The most common (and probably most effective) way is to correct for Principal Components derived from PCA on genome-wide SNP data. See: http://www.nature.com/ng/journal/v38/n8/abs/ng1847.html
Be sure to check that the principal components actually represent ancestry (for example by correlating them with geography), because if you have any batch effect or systematic quality differences in your dataset, the PCA is likely to pick this up as well (batch effects should be accounted for as well by the way!).
Another commonly used correction is genomic control, which decreases test-statistics for all association tests, but has lately been under scrutiny, because it tends to overcorrect when you have many real signals.