Building Compressed Causal Models of the World

A given causal system can be represented in a variety of ways. How do agents determine which variables to include in their causal representations, and at what level of granularity? Using techniques from Bayesian networks, information theory, and decision theory, we develop a formal theory according to which causal representations reflect a trade-off between compression and informativeness, where the optimal trade-off depends on the decision-theoretic value of information for a given agent in a given context. This theory predicts that, all else being equal, agents prefer causal models that are as compressed as possible. When compression is associated with information loss, however, all else is not equal, and our theory predicts that agents will favor compressed models only when the information they sacrifice is not informative with respect to the agent’s anticipated decisions. We then show, across six studies reported here (N=2,364) and one study reported in the supplemental materials (N=182), that participants’ preferences over causal models are in keeping with the predictions of our theory. Our theory offers a unification of different dimensions of causal evaluation identified within the philosophy of science (proportionality and stability), and contributes to a more general picture of human cognition according to which the capacity to create compressed (causal) representations plays a central role