Paper Summary
Share...

Direct link:

A Hierarchical Mixed Membership Stochastic Blockmodel for Mediation

Sun, April 15, 2:45 to 4:15pm, Westin New York at Times Square, Floor: Ninth Floor, Palace Room

Abstract

In the educational and social sciences, interventions sometimes aim to change the ways in which individuals interact as means to improve some other outcome. For example, educational professional development programs often focus on the ways in teachers communicate or consult with each other to ultimately improve teaching quality. In these studies, the social network within each school provides insight into the mechanisms that affect individual outcomes and acts as a power mediator between the intervention and outcome.

Consider a professional development program whose aim is to bring teachers together to communicate about instruction. Assuming the program is effective, the communication network of these teachers is now different than it was before. The social network therefore is the first observable outcome of the intervention. Of course, the ultimate goal is to improve instruction, so the social network actually operates as a mediator. Now that teachers have created these relationships (and/or changed existing relationships), these new networks can create a path through which teacher quality improves.

Sweet et al. (2013) and Sweet et al. (2014) introduced statistical network models for network-level interventions, building on a small literature on multilevel social network models. In their models, they introduce models for samples that accommodate treatment effects. These models naturally lend themselves to education research where networks of teachers (or students) exist within each school.

We introduce a Bayesian framework for modeling social networks as mediators; we combine statistical network models for network-level interventions with Bayesian mediation models (Yuan and MacKinnon, 2009). Instead of a mediating variable, we develop models where the social network is the mediator. As a proof of concept, we introduce a mediation model where the mediator is a hierarchical mixed membership stochastic blockmodel. Thus, the intervention affects the shape of the networks and the shape of the networks affects a node-level variable (e.g. teacher instruction).

We use this model to show that network structure mediates the effect of coaches on beliefs about mathematics instruction.

Author