When correlation equals causation: A behavioral and computational account of second-order correlation learning in children

The present investigation reports two experiments that examined 2- and 3-year-old children’s ability to use second-order correlation learning—in which a learned correlation between two pairs of features (e.g., S and T, S and R) is generalized to the noncontiguous features (i.e., T and R)—to make causal inferences. Previous findings showed that 20- and 26-month-olds can use second-order correlation learning to learn about static and dynamic features in category and non-category contexts. The present investigation extends these findings by showing that a key prerequisite for detecting an indirect correlation between an objects’ surface feature and its causal efficacy is that children encode direct correlations on which the second-order correlation is based. Finally, to demonstrate that this reasoning does not require specialized causal-reasoning mechanisms, we also present the results of an autoencoder connectionist model, which provides a mechanistic account of the observed behavioral results.