Statistical Inference from Power Law Distributed Web-Based Social Interactions

Abstract:     
Purpose – to propose a method for statistical inference on data from power law distributions in order to explain behavior and social phenomena associated with web-based social spaces such as discussion forums, question-and-answer sites, web 2.0 applications and the like.

Design/methodology/approach – The paper starts by highlighting the importance of explaining behavior in social networks. Next, the power law nature of social interactions is described and a hypothetical example is used to explain why analyzing sub-sets of data might misrepresent the relationship between variables having power law distributions. Analysis requires the use of the complete distribution. We propose logarithmic transformation prior to correlation and regression analysis and show why it works using the hypothetical example and field data retrieved from Microsoft's Netscan project.

Findings – The hypothetical example emphasizes the importance of analyzing complete datasets harvested from social spaces. The Netscan example shows the importance of the logarithmic transformation for enabling the development of a predictive regression model based on the power law distributed data. Specifically, we show that the number of new and returning participants are the main predictors of discussion forum activity.

Originality/value – This paper offers a useful analysis tool for anyone interested in social aspects of the Internet as well as corporate intra-net systems, knowledge management systems or other systems which support social interaction such as cellular phones and mobile devices. It also explains how to avoid errors by paying attention to assumptions and range restriction issues.

Keywords: information markets, power law distribution, statistical analysis, motivation for participation, Netscan data

JEL Classifications: C12, C80

Accepted Paper Series