Causal Relationships between Social Networks and Health: A Comparison of Three Modeling Strategies
Emily Ruppel, Stephanie Child, Claude Fischer, Marian BotchwayWhile prior research documents strong associations between social network characteristics and health outcomes, untangling causation and correlation has been a research priority for decades in the field of networks and health. Recent advances in the collection and analysis of panel data have brought us closer to answering these questions. This paper uses three waves of data from the UC Berkeley Social Networks Study (UCNets) to investigate connections between nine network variables and two global health outcomes, psychological well-being and self-rated health. We address the challenges of causal inference by explicitly comparing three modeling strategies: ordinary least squares (OLS) regression, regression with lagged dependent variables (LDVs), and hybrid fixed and random effects models. Results suggest that both OLS and LDV models are at risk of overestimating the causal effects of network characteristics on health. Hybrid models make possible the decomposition of between-individual effects and within-individual effects and show that associations between networks and health primarily operate between individuals, as opposed to changes to networks causing within-individual changes in health during the study period. These findings have methodological and substantive implications: First, they demonstrate the importance of panel data to scholarship on networks and health and illustrate hybrid models as an underutilized technique for untangling causation and correlation in this field. Second, they suggest (with some caveats) that networks’ causal effects on health may be more limited than previously thought, meaning that future research should strive to identify social, cultural, or individual factors that influence both networks and health, producing the between-individual associations established in past research and replicated in the current study.