Nodes from Arrows: Independence Assumptions in Human Causal Induction

-----

Causal inferences occur in contexts that are like rivers—one never steps into the same context twice. A key question, therefore, is how knowledge about a putative cause generalizes across the many stimulus configurations in which it occurs. According to the Causal Power view (Cartwright, 1987; Cheng, 1997), learners filter statistical regularities through a lens of a priori but defeasible assumptions about invariant causal capacities. Building on this view, we propose that violations of the invariance assumption, i.e., causal interactions, prompt reasoners to revise their representation of the cause variable, in order to preserve the invariance condition (Liljeholm, 2015). This is in contrast with associative learning theories, which either fail to detect interactions (e.g., Rescorla & Wagner, 1972) or reject invariance entirely, relying solely on feature similarity for generalization (e.g., Pearce, 1987). Our approach also diverges from recent hierarchical Bayesian models that propose a generic adoption of any arbitrary integrating function that is supported by the available data (e.g., Lucas & Griffith, 2024). To test our hypothesis, we designed a causal learning task that included both invariant and interacting causes – in each case, transfer tests confirmed that participants revise their representations of putative causal variables but stay faithful to the generic invariance assumption, consistent with the Causal Power view. A causal network is sketched and found to outperform both associative and Bayesian accounts. Future work will focus on integration with Deep Learning, which largely lacks causal representations.

-----