Data integration for improved psychometric phenotypes in gene association tests

Tests of genetic associations with psychological traits require substantial power due to the expected small effect sizes. Two proposed avenues for gaining power in genetic tests are increasing sample sizes and... [ view full abstract ]

Tests of genetic associations with psychological traits require substantial power due to the expected small effect sizes. Two proposed avenues for gaining power in genetic tests are increasing sample sizes and improving phenotypic measures (Robinson, Wray, & Visscher, 2014). Common practice is to meta-analyse the results of gene-finding studies carried out across multiple cohorts through genetic consortia projects, thereby increasing sample-size. However, systematic differences in measurement and phenotype definition across cohorts may sacrifice the quality of the phenotypic measurement and reduce the power gain resulting from combining cohorts. We propose integrative data analysis (IDA) via psychometric measurement models for simultaneously increasing sample size and improving phenotype resolution. IDA for psychological phenotypes involves fitting a measurement model to pooled item-level data from all studies (Curran & Hussong, 2009). We present a bi-factor model that isolates a common phenotype score among the full item set while separating out cohort-specific measurement differences. We investigate the utility of IDA for tests of genetic effects in pooled studies with different phenotype item content, reflecting the common situation of different questionnaires across consortia partners. A simulation study under a series of underlying phenotype measurement conditions demonstrates that the bi-factor model for IDA can increase power as much as 11%, compared to meta-analysis using sum scores of truncated item subsets in each cohort. Power gains are greater when measurement differences exist across partners, as these differences are captured in the IDA model but not in the formation of sum scores. The benefits of IDA are maximized with the collection of a very small de novo sample with complete data, which we call a phenotypic reference panel. Implications for GWAS meta-analyses are discussed.