A comparison of methods to elicit causal structure

We compare two methods to elicit graphs from people that represent the causal structure of common artifacts. One method asks participants to focus narrowly on local causal relations and is based on the "make-a-difference" view of causality, specifically on an interventional theory of causality and so we call it "Intervention." It asks subjects to answer a series of counterfactual questions. The second method draws directly from the graphical aspect of Causal Bayesian Networks and allows people to consider causal structure at a more global level. It involves drawing causal graphs using an online interface called "Loopy." This method does not depend on a definition of causal relatedness. We use signal detection theory to analyze the likelihoods of people generating correct and incorrect causal relations (hit rates and false alarm rates, respectively) using each method. The results show that the intervention method leads people to generate more accurate causal models.