To interpret results, we ordered SNPs by primary test p-value and then formed r2 bins across the 226 SNPs, by chromosome, as follows. The correlation coefficients were computed separately in cases (r2case) and controls (r2control). The most significant SNP and all SNPs having min(r2case , r2control) ≥ 0.8 with it defined the first bin; these were removed from the binning process, which continued with the next most significant SNP that was not already binned. By ordering the SNPs by significance to define the bins, we enable the bins to better reflect the correlations with the most significant SNPs. Using min(r2case , r2control) is slightly more conservative (potentially leading to more bins) than using r2 from the combined sample, but in practice makes little difference. Alternative algorithms may be efficient for selecting tag SNPs, for example the greedy algorithm of (Carlson et al., 2004). However, such an algorithm might then relegate two SNPs into separate bins even though they are both correlated with a very significant SNP, giving the illusion that they are two distinct signals.
