Limits...
Greater power and computational efficiency for kernel-based association testing of sets of genetic variants.

Lippert C, Xiang J, Horta D, Widmer C, Kadie C, Heckerman D, Listgarten J - Bioinformatics (2014)

Bottom Line: Set-based variance component tests have been identified as a way to increase power in association studies by aggregating weak individual effects.We compared a standard statistical test-a score test-with a recently developed likelihood ratio (LR) test.On synthetic data, we find that the LR test yielded up to 12% more associations, consistent with our results on real data, but also observe a regime of extremely small signal where the score test yielded up to 25% more associations than the LR test, consistent with theory.

View Article: PubMed Central - PubMed

Affiliation: eScience Research Group, Microsoft Research, Los Angeles, CA, 90024 and eScience Research Group, Microsoft Research, Redmond, WA, 98052, USA.

Show MeSH
Power on synthetic data for each method in each setting, for the lowest signal strength, . Fraction of tests deemed significant across various significance levels for each method is shown on the vertical axis. The threshold for significance is shown on the horizontal axis. Other signal strengths are shown in Figure 2 and Supplementary Figures S1 and S2
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4221116&req=5

btu504-F1: Power on synthetic data for each method in each setting, for the lowest signal strength, . Fraction of tests deemed significant across various significance levels for each method is shown on the vertical axis. The threshold for significance is shown on the horizontal axis. Other signal strengths are shown in Figure 2 and Supplementary Figures S1 and S2

Mentions: After establishing control of type I error, we then systematically investigated power, using four different levels of effect size for the causal SNPs (which were precisely those SNPs being tested), of , and across a range of significance thresholds , and (the same thresholds used for the type I error experiments). We found that for the lowest signal strength (), the score test yielded slightly more power than the LR test, consistent with the notion that the score test is locally optimal (Fig. 1). For the other signal strengths (), we found that the LR test yielded more power than the score test at each level (Fig. 2, Supplementary Figs S1 and S2), and in aggregate (Supplementary Fig. S3). The setting with the largest gain for the score test ( common variants, binary phenotype) showed a 25% relative gain in power for the score test over the LR test. However, this setting has so little signal that even for the score test, power was only 4%. The setting with the largest gain for the LR test ( common variants, Gaussian phenotype) showed a 12% relative gain in power for the LR over the score test, consistent with our real-data experiments.Fig. 1.


Greater power and computational efficiency for kernel-based association testing of sets of genetic variants.

Lippert C, Xiang J, Horta D, Widmer C, Kadie C, Heckerman D, Listgarten J - Bioinformatics (2014)

Power on synthetic data for each method in each setting, for the lowest signal strength, . Fraction of tests deemed significant across various significance levels for each method is shown on the vertical axis. The threshold for significance is shown on the horizontal axis. Other signal strengths are shown in Figure 2 and Supplementary Figures S1 and S2
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4221116&req=5

btu504-F1: Power on synthetic data for each method in each setting, for the lowest signal strength, . Fraction of tests deemed significant across various significance levels for each method is shown on the vertical axis. The threshold for significance is shown on the horizontal axis. Other signal strengths are shown in Figure 2 and Supplementary Figures S1 and S2
Mentions: After establishing control of type I error, we then systematically investigated power, using four different levels of effect size for the causal SNPs (which were precisely those SNPs being tested), of , and across a range of significance thresholds , and (the same thresholds used for the type I error experiments). We found that for the lowest signal strength (), the score test yielded slightly more power than the LR test, consistent with the notion that the score test is locally optimal (Fig. 1). For the other signal strengths (), we found that the LR test yielded more power than the score test at each level (Fig. 2, Supplementary Figs S1 and S2), and in aggregate (Supplementary Fig. S3). The setting with the largest gain for the score test ( common variants, binary phenotype) showed a 25% relative gain in power for the score test over the LR test. However, this setting has so little signal that even for the score test, power was only 4%. The setting with the largest gain for the LR test ( common variants, Gaussian phenotype) showed a 12% relative gain in power for the LR over the score test, consistent with our real-data experiments.Fig. 1.

Bottom Line: Set-based variance component tests have been identified as a way to increase power in association studies by aggregating weak individual effects.We compared a standard statistical test-a score test-with a recently developed likelihood ratio (LR) test.On synthetic data, we find that the LR test yielded up to 12% more associations, consistent with our results on real data, but also observe a regime of extremely small signal where the score test yielded up to 25% more associations than the LR test, consistent with theory.

View Article: PubMed Central - PubMed

Affiliation: eScience Research Group, Microsoft Research, Los Angeles, CA, 90024 and eScience Research Group, Microsoft Research, Redmond, WA, 98052, USA.

Show MeSH