line) when there is bias due to population stratification. Error bars display a 95% jackknife confidence interval over 20,000 SNPs. See Extended Data Figs. 3 and 4 for plots including the standard GWAS estimator and a sibling-only scenario (that is, no singletons).
