Sparsity Misspecification and Robust Covariate Effect Estimation for Sparse Social Networks

• Activity Number: 61
• Type: Topic Contributed
• Date/Time: Sunday, August 3, 2014 : 4:00 PM to 5:50 PM
• Sponsor: IMS
• Author(s): Alexander D'Amour*+ and Edoardo M. Airoldi
• Companies: Harvard and Harvard
• Keywords: social network analysis ; stochastic processes ; partial likelihood ; truncation

Abstract

In social interaction datasets, we often wish to estimate how covariates influence the observed patterns of interactions between actors. For example, we may wish to explain the dynamics of collaborations between inventors as a function of their corporate affiliations, areas of expertise, or geographic location. A natural approach is to treat each pair of actors as a conditionally independent replicate, and to apply estimation and prediction procedures that are popular with more traditional non-network datasets, for example, Poisson or Cox proportional hazards regression.

Unfortunately, this approach is inadequate for describing the known sparsity of real social interaction data. We show that this misspecification degrades the usefulness and interpretability of traditional estimators when the object of the analysis is to compare or share information across network samples of differing size. We then propose a modeling framework that defines interpretable covariate effects irrespective of the sparsity of the interaction process. Finally, we derive a procedure that robustly estimates these effects by trading a small loss of efficiency for valid estimates and large computational gains.