Pragmatic Explanation of Causal Inference from Correlational Statements
Abstract
People sometimes infer causal relationships from correlational evidence. For example, upon hearing cancer is associated with smoking, listeners might conclude smoking causes cancer. Without background information to inform priors, readers tend to interpret subjects of associated with as effects, rather than causes. We replicate this preference across multiple correlational predicates, and then derive it within a rational speech acts model of pragmatic inference. Central to our analysis is the observation that English asymmetrically constrains speakers: a speaker who plans to utter an effect in subject position has fewer, higher-cost alternatives compared to a speaker who plans to utter a cause in subject position. This asymmetry drives the pragmatic listener to infer that the use of a symmetric predicate signals that the subject is an effect. Our analysis additionally accounts for some of the variability in preferences for the subject-as-effect interpretation across different correlational predicates.