Causal Relationships between Personal Networks and Health: A Comparison of Three Modeling Strategies

Abstract

Prior research documents associations between personal network characteristics and health, but establishing causation has been a long-standing research priority. To evaluate approaches to causal inference in egocentric network data, this article uses three waves from the University of California Berkeley Social Networks Study (N = 1,159) to investigate connections between nine network variables and two global health outcomes. We compare three modeling strategies: cross-sectional ordinary least squares regression, regression with lagged dependent variables (LDVs), and hybrid fixed and random effects models. Results suggest that cross-sectional and LDV models may overestimate the causal effects of networks on health because hybrid models show that network–health associations operate primarily between individuals, as opposed to network changes causing within-individual changes in health. These findings demonstrate uses of panel data that may advance scholarship on networks and health and suggest that causal effects of network support on health may be more limited than previously thought.

Distinct aspects of human connection associated with subjective well-being

“…The KHB method suggests that evaluations of networks exercise independent effects on SWB; they do not purely mediate the relationship between network descriptions and SWB. Other researchers, too, have found few, small, or ambiguous causal effects of “objective” network traits on SWB in panel studies ( Santini et al, 2020 ; Bui, 2020 ; Ruppel et al, 2022 ; see Gariépy et al, 2016 ). Yet even the effects of subjective evaluations may have been produced in part by reciprocal effects ( Supplement , Appendix E ).…”

“…Meta-analyses criticize the varied quality of extant experiments, which have produced a range of contradictory results ( Hogan et al, 2002 ; Hunter et al, 2019 ). The limited success of these clinical interventions may underline problems in establishing causality using observational data (e.g., Schaefer et al, 2011 ; Moore et al, 2018 ; Ruppel et al, 2022 ), though the failures of these interventions may also be related to their artificial character.…”

“…(2) Untested third factors threaten causal inference. Panel data help eliminate these confounders but have so far yielded modest estimates of causality (e.g., Hajek et al, 2016 ; Yang and Park, 2019 ; Ruppel et al, 2022 ). Confounding variables which change between waves, such as genetic or childhood factors that unfold in later life or coincident events between waves, can muddle panel analyses.…”

“…Its data yield network correlations with SWB independent of egos’ evaluations; its three waves of data provide estimates of both between- and withinperson effects; and it covers a broader age range than do most studies—young adults and late-middle-aged adults. Previous studies have also used UCNets (e.g., Child and Lawton, 2019 , 2020 ; Offer, 2020 ; Bilecen and Vacca, 2021 ; K. Lee, 2021 ; Ruppel et al, 2022 ). Our work replicates these analyses to a degree, but expands on them in two ways: First, we use all three waves of the UCNets data, which few studies have done thus far, and which allows us to undertake more sophisticated causal analyses.…”