Detection of Experimental Effects in Social Network Analysis

Descriptive and inferential statistical techniques exist for the analysis of social networks, but to date the inferential methods have been limited to the comparison of one network to its hypothesized population parameters (analogous to a one-sample t-test), or the comparison of multiple relational structures measured on the same group of actors (analogous to a correlation coefficient). In this paper, we explore techniques for comparing network structures when each network comprised an entirely different set of actors (analogous to a two-sample t-test or a between-subjects analysis of variance). Such methods for between-group comparisons are critical in theory testing, where a researcher varies an experimental factor for the purposes of studying its impact on some dependent variable (e.g. the resulting network structure). This ability to test the significance of manipulated factors (like ANOVA) would provide another important means by which social network analyses would be an aid to researchers. We propose several such statistical methods for comparing network interactions. There are many areas of substantive research in which groups are examined under different operating conditions, and it seems reasonable to study these inter-related actors as players in a network, and test the significance of the factors that had been experimentally manipulated, thereby capitalizing on the strengths of both network analyses and the logic of analysis of variance and experimental design. In this paper, we illustrate our proposed methods on data representing coalitions formed under different experimental conditions.