Title: Causal inference with directed acyclic graphs: A case study in psychosis
Authors: Giusi Moffa - Institute of clinical epidemiology and biostatistics, University of Basel; and UCL Division of Psychiatry (Switzerland) [presenting]
Jack Kuipers - ETH Zurich (Switzerland)
Abstract: Directed acyclic graphs (DAGs) are common tools to describe causal mechanisms across different fields, ranging from social science to biology. Traditionally they have been used in forward causation to estimate the effects of causes given a postulated causal structure informed through domain experts. Thanks to computational progress in structure learning for Bayesian networks, DAGs have now also gained popularity in reverse causation. Inferring the DAG structure from observational data allows us to gain insights about putative causal mechanisms, though only under very strict assumptions. Given the networks we learn from the data we can then derive putative intervention effects. However, to ensure robust inference it is essential to account for the uncertainty in the estimation of the DAG structure. This is now possible thanks to substantial advance in sampling of Bayesian networks from their posterior distribution given the data. As a result, we can follow a fully Bayesian approach to derive a posterior distribution of putative causal effects. We focus specifically on binary variables and present a case study in Psychosis. The method applies both to cross-sectional data as well as to longitudinal data, when we consider dynamic Bayesian networks.