Browsing by Author "Ruotsalainen, Sanni"

Genome-wide association studies have identified hundreds of genomic loci associated with a wide range of human conditions and quantitative traits, such as cholesterol level and diabetes. However, most of these studies have focused on analysing single traits, even the studies involving multiple related traits. Growing evidence for pleiotropy, where the same genetic locus is associated with multiple traits, supports the idea that multivariate methods could provide a remarkable boost in statistical power compared to univariate methods. In this thesis the main research question is to compare the multivariate Wald test to the corresponding univariate test, and to see when multivariate testing is more useful. My second research question is to compare the multivariate Wald test and another multivariate method called Canonical Correlation Analysis (CCA), and to see if they yield the same result. To examine these topics I performed a simulation study in which I simulated data set with 1,000 genotypes and 1,000 individuals. In addition I simulated bivariate phenotypes that were differently correlated with each other, and the genotypes. I performed the univariate Wald test for each trait against each genotype, and the multivariate Wald test for each trait pair against each genotype. I also performed the corresponding CCA to compare those results with the Wald test. In addition to the simulation study I performed the similar analyses for real data from The National FINRISK Study. I used three different blood lipid measurements, HDL-cholesterol, LDL-cholesterol and triglycerides as example traits, and 157 genomic loci previously known to associate with blood lipid levels. These blood lipid levels were appropriate example traits for this study because they are correlated differently with each other, and they are differently associated with the 157 genomic loci used here. Therefore I found many different combinations of correlation between traits, and directions of genetic effects for different traits. Based on my simulation studies I can say that the multivariate testing is never much worse in terms of power to detect associations than the corresponding univariate tests, and in some cases it is much more powerful. Thus there is no reason not to do the multivariate analysis first in case of studying multiple related traits. Multivariate testing is more powerful in cases where the correlation between the traits is large and the genetic effects for the traits show opposite directions compared to the trait correlation. The least effective multivariate testing is compared to univariate testing when the correlation between the traits is small, and the directions of genetic effects is consistent with the trait correlation. Based on my results multivariate Wald test and CCA yield the same results, with some minor approximation differencies in small sample sizes.