Structural Equation Modeling: What Is It For and Why Is It Useful for Genetics?

The development of statistical methods for estimating genetic correlations from genome-wide association study (GWAS) summary statistics have produced ever-expanding “atlases” of genetic correlations among traits. As a... [ view full abstract ]

The development of statistical methods for estimating genetic correlations from genome-wide association study (GWAS) summary statistics have produced ever-expanding “atlases” of genetic correlations among traits. As a result, widespread pleiotropy is now taken to be the rule rather than the exception for complex human phenotypes. Structural equation modeling (SEM) is a statistical framework for testing hypotheses about the data-generating processes that give rise to observed patterns of variation and covariation. In this talk, I review the utility of SEM -- what is it typically used for and why it is useful for answering questions in genetics? Paying particular attention to the role of latent variables, I describe classes of problems typically addressed by SEM, such as assessing convergent and discriminant validity and ameliorating the consequences of imperfect measurement. As applied to genetic correlations among traits, these classic “psychometric” problems are, in fact, critical to our understanding of genetic architecture. Moreover, many psychological, behavioral, and disease phenotypes of interest to researchers in complex trait genetics were developed as latent constructs in an SEM framework and do not correspond neatly to a single, univariate measurement. Several previous methods have been developed to capitalize on genetic associations among traits to, e.g., boost statistical power for genetic discovery, but these methods have assumptions that can limit their applicability. Thus, there is a lacuna in the methodological landscape, one that we aim to fill with the method we introduce in this symposium – Genomic Structural Equation Modeling.