| This thesis concerns the drawing of inferences about variables in complex systems of causal relationships---in particular, inferences about the probability that an unobserved variable has a certain value (e.g., the probability that a symptom or a cause is present) given information about the values of other variables to which it is causally related, directly or indirectly. A normative account of this sort of inference was derived straightforwardly from graphical causal model theory (Pearl, 2000; Spirtes et al., 2000). Central to this account is a principle known as the causal Markov condition, which specifies the conditions under which variables are relevant to inferences about one another. In many cases it predicts that indirectly related variables are irrelevant to (or independent of, or "screened off" from) one another. The normative account was contrasted with an adaptive hypothesis, on which reasoners allow for possible hidden common causes and hidden paths between known variables by assigning inferential support to observed variables more liberally than predicted by the causal Markov condition. This phenomenon is termed nonindependence.; Several experiments were conducted in which participants drew inferences about unobserved variables in complex causal models. Results of all experiments showed nonindependence, consistent with the adaptive hypothesis. Experiments 2--4 were designed to distinguish between three forms of the adaptive hypothesis, each of which specifies a different distribution of inferential support over normatively irrelevant (screened-off) variables. Though results of these experiments were mixed, they tended to favor a proximity theory, in which inferential support is assigned to normatively screened-off variables as a function of their proximity, in known causal structure, to the variable about which an inference is being made. Experiment 5 yielded some suggestive evidence that reasoners with expertise in different scientific disciplines (astronomy, cell biology) show different degrees of nonindependence.; An approach to modeling natural inference from complex causal models is sketched. Also, several key phenomena predicted by recent psychological theories built on the graphical causal model framework are reviewed, and it is shown that the causal Markov condition, which is central to this framework, may be too general for the explanatory work it has been asked to do. |
