DRAFT 3/4/07 Bayesian Generic Priors for Causal Learning

We present a Bayesian model of causal learning that incorporates generic priors on distributions of weights representing potential powers to either produce or prevent an effect. These generic priors favor necessary and sufficient causes. The NS power model couples these priors with a causal generating function derived from the power PC theory (Cheng, 1997). We test this and other alternative Bayesian models using the strategy of computational cognitive psychophysics, fitting multiple data sets in which several parameters are varied parametrically across multiple types of judgments. The NS power model accounts for a wide range of data concerning judgments of both causal strength (the power of a cause to produce or prevent an effect) and causal structure (whether or not a causal link exists). For both types of causal judgments, a generic prior favoring a cause that is jointly necessary and sufficient explains interactions involving causal direction (generative versus preventive causes). For structure judgments, an additional prior that a new candidate cause will be deterministic (i.e., sufficient or else ineffective) explains why people’s causal structure judgments are based primarily on causal power and the base rate of the effect, rather than sample size. Alternative Bayesian formulations that lack either causal power assumptions or generic priors for necessity and sufficiency proved inadequate. Broader implications of the Bayesian framework for human learning are discussed.