Superpopulation Inference in Social Network Analysis

Abstract

Social network data, which record pairwise interactions between individuals, have generated significant interest in recent years as a tool for investigating complex social processes. Many social scientists approach large-scale social network data with the hope that this data source can help them develop theory that is widely applicable across different social situations. This sort of generalization requires the notion of a superpopulation that simultaneously governs outcomes in different contexts, so that inferences from one context can be transferred to another. Classical approaches to superpopulation inference depend on a well-specified model, and analysis of these methods is built on an analogy between large samples and the superpopulation. Unfortunately, social network data have a number of idiosyncrasies that undermine this approach. In this thesis, we focus on complications introduced by the sparsity of social networks. Sparsity is a difficult property to model accurately, and when a social network process is sparse, samples of different size are inhomogeneous, breaking the classical equivalences between large samples and superpopulations. In this thesis, we introduce theory and methods for drawing well-defined superpopulation inferences from sparse social network data. In particular, we develop a general framework for understanding whether an estimation procedure supports the appropriate superpopulation generalization, and in the context of social network analysis, develop methodology for drawing generalizable predictive and causal inferences.