Social Contagion Theory: Examining Dynamic Social Networks and Human Behavior

Here, we review the research we have done on social contagion. We describe the methods we have employed (and the assumptions they have entailed) in order to examine several datasets with complementary strengths and weaknesses, including the Framingham Heart Study, the National Longitudinal Study of Adolescent Health, and other observational and experimental datasets that we and others have collected. We describe the regularities that led us to propose that human social networks may exhibit a “three degrees of influence” property, and we review statistical approaches we have used to characterize inter-personal influence with respect to phenomena as diverse as obesity, smoking, cooperation, and happiness. We do not claim that this work is the final word, but we do believe that it provides some novel, informative, and stimulating evidence regarding social contagion in longitudinally followed networks. Along with other scholars, we are working to develop new methods for identifying causal effects using social network data, and we believe that this area is ripe for statistical development as current methods have known and often unavoidable limitations.

💡 Research Summary

The paper provides a comprehensive review of the authors’ research program on social contagion, focusing on how behaviors, affective states, and health outcomes spread through dynamic social networks. Central to the work is the construction of the “FHS‑Net,” a longitudinal network dataset derived from the Framingham Heart Study (FHS). By digitizing handwritten tracking sheets, the authors reconstructed 12,067 individuals and 53,228 ties (spouses, siblings, friends, coworkers, neighbors) over 32 years, with repeated measurements of body‑mass index, smoking, alcohol use, happiness, depression, and many other variables across seven exam waves. Complementary datasets—including the AddHealth school‑based network, online social media extracts, and several experimental studies—are also described, each with distinct strengths and weaknesses.

The authors first assess whether traits cluster in the observed networks more than expected by chance. They employ a permutation test that preserves the exact network topology and overall trait prevalence while randomly shuffling trait assignments across nodes. By comparing the observed risk ratios at each geodesic distance (1‑step, 2‑step, 3‑step) to the distribution generated from thousands of permuted networks, they demonstrate statistically significant clustering up to three degrees of separation for a wide range of outcomes (obesity, smoking, happiness, loneliness, etc.). This empirical regularity underpins the “three degrees of influence” hypothesis, suggesting that a person’s behavior can be associated with that of friends, friends‑of‑friends, and even friends‑of‑friends‑of‑friends.

To move beyond descriptive clustering, the authors fit longitudinal dyadic regression models. These models treat each ego‑alter pair as an observation and include lagged alter outcomes as predictors of ego outcomes, thereby estimating peer effects while controlling for ego’s prior state and covariates. The approach accommodates both binary outcomes (e.g., obesity status) and continuous measures (e.g., BMI). Critical assumptions are highlighted: (1) network ties are exogenous, (2) unobserved confounders are either absent or adequately controlled, and (3) the timing of exposure and outcome is correctly specified. The authors acknowledge that these assumptions are often violated in observational network data and discuss potential biases such as homophily, shared environment, and simultaneous adoption.

To address identification challenges, the paper introduces several strategies. First, the directionality of ties (e.g., “named friend” vs. “namer”) is exploited; asymmetric ties can help disentangle influence from selection because only the direction in which influence can flow is considered. Second, geographic information (home addresses, workplace locations) is used to separate social influence from common exposure to local environmental factors. Third, the authors discuss the value of experimental and quasi‑experimental designs—such as randomized assignment of network positions or induced exposure—to obtain more credible causal estimates. They cite their own experiments on altruism, political mobilization, and cooperation, where network structures were deliberately manipulated, yielding stronger causal inference.

Throughout, the authors are transparent about methodological limitations. They discuss missing data (unobserved alters, incomplete tie ascertainment), sampling biases (the FHS cohort is predominantly white, relatively affluent, and highly educated), and the difficulty of estimating standard errors in network‑dependent data. They argue that while current methods (permutation tests, longitudinal regression, directional tie analysis) represent the best available tools, they are not “final” and call for continued statistical innovation. Suggested avenues include Bayesian hierarchical models for network data, structural equation modeling that incorporates latent homophily, and machine‑learning approaches (e.g., graph neural networks) that can capture complex, non‑linear diffusion processes.

In conclusion, the paper synthesizes a decade of empirical work showing that a variety of health‑related and behavioral traits exhibit measurable clustering and apparent peer effects across up to three degrees of separation. It balances enthusiasm for the robustness of these findings with a sober appraisal of the inferential challenges inherent in observational network research. By outlining both the substantive evidence for social contagion and the methodological frontiers that remain, the authors provide a roadmap for future research aimed at rigorously quantifying and eventually leveraging network influence for public‑health interventions.