Table 1.

Selection of brain measures for genetic analysis. In ENIGMA, the brain measures chosen for analysis had to be feasible to measure consistently and efficiently at a large number of sites, according to agreed protocols (available at enigma.ini.usc.edu). As power is limited in GWAS, various tactics may be useful in the future to boost the power to find genetic associations. Some of these are categorized here

Power enhancement approach	Principle	Pros and cons
1. Enhance the dataset	Increase the sample size (some GWAS studies now assess 100,000+ subjects (Lango Allen et al. 2010; Speliotes et al. 2010))	Identifies variants with smaller effect sizes, but is more costly
	Increase genomic coverage/sequencing	Picks up rarer variants, but requires even more subjects for power of low frequency variants. Also is more costly than genotyping common SNPs through genotyping arrays, but the cost is rapidly decreasing
	Increase the range of phenotypes studied	May be able to find a high effect size phenotype, but also need to correct for the number of measures assessed, which may be large (e.g., in “voxel-based” GWAS; Stein et al. 2010; Ge et al. 2012); if too many are assessed, power is low
2. Data reduction	Focus on candidate SNPs/genes, candidate pathways, candidate phenotypes	Avoids heavy statistical correction, but may miss unexpected variants or phenotypes
2.1. Based on classical genetics principles	Heritability screening—remove or de-emphasise measures with low heritability	This may empower genomic screens of complex phenotypes (e.g., genome-wide connectome-wide screens; Jahanshad et al. 2013a), see ENIGMA-DTI (Jahanshad et al., Jahanshad et al. 2013b).
	Genetic Clustering—find parts of an image or 3D cortical surface with common underlying genetic determination	GWAS on the resulting “genetic clusters” appears to have higher power than standard voxel-based approaches (Chiang et al. 2011, 2012; Chen et al. 2011; 2012)
2.2. Based on relevance to disease	Endophenotype ranking value (ERV; Glahn et al. 2012), aims to rank biomarkers in terms of their promise as endophenotypes for any heritable illness.	Balances the strength of the genetic signal for the endophenotype and the strength of its relation to the disorder of interest
2.3. based on using multivariate statistics	Use multiple predictors in the genome or image or both (reviewed in Hibar et al. 2011a; Thompson et al. 2013; Meda et al. 2012); Sparse regression (Vounou et al. 2010, 2012; Ge et al. 2012; Silver et al. 2012), compressive sensing, parallel ICA, machine learning methods	Can search both the image and the genome simultaneously Difficult to apply to distributed remote datasets for meta-analysis and may be difficult to interpret the biological relevance of the signal
3. Multimodality approaches Exploit joint information in several imaging modalities or other biomarkers at the same time	Multimodal data fusion using ICA (Calhoun et al. 2009) Seemingly Unrelated Regression to pool information across simultaneous models (SUR; Jahanshad et al. 2013c)	More work required to analyse mutliple data modalities at once (e.g., anatomical MRI and DTI)