A Review of Automated Procedures for the Discovery of Causal Relations and the Use of Causal Hypotheses in Prediction

Clark Glymour

The discovery tasks that have been best studied in statistical and computer science concern the extraction of statistical or rule based regularities from various sorts of data streams. In the natural and social sciences, however, practitioners want (and often claim) insight into the causal dependencies and mechanisms that produce observed patterns and role instantiations. In policy contexts information is wanted that will enable correct prediction about relevant outcomes of interventions or actions in a system or population. Both epidemiology and economics are concerned with what will happen if alternative possible policies are broadly applied; the same is true of businesses, governments and many other organizations. Statistical or rule based patterns do not directly give any such information. For example, lung cancer is correlated with a history of slightly discolored fingers, but many interventions to prevent discoloration--requiring everyone to wear gloves, for instance--will have no effect on lung cancer or on lung cancer rates in the population so treated. In many contexts causal knowledge is limited, incomplete, or simply erroneous. The standard methods to extend causal knowledge involve experimental procedures in which various conditions are deliberately manipulated. But in many contexts--astronomy, geology, meteorology, epidemiology, human affairs, parts of engineering, parts of biology, and occasionally even in chemistry and physics--the requisite experiments cannot (or cannot all) performed for practical or economic or ethical reasons. So what is wanted are reliable methods for combining limited prior knowledge with nonexperimental (or partially non-experimental) data to extend causal knowledge or at least to guide us in deciding which experiments are worth performing; in other words, reliable methods are wanted for extracting causal explanations of observed statistical or role based patterns. In addition, algorithms are wanted for predicting, whenever possible, the outcomes of interventions given partial and possibly quite incomplete causal knowledge and associate partial specification of probabilities.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.