Manipulating prior causal beliefs affects the formation of apparent causal illusions.

Non-contingent learning data can produce an apparent "illusory" or "false" perception of causality ("causal illusion"), especially if a potential cause and effect frequently co-occur - a phenomenon known as the "outcome-density bias." In contrast to the prominent bias view, Bayesian models of causal induction explain these causal illusions as the result of a rational learning process in which observed data fail to fully override non-zero causal priors. Convincing evidence for this rational explanation requires an experimentally manipulated effect of causal priors - which has so far been lacking. We report four experiments (N=1860) supporting the Bayesian interpretation. We successfully manipulated participants' causal priors either through visually conveyed causal mechanism information or through statistical base-rate information, and found that both manipulations influenced the degree to which participants reported a causal relation after having processed non-contingent learning data. The results were in line with Bayesian updating: lower causal priors yielded lower post-learning beliefs in a causal connection, effectively reducing the "causal illusion." The results strengthen computational models implementing a rational Bayesian view of causal induction.