An Equilibrium Model of Experimentation on Networks

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

This project will investigate how the balance between discovery and diffusion depends on the nature of connections between people. Do more connections lead to faster technological progress? Is society better off in a loosely connected network or with many tight clusters? This project has two important features that differentiate it from the wider literature. First, by studying both diffusion and discovery within a single model, the project analyzes the interconnection between the two forces and how this depends on the social network. Second, the project considers forward-looking agents, whose incentives for original discovery are crowded out by the possibility of learning from others. The resulting model can be used to understand how policy changes (e.g., subsidizing innovation) affect society’s technological progress. The project also speaks to a growing empirical literature that studies how people learn about innovations from their neighbors (e.g. new production techniques or new consumer products). In their classic paper, Bala and Goyal (1998) study agents who learn from their neighbors in a network. To make analytical progress, they restrict attention to myopic, non-Bayesian agents, shortcutting strategic considerations and allowing them to solve the model as a sequence of static decision problems. In contrast, this project studies agents who are forward-looking and Bayesian, so social learning (both past and future) crowds out private experimentation. The key simplifying assumption is that agents learn via a perfect good news process. While each agent faces a rich strategy space, her “social learning curve” is described by a simple function of time, and her best response reduces to choosing a single number: the total amount of individual experimentation, as captured by a cutoff time. The model thus recovers the tractability of the reduced-form models of experimentation in a model of Bayesian learning and uses this to provide a clean characterization of initial experimentation and subsequent contagion. The project investigates learning and welfare as a function of the density and structure of the network. It studies two measures of network density: (i) the size of the core in a core-periphery network and (ii) the degree of random regular networks generated by the configuration model. For either measure, preliminary results suggest that aggregate (asymptotic) information decreases in network density, whereas welfare is hump-shaped in network density, with intermediate networks striking a balance between motivating individual discoveries and their social diffusion. The project will also investigate network structure, including models where links are one-directional (e.g., Twitter), bi-directional (e.g., LinkedIn) and clustered (e.g., Facebook). It will study the effect of such network architecture on learning and welfare in the context of tree networks. Trees both approximate large random networks and are highly tractable, allowing the researchers to characterize the social diffusion of information by simple ordinary differential equations. Collectively, this project will paint a clear picture about learning dynamics, information aggregation, and welfare in networks of forward-looking agents. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.