Reply to Zhang: Causal identification, social networks, and biological aging

Zhang (1) raises several methodological concerns about our study (2), most of which were raised and addressed during peer review. We welcome the opportunity to clarify our contribution and discuss Zhang’s larger methodological critique of observational studies.

At the core of Zhang’s letter lies a central question: Can observational studies of social relationships contribute to causal understanding when strong identification is difficult to achieve? In our view, the answer is “yes, with clear caveats.” When true causal inference is infeasible, the relevant task is to determine whether an observed association survives the most plausible alternative explanations. We believe we have demonstrated that it does.

On reverse causality, while true that our longitudinal follow-up relies on self-reported health rather than epigenetic clocks, temporal ordering remains informative. Recall that adjusting for the presurvey Charlson comorbidity index and the lifetime multimorbidity index did not attenuate the association (Table 4). On confounding (especially neuroticism), we have shown that our models controlling for affective orientation toward others did not affect estimates (Table 4). The same result is obtained when we address Zhang’s specific concerns on “fatigue, chronic pain, or cognitive slowing” as potential confounders. We assessed robustness to unobserved confounding using konfound analysis (SI Appendix, Table S14) (3), serving a purpose similar to E-values (4), while benchmarking required confounding against observed covariates. Regarding epigenetic clock limitations (5), we randomized samples across plates and adjusted for batch effects and estimates of cell composition in our regression models. Note that cross-tissue validation suggests that relative rankings remain stable in saliva (6).

Zhang points to quasiexperimental approaches (7), but they identify narrow causal effects under specific forms of exogenous variation. Instrumental-variable estimates based on spousal death identify a local average treatment effect for widowhood, not for negative ties across heterogeneous network roles. Roommate randomization cannot speak to kin ties—where our strongest results appear—because family relationships are never randomly assigned. For such social ties, causal understanding is more likely to emerge from convergent evidence than from any single decisive design.

Challenges regarding causal inference are not unique to our study, and recall the debate about the social contagion of obesity (8). Cohen-Cole and Fletcher (9) showed that similar methods could detect apparent “contagion” in height. We take that critique seriously, which is why our paper includes height as a negative control, showing no significant association. The broader lesson from that debate is not that observational network research is invalid but that causal claims must be evaluated against both alternative explanations and plausible mechanisms. Our paper documents converging associations across epigenetic clocks, inflammatory biomarkers, anthropometric measures, and psychiatric symptoms, and interprets them in light of theoretically informed and empirically established mechanisms linking chronic social stress to HPA-axis dysregulation and allostatic load (10, 11). These mechanisms do not depend on perfect exogenous variation; they reflect stress that is real, repeated, and physiologically consequential.

Social ties resist clean identification because they are embedded, obligatory, and persistent. These features are not nuisances to be instrumented away; they are precisely why such relationships matter for health.