Causes, coincidences, and theories.

摘要

The ability to infer causal relationships is central to the growth of both scientific and everyday knowledge. The fundamental problem in explaining how people infer causal structure from data is understanding how we learn so much from so little. I approach this problem by viewing causal induction as a rational statistical inference comparing causal models generated by a causal theory, and present a formal framework that can be used to develop computational models of human judgments. In this framework, which I call theory-based causal induction, a theory is defined as a hypothesis-space generator: it generates hypotheses about the causal structure underlying a set of events in the same way that a grammar generates hypotheses about the syntactic structure underlying a sentence. These hypotheses are then evaluated by Bayesian inference. This framework allows theories to impose strong constraints on the causal structures considered by a learner, making it possible to infer causal relationships from only small amounts of data. I show how this framework can be used to explain human judgments about causal relationships from contingency data, inferences about physical systems that operate in discrete and continuous time, and the role of coincidences in causal discovery. This analysis has implications for understanding the strengths and limitations of causal graphical models (also known as Bayesian networks) as an account of human cognition, the domain-specificity of causal reasoning, and the role of mechanism knowledge in causal induction.