Testing Models of Social Learning on Networks: Evidence from a Lab Experiment in the Field

Agents often use noisy signals from their neighbors to update their beliefs about a state of the world. The effectiveness of social learning relies on the details of how agents aggregate information from others. There are two prominent models of information aggregation in networks: (1) Bayesian learning, where agents use Bayes’ rule to assess the state of the world and (2) DeGroot learning, where agents instead consider a weighted average of their neighbors’ previous period opinions or actions. Agents who engage in DeGroot learning often double-count information and may not converge in the long run. We conduct a lab experiment in the field with 665 subjects across 19 villages in Karnataka, India, designed to structurally test which model best describes social learning. Seven subjects were placed into networks with common knowledge of the network structure, designed to maximize our experimental power to distinguish between these two models of learning. Subjects attempted to learn the underlying (binary) state of the world, having received independent identically distributed signals in the first period. Thereafter, in each period, subjects made guesses about the state of the world, and these guesses were transmitted to their neighbors at the beginning of the following round. We structurally estimate a model of Bayesian learning, relaxing common knowledge of Bayesian rationality by allowing agents to have incomplete information about other players’ types (Bayesian or DeGroot updating types). Our estimates show that, despite the flexibility in modeling learning in these networks, agents are robustly best described by DeGroot-learning models wherein they take a simple majority of previous guesses in their neighborhood.