Constraints and Nonconstraints in Causal Learning: Reply to White (2005) and to Luhmann and Ahn (2005)

Causal Knowledge and the Confusion Between Input and Output Distinguishing between a singularist and a regularity account of causal reasoning, White (2005) writes, The operation of a causal power is not a matter of probability or frequency. . . . In fact a power might never be exercised at all. . . . People can ascribe a causal power to a thing on the basis of knowledge of its nature even if they have no evidence for the operation of the power in question.” (p. 677) Three confusions concerning the problem of causal learning are involved here. First, White (2005) allows abstract prior causal knowledge in the input. Ascribing causal power from “knowing the nature” of a thing simply means that one is applying abstract prior causal knowledge. For the problem of causal learning, assuming that the input already contains the desired unknown—the domain-specific causal knowledge—would create a circular argument (e.g., Ahn, Kalish, Medin, & Gelman, 1995; Shultz, 1982; White, 2000). It seems that White is addressing a different research problem. Second, the fact that reasoners use abstract prior causal knowledge does not mean that regularity information is unnecessary for causal learning. One cannot know, for example, that the wind 2 The comparison between methods of administration suggested by White (2005) might be interpreted to mean that only the left column in Figure 2 is relevant to the assessment of a releasing condition and that our theory includes irrelevant information. It should be clear, however, that the right column is relevant: Without that column, one would be unable to tell whether ingestion is a simple cause of the side effect (Pattern A), noncausal by itself (Pattern B), or a releasing condition for the medicine to produce the side effect (Patterns B and C together). 3 To determine whether one conjunctive cause has more power than another (e.g., whether swallowing a medicine is more effective than injecting it), one would measure the power of each conjunctive cause separately using the standard power PC inference process and then compare the outputs. It is clear that such comparisons do not imply a process that contradicts the standard one. Figure 1. Testing ingestion as a releasing condition while holding “pill” constantly present. Figure 2. Outcome patterns for different methods of administering medicine holding the agent (pill or solution) constantly present. side effect; o no side effect. 696 COMMENTS