Observational Studies of Peer Effects:

• Activity Number: 494 - Identifying and Addressing Sources of Bias in Causal Inference
• Type: Topic Contributed
• Date/Time: Wednesday, July 31, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Dean Eckles* and Eytan Bakshy
• Companies: MIT and Facebook
• Keywords: causal inference; peer effects; social networks; node embeddings; information diffusion

Abstract

Peer effects, in which the behavior of an individual is affected by the behavior of their peers, are posited by multiple theories in the social sciences. Other processes can also produce behaviors that are correlated in networks and groups, thereby generating debate about the credibility of observational (i.e. nonexperimental) studies of peer effects. Using a large experiment as a "gold standard", we evaluate methods for causal inference for peer effects with observational data. First, we consider potentially high-dimensional adjustment for individuals' prior behavior. Naive observational estimators overstate peer effects by 320%, but high-dimensional models adjusting for over 3,700 past behaviors provide additional bias reduction, such that the full model reduces bias by over 97%. Second, we consider methods that use the observed social network for adjustment, such as through estimating latent positions of nodes in the network.