Constraint-based analysis for causal discovery in population-based biobanks

Abstract

Availability of large genetic databases has led to the development of powerful causal inference methods that use genetic variables as instruments to estimate causal effects. Such methods typically make many assumptions about the underlying causal graphical model, are limited in the patterns they search for in the data, and there is no guide for systematic analysis of a large database. Here, we present cGAUGE, a new pipeline for causal Graphical Analysis Using GEnetics that utilizes large changes in the significance of local conditional independencies between the genetic instruments and the phenotypes. We detect cases where causal inference can be performed with minimal risk of horizontal pleiotropy. Moreover, we search for new graphical patterns to reveal novel information about the underlying causal diagram that is not covered by extant methods, including new direct links, colliders, and evidence for confounding. We present theoretical justification, simulations, and apply our pipeline to 70 complex phenotypes from 337,198 subjects from the UK Biobank. Our results cover 102 detected causal relationships, of which some are new and many are expected. For example, we detect a direct causal link from high cholesterol to angina and a feedback loop between angina and myocardial infarction. We also corroborate a recent observational link between asthma and Crohns disease. Finally, we detect important features of the causal network structure including several causal hubs such as intelligence and waist circumference.