Abstract
Most inference using social network models assumes that the presence or absence of all relations is known. This is rarely the case. Most social network analysis ignores the problem of missing data by including only actors with complete observations.
In this paper we use a statistical model for the underlying social network to demonstrate that the computationally parsimonious complete case approach can lead to different conclusions from an approach utilizing all observations. We also show that the overall fit to the data is improved by extending the model to represent differences between respondents and non-respondents.
The ideas are motivated and illustrated by an analysis of a friendship network from the National Longitudinal Study of Adolescent Health.