Methods for Large-Scale Genetic Association Studies.

Abstract

With the increasing availability and decreasing costs of high-throughput genotyping, contemporary genetic association studies now incorporate vast quantities of information. Major advances in genotyping technology have led to higher throughput at lower costs, and greater accuracy and completeness. These advances bring with them new questions, including 1) how to best adjust for the multiple testing problem given the likely correlation between tests involving dense markers, and 2) what levels of genotyping quality can be expected, and what levels can be tolerated in an association testing framework. We first address the issue of adjustment for the many tests performed, given the high levels of correlation that are typical of association studies involving dense markers. We present PACT (P-value Adjusted for Correlated Tests), an estimator analogous to Bonferroni or Sidak adjustment which accounts for the correlation between tests. We show through simulation that PACT can attain the accuracy and power of permutation tests thousands of times faster. We next extend our work on PACT so that it may be applied to meta-analyses involving correlated tests. We describe extensions to four common study designs, and show through simulation that these methods provide valid tests with greater power than methods which do not account for correlation. Finally, we investigate the nature of genotyping error and missing data for a variety of common SNP genotyping platforms in two datasets where replicate genotyping has been performed. We find that the rates of error and missingness vary depending on an individual's true genotype, and that heterozygotes and minor allele homozygotes are more prone to errors and missingness on most platforms. We show that differential rates of genotype error and missing data can invalidate the commonly used test of equal allele frequencies. We use simulation to assess the impact of the observed distribution of errors and missing data on false-positive rates in a genome-wide association context. We find that the impact varies depending on 1) whether the underlying association test is an allele frequency test, a test for trend, or a family-based association test and 2) whether appropriate quality control measures are applied.