One way to avoid the potential loss of power associated with univariate conceptualisations of complex heterogeneous traits, is to adopt a multivariate method, which accommodates the originally multivariate nature of the phenotypic measure. Exploratory multivariate strategies, developed in GWAS context, include MultiPhen [11], and canonical correlation analysis [12], which is included in the GWAS software PLINK [13] (as canonical correlation analysis is identical to MANOVA with one GV treated as additive codominant (i.e., covariate), we use the term MANOVA here). MultiPhen uses ordinal regression to regress 0/1/2-coded GVs on a collection of phenotypes of any measurement nature (i.e., continuous, dichotomous, ordinal), and applies one omnibus test to test whether all regression weights in the model are together significantly different from zero. MultiPhen has been shown to outperform MANOVA when minor allele frequency (MAF) is low and the phenotypes are case-control status or non-normally distributed continuous variables [11]. Under most circumstances, however, MultiPhen and MANOVA yield very similar results in terms of power to detect causal GVs.
