The problem. A plain association test assumes samples are independent. In real cohorts they are not: people share ancestry and sometimes family, and that structure creates correlations that look like association signal but are not. Mixed models can absorb this by modeling a genetic relatedness matrix, but the exact computation was too slow to run at every variant across a genome.

The idea. GEMMA makes the standard linear mixed model efficient enough for genome-wide use. It fits a random effect built from a kinship matrix, which captures how related each pair of samples is, so the shared background is accounted for and only the variant-specific signal is tested. The contribution is mostly computational: a reformulation that avoids redoing the expensive matrix work at every variant, turning a method that was correct but impractical into one you can actually run.

Why it matters. Structure is the central confounder in association studies, the thing that produces false hits if you ignore it. This is the same worry PLINK checks for with stratification, handled here inside the model rather than as a filter. The relatedness matrix idea also connects to the heritability work in the next paper, which reuses it for a different question.

Verdict. A foundational method for structure-aware association testing. Read it for why mixed models became the default, and for the trick that made them fast.