We investigated selection and analysis of tag SNPs for genome-wide association studies by specifically examining the relationship between investment in genotyping and statistical power. Do pairwise or multimarker methods maximize efficiency and power? To what extent is power compromised when tags are selected from an incomplete resource such as HapMap? We addressed these questions using genotype data from the HapMap ENCODE project, association studies simulated under a realistic disease model, and empirical correction for multiple hypothesis testing. We demonstrate a haplotype-based tagging method that uniformly outperforms single-marker tests and methods for prioritization that markedly increase tagging efficiency. Examining all observed haplotypes for association, rather than just those that are proxies for known SNPs, increases power to detect rare causal alleles, at the cost of reduced power to detect common causal alleles. Power is robust to the completeness of the reference panel from which tags are selected. These findings have implications for prioritizing tag SNPs and interpreting association studies.