The 'Unfriending' Problem: The Consequences of Homophily in Friendship Retention for Causal Estimates of Social Influence

An increasing number of scholars are using longitudinal social network data to try to obtain estimates of peer or social influence effects. These data may provide additional statistical leverage, but they can introduce new inferential problems. In particular, while the confounding effects of homophily in friendship formation are widely appreciated, homophily in friendship retention may also confound causal estimates of social influence in longitudinal network data. We provide evidence for this claim in a Monte Carlo analysis of the statistical model used by Christakis, Fowler, and their colleagues in numerous articles estimating “contagion” effects in social networks. Our results indicate that homophily in friendship retention induces significant upward bias and decreased coverage levels in the Christakis and Fowler model if there is non-negligible friendship attrition over time.

💡 Research Summary

The paper addresses a subtle but critical source of bias in longitudinal social network analyses that aim to estimate peer or social influence—namely, homophily in the retention of friendships, often referred to as the “unfriending problem.” While the literature has long recognized that homophily in friendship formation can confound causal inference, the authors argue that the same mechanism operating during the maintenance (or dissolution) of ties can also generate spurious contagion estimates. To demonstrate this, they conduct an extensive Monte‑Carlo simulation study modeled on the statistical framework popularized by Christakis and Fowler (CF).

The CF model treats two waves of network data as a panel: it conditions on dyads that remain friends between wave t and wave t + 1, then regresses the change in an individual’s outcome on the contemporaneous change in the friend’s outcome, controlling for baseline characteristics. The key assumption is that, conditional on friendship persistence, the friend’s change is exogenous to the ego’s change. The authors show that this assumption is violated when the probability of friendship persistence itself depends on similarity in outcomes—a form of homophily in tie retention.

In the simulation, synthetic networks are generated with 1,000 nodes and a set of observable traits (e.g., age, health status). Friendship formation follows a logistic function of trait similarity, reproducing realistic homophily levels. At each time step, an outcome variable (e.g., body‑mass index) evolves according to a pre‑specified structural equation that may include a true peer influence parameter β (set to 0 in half of the runs). Crucially, after the outcome shock, each dyad faces a retention probability that is a decreasing function of the absolute difference in outcome change; thus, dyads that diverge are more likely to dissolve. The overall attrition rate is varied from 5 % to 30 % to reflect empirical settings.

Applying the CF estimator to each simulated dataset yields two striking patterns. First, when any non‑trivial attrition is present, the estimated β̂ is systematically upward‑biased. For example, with a 20 % attrition rate and moderate homophily in retention, the mean bias reaches +0.15 even when the true β = 0. Second, the nominal 95 % confidence intervals cover the true parameter far less often than advertised—coverage drops below 70 % in many scenarios. This under‑coverage is driven by the selective survival of dyads that are already similar, which inflates the apparent synchrony of outcome changes and mimics contagion. The bias intensifies as the retention homophily parameter grows, confirming that more clustered, highly homophilous networks are especially vulnerable.

The authors discuss the implications for the large body of empirical work that has reported “contagion” effects in health, obesity, smoking, and political behavior using the CF approach. Their findings suggest that many of these reported effects could be partially, if not wholly, artifacts of unmodeled tie‑dissolution processes. They propose several methodological remedies: (1) explicitly model the retention process, for instance by estimating a separate dyadic logistic model for friendship survival and using inverse‑probability weights in the contagion regression; (2) adopt joint modeling frameworks such as structural equation models or latent‑variable dynamic network models that simultaneously capture tie formation, tie dissolution, and outcome evolution; (3) conduct sensitivity analyses that vary the assumed retention‑homophily strength to assess robustness; and (4) where possible, collect higher‑frequency panel data to directly observe the timing of tie loss relative to outcome changes.

In sum, the paper makes a compelling case that homophily in friendship retention is a non‑negligible source of bias for causal estimates of social influence in longitudinal network data. By rigorously demonstrating the magnitude of the bias through simulation, it calls for a re‑examination of prior “contagion” findings and for the adoption of more sophisticated statistical techniques that account for the full life‑cycle of social ties. This contribution deepens our understanding of the methodological challenges inherent in network‑based causal inference and provides a clear roadmap for future research to obtain more credible estimates of peer effects.