Although other statistics from association results, such as regression coefficient beta, t statistics (beta divided by the standard error), and raw p-values, can index association strengths between SNPs and the phenotypes could be used as class indicators, we concluded that they are less or equivalently appropriate as class indicators compared to −log10(p). Raw p-values may not be appropriate because they may result in classes overly driven by SNPs with lower association strengths, where p-values are more distinguishable than p-values of higher association strengths. Given that the direction of effect of each SNP is determined by an arbitrary decision on the reference allele, and because we designed the analysis to classify SNPs based on their patterns of association strengths, signs of raw beta coefficients may introduce an unnecessary dimension into classification. Absolute values of beta coefficients still may not correctly represent the association strengths without taking into account their standard errors. Absolute values of beta coefficients divided by their standard errors, absolute values of t statistics, may better represent association strengths than raw beta. However, we expect that using absolute values of t statistics would be equivalent to using log10(p).
