To avoid inflated type I error and reduced power in genetic association studies, it is necessary to adjust properly for population stratification and known/unknown subject relatedness. It would be interesting to compare the performance of a principal component-based approach with a linear mixed model. Furthermore, with the availability of genome-wide sequencing data, the question of whether it is preferable to use common variants or rare variants for such an adjustment remains largely unknown. In this paper, we use the Genetic Analysis Workshop 18 data to empirically investigate these issues. We consider both a quantitative trait and a binary trait.