Causal Models and How to Refute Them

Directed acyclic graph models (DAG models, also called Bayesian networks) are widely used in the context of causal inference, and they can be manipulated to represent the consequences of intervention in a causal system. However, DAGs cannot fully represent causal models with confounding; other classes of graphs, such as ancestral graphs and ADMGs, have been introduced to deal with this using additional kinds of edge, but we show that these are not sufficiently rich to capture the range of possible models. In fact, no mixed graph over the observed variables is rich enough, regardless of how many edges are used. Instead we introduce mDAGs, a class of hyper-graphs appropriate for representing causal models when some of the variables are unobserved. Results on the Markov equivalence of these marginal models show that when interpreted causally, mDAGs are the minimal class of graphs which can be sensibly used. Understanding such equivalences is critical for the use of automatic causal structure learning methods, a topic in which there is considerable interest. We elucidate the state of the art as well as some open problems.