Job Market Paper

Abstract. We study network intervention when agents have incomplete information about the network structure and intervention is observed locally. Unlike complete information settings where the optimal intervention is determined by structural centrality, under incomplete information and limited observability, the optimal intervention is determined by the interaction of network position and information propagation. An observer of this removal must reason about whether their neighbors have seen the removal as well and this leads to an information cascade of inferences about co-observation. We characterize this information cascade through belief weighted walks that traces chains of co-observation. Under symmetric priors we show that, the information cascade reduces to a degree based kernel and when this kernel is homogeneous, the optimal targeting reduces to two local statistics. The degree of the removed agent and the co-observation potential of their neighbors. Global targeting consequences thus can be inferred entirely from local information under symmetric priors.