The traditional approach to causal discovery relies on experimentation: if an externally imposed change of variable A (e.g., a medical treatment) results in a change of variable B (e.g., health outcome), we conclude that A causes B. Over the last decades, an alternative approach has been suggested: under suitable assumptions, causal discovery is possible from purely observational data only, i.e., without relying on experiments. Constraint-based causal discovery methods have been proposed that promise to learn causal structure from purely observational data. However, strong assumptions and huge sample sizes are required for such approaches to yield reliable results. In this talk, I introduce Joint Causal Inference (JCI), a novel framework for causal discovery that elegantly unifies both approaches, thereby paving the way to reliable automatic causal discovery from big data. Compared with existing constraint-based approaches for causal discovery from multiple data sets, JCI offers several advantages: it allows for several different types of interventions in a unified fashion, it can learn intervention targets, it systematically pools data across different datasets which improves the statistical power of independence tests, and most importantly, it improves on the accuracy and identifiability of the predicted causal relations.
© 2017 K. Schulz