Causal Reasoning Through Intervention
York Hagmayer, Steven A. Sloman, David A. Lagnado, Michael R. Waldmann · 2007
Abstract In causal reasoning, the observation of an event supports different inferences than an intervention that generates the event. An intervention breaks the connection between the manipulated event and its normal causes. Therefore, in contrast to an observation, an intervention prevents diagnostic inferences about the causes of the event. This chapter shows how causal Bayes nets can be used to model observations and interventions. Empirical findings are then presented demonstrating that people are highly sensitive to this distinction: (i) Their inferences conform to the distinction when reasoning counterfactually; (ii) they derive different predictions for novel observations and previously unobserved interventions when the underlying causal model entails different outcomes; and (iii) their decisions differ depending on whether an evidential statistical relation between an action and an outcome will be sustained if the action is deliberately chosen (i.e. intervened on). Other non-causal theories of learning and reasoning are not able to explain these findings.