Uses and Limitations of GEEs and GLMs for Social Network Data

• Activity Number: 70
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Social Statistics Section
• Author(s): Elizabeth Ogburn*+
• Companies: Harvard University
• Keywords: social networks ; causal inference

Abstract

It is tempting to use familiar statistical machinery like generalized linear models, generalized estimating equations, and spatial autoregressive models to estimate causal effects using social network data. However, as some researchers have noted, in many network settings these models are inconsistent or uninterpretable. I describe three distinct sources of dependence in social network data, explain why network dependence is generally incompatible with the assumptions of the standard models listed above, and give conditions under which the assumptions of generalized linear models and generalized estimating equations will be met even in the presence of network dependence. I explore the limitations of these methods even when the necessary assumptions are met and describe alternatives that take account of network dependence directly.