Linked DNA variants reduced estimated independent tests in genetic association study
Accounting for linked DNA variants cut the estimated number of independent tests from 500,000 to 67,000 for an Affymetrix panel and from 317,000 to 82,000 for an Illumina panel in the CEPH Utah population. The study, published in BMC Genomics on 31 October 2008, proposed statistical thresholds that reflect those links. Duggal P and colleagues examined genome wide association studies, which search across the genome for common genetic variants that influence health and disease. Their approach adjusted for the fact that many tested variants are correlated.
Testing hundreds of thousands of genetic markers creates opportunities for chance findings. Researchers therefore need to control the family-wide Type 1 error: the risk of at least one false positive across the tests. A Bonferroni correction tightens the statistical cutoff according to the number of tests. But counting every marker separately would overcorrect when the markers were not independent, the authors said.
The markers were single nucleotide polymorphisms, or SNPs, differences at individual positions in DNA. Many fall within blocks of linkage disequilibrium, where variants occur together more often than would be expected by chance. The authors approximated the number of independent SNPs by counting 1 SNP per block and adding all SNPs outside those blocks. This gave them an estimated number of independent tests to use in the correction.
For the Affymetrix 500 K and Illumina 317 K panels, they recommended p-value thresholds of 10^-5 for "suggestive" associations, 10^-7 for "significant" associations and 10^-8 for "highly significant" associations. These statistical cutoffs were intended to control the risk of false positives while accounting for correlated markers.
For the Phase II HapMap CEPH Utah and Yoruba populations, the recommended thresholds were different: 10^-6 for "suggestive", 10^-7 for "significant" and 10^-9 for "highly significant" associations. The recommendations were based on an approximation of the effective number of independent SNPs, rather than treating every SNP tested as a separate, independent test.