When making a decision based on observational data, a person’s choice depends on her beliefs about which correlations reflect causality and which do not. We model an agent who predicts the outcome of each available action from observational data using a subjective causal model represented by a directed acyclic graph (DAG). An analyst can identify the agent’s DAG from her random choice rule. Her choices reveal the chains of causal reasoning that she undertakes and the confounding variables she adjusts for, and these objects pin down her model. When her choices determine the data available, her behaviour affects her inferences, which in turn affect her choices. We provide necessary and sufficient conditions for testing whether such an agent’s behaviour is compatible with the model.