Can longitudinal generalized estimating equation models distinguish network influence and homophily? An agent-based modeling approach to measurement characteristics

Authored by Meghan Hutchins, Kori Sauser Zachrison, Theodore J Iwashyna, Achamyeleh Gebremariam, Joyce M Lee

Date Published: 2016

DOI: 10.1186/s12874-016-0274

Sponsors: United States National Institutes of Health (NIH) Center for Social Epidemiology and Population Health

Platforms: Java

Model Documentation: Other Narrative

Model Code URLs: http://dx.doi.org/10.5061/dryad.v3s0k

Abstract

Background: Connected individuals (or nodes) in a network are more likely to be similar than two randomly selected nodes due to homophily and/or network influence. Distinguishing between these two influences is an important goal in network analysis, and generalized estimating equation (GEE) analyses of longitudinal dyadic network data are an attractive approach. It is not known to what extent such regressions can accurately extract underlying data generating processes. Therefore our primary objective is to determine to what extent, and under what conditions, does the GEE-approach recreate the actual dynamics in an agent-based model. Methods: We generated simulated cohorts with pre-specified network characteristics and attachments in both static and dynamic networks, and we varied the presence of homophily and network influence. We then used statistical regression and examined the GEE model performance in each cohort to determine whether the model was able to detect the presence of homophily and network influence. Results: In cohorts with both static and dynamic networks, we find that the GEE models have excellent sensitivity and reasonable specificity for determining the presence or absence of network influence, but little ability to distinguish whether or not homophily is present. Conclusions: The GEE models are a valuable tool to examine for the presence of network influence in longitudinal data, but are quite limited with respect to homophily.

Tags

Obesity Contagion dynamic networks Large social network Instrumental variables