Network Effects in Contagion Processes: Identification and Control

In this paper, we study the problem of identifying network effects in contagion processes and present an application to the propagation of influenza in the United States. In particular, using data on the evolution of infections over time, the travel intensity between states as well as environmental conditions we first provide a framework to identify the true network effect of traveling between states. Any identification strategy in this context needs to handle the following challenges: the reflection problem and the time correlation problem. The reflection problem arises from the observation that when sampling from the contagion process is frequent (in our case, weekly), the (potential) endogenous network effect cannot be discriminated from the correlation effect (such as that due to similar environmental conditions). The time-correlation effect stems from the observation that contagion processes are naturally characterized by correlation across different lags. We propose an instrumental variable approach, based on a spatiotemporally lagged versions of the observed data, and we show that our approach effectively tackles the aforementioned issues both theoretically and through a series of robustness checks. Finally, we use our estimates to propose and evaluate the performance of intervention and control policies, illustrating the benefits of network-based interventions.