An exploration of the Facebook social networks of smokers and non-smokers

BACKGROUND Social networks influence health behavior, including tobacco use and cessation. To date, little is known about whether and how the networks of online smokers and non-smokers may differ, or the potential implications of such differences with regards to intervention efforts. Understanding how social networks vary by smoking status could inform public health efforts to accelerate cessation or slow the adoption of tobacco use. OBJECTIVES These secondary analyses explore the structure of ego networks of both smokers and non-smokers collected as part of a randomized control trial conducted within Facebook. METHODS During the trial, a total of 14,010 individuals installed a Facebook smoking cessation app: 9,042 smokers who were randomized in the trial, an additional 2,881 smokers who did not meet full eligibility criteria, and 2,087 non-smokers. The ego network for all individuals was constructed out to second-degree connections. Four kinds of networks were constructed: friendship, family, photo, and group networks. From these networks we measured edges, isolates, density, mean betweenness, transitivity, and mean closeness. We also measured diameter, clustering, and modularity without ego and isolates. Logistic regressions were performed with smoking status as the response and network metrics as the primary independent variables and demographics and Facebook utilization metrics as covariates. RESULTS The four networks had different characteristics, indicated by different multicollinearity issues and by logistic regression output. Among Friendship networks, the odds of smoking were higher in networks with lower betweenness (p = 0.00), lower transitivity (p = 0.00), and larger diameter (p = 0.00). Among Family networks, the odds of smoking were higher in networks with more vertices (p = .01), less transitivity (p = .04), and fewer isolates (p = .01). Among Photo networks, none of the network metrics were predictive of smoking status. Among Group networks, the odds of smoking were higher when diameter was smaller (p = .04). Together, these findings suggested that compared to non-smokers, smokers in this sample had less connected, more dispersed Facebook Friendship networks; larger but more fractured Family networks with fewer isolates; more compact Group networks; and Photo networks that were similar in network structure to those of non-smokers. CONCLUSIONS This study illustrates the importance of examining structural differences in online social networks as a critical component for network-based interventions and lays the foundation for future research that examines the ways that social networks differ based on individual health behavior. Interventions that seek to target the behavior of individuals in the context of their social environment would be well served to understand social network structures of participants.