As we often do not know how a causal GV impinges on a phenotype, determining the most informative operationalisation of a trait for gene-finding purposes poses a challenge. Multiple studies [4]–[7] have shown that phenotypic data reduction, such as case-control status phenotypes or sum scores calculated across all distinguishable phenotypes, results in a considerable loss of statistical power to detect GVs in all but the special circumstance that 1) a single phenotypic dimension underlies the variance-covariance structure of the multivariate phenotypes (i.e., single common factor model), and 2) the GV directly affects this dimension (schematic representation Figure 1a). In this ideal unidimensional model, the underlying phenotypic dimension mediates the relationship between the GV and the multivariate phenotypes, and the univariate sum score is a good approximation of this dimension. However, many other genotype-phenotype models are plausible. For instance, the model could be multi-dimensional rather than unidimensional (Figure 1b–1c), and the GV effect could be specific to one of the phenotypes, rather than on the latent dimension (Figure 1d–1e). Recently, the field of psychology has witnessed a shift towards network models,
