The process of causal discovery concerns the mechanism by which an intelligent system learns to differentiate sequences of events on which the system bases explanation and control (i.e., causal relations) from the indefinitely many that are incidental (i.e., noncausal sequences). Just as the goal of an intelligent visual system is to infer a visuospatial representation of the distal 3-D world rather than features of a 2-D retinal image, the goal of an intelligent causal reasoner is to infer causal relations in the distal world rather than features of proximal stimuli such as covariation. This goal is unreachable without an a priori conviction that there exist causal relations in the world -- invariant relations that allow the prediction of the consequences of actions. The causal power theory of probabilistic contrasts (Cheng, 1997; Novick & Cheng, 2000) assumes that covariations are interpreted in terms of such unobservable invariant relations. This theory explains psychological phenomena regarding the discovery of simple and conjunctive causal relations that are inexplicable by a purely covariational approach, because the latter foregoes inference regarding distal relations. For the same reason, conventional statistics is inappropriate for testing causal hypotheses. .