We survey the literature on the economic consequences of the structure of social networks. We develop a taxonomy of “macro” and “micro” characteristics of social-interaction networks and discuss both the theoretical and empirical findings concerning the role of those characteristics in determining learning, diffusion, decisions, and resulting behaviors. We also discuss the challenges of accounting for the endogeneity of networks in assessing the relationship between the patterns of interactions and behaviors.

We present a dynamic model of network formation where nodes find other nodes with whom to form links in two ways: some are found uniformly at random, while others are found by searching locally through the current structure of the network (e.g., meeting friends of friends). This combination of meeting processes results in a spectrum of features exhibited by large social networks, including the presence of more high- and low-degree nodes than when links are formed independently at random, having low distances between nodes in the network, and having high clustering of links on a local level. We fit the model to data from six networks and impute the relative ratio of random to network-based meetings in link formation, which turns out to vary dramatically across applications. We show that as the random/network-based meeting ratio varies, the resulting degree distributions can be ordered in the sense of stochastic dominance, which allows us to infer how the formation process affects average utility in the network.

We report experimental results from long sequences of decisions in environments that are theoretically prone to severe information cascades. Observed behaviour is much different—information cascades are ephemeral. We study the implications of a theoretical model based on quantal response equilibrium, in which the observed cascade formation/collapse/formation cycles arise as equilibrium phenomena. Consecutive cascades may reverse states, and usually such a reversal is self-correcting: the cascade switches to the correct state. These implications are supported by the data. We extend the model to allow for base rate neglect and find strong evidence for overweighting of private information. The estimated belief trajectories indicate fast and efficient learning dynamics.