The Role of Causal Models in Statistical Reasoning

When making judgments based on statistical data, people have been shown to be poor at probabilistic reasoning, specifically Bayesian inference. While recent research shows people perform better when information is provided in a natural frequency format, we find this result’s explanatory reach limited, both for explaining people’s judgment failures and as a theory of human reasoning under uncertainty. Most prior studies demonstrating probabilistic reasoning deficits gave their subjects probabilistic inference problems that did not explain the causal mechanisms behind the provided statistics. Our research shows that when questions are posed that explain the causal structure of the domain, subjects perform significantly better. Specifically, base rate neglect can be made to virtually disappear when the content of a question reflects the true causal structure of the domain. We propose that causality is essential to probabilistic reasoning, and that without the opportunity to incorporate statistical data into a consistent theory of the causal structure of a domain, the typical person will have trouble performing normatively correct Bayesian inference.