• Two- to 4-year-olds can use second-order correlation learning to make causal inferences; • Two- to 4-year-old children can detect, extract, and encode second-order correlations in a category-like context involving several objects; • A connectionist computational model demonstrates that second-order correlations can emerge through shared representational overlap. Second-order correlation learning—or the capacity to infer an indirect relation between features based on separate direct relations—has been studied extensively in children older than 4 years of age. Comparatively less research has examined whether children 4 years of age and younger can engage in this process. One of the few studies that assessed second-order correlation learning in children younger than this age showed that 2- to 3-year-olds can use this form of learning to make causal inferences in a discrimination context (e.g., Benton et al., 2021 ). The present study extended this work in two key ways. First, it tested whether 2- to 4-year-olds could detect and encode second-order correlations that are embedded in a cognitively more demanding category-like context consisting of multiple objects and then use those correlations to make causal inferences. Second, it explored a possible mechanistic basis of second-order correlation learning through a connectionist computational model. The results of the behavioral study revealed that children can detect second-order correlations among several distinct objects and use them to make causal inferences, and the computational model captured children’s performance and elucidated a possible mechanism through which such correlations might emerge in the mind. Taken together, the results demonstrate that children as young as two can use second-order correlations to guide causal inference in more complex contexts and that this ability may emerge from more general processes of the mind based on representational overlap.
Deon T. Benton (Sat,) studied this question.