Network threats to causal inference: Variations in network position by participation in randomized controlled trials.

Researchers and practitioners rely on randomized controlled trials (RCTs) to make causal inferences. However, most people who participate in RCTs are part of multiple, overlapping social networks that shape their behaviors and attitudes. As a result, variations in trial participants' and non-participants' network positions may impede the generalizability of a RCT's conclusions. The current project evaluates the extent and impact of these variations by comparing the network positions of RCT participants to RCT non-participants. As an informative case study, we considered a workplace-based RCT at a large hospital where a subset of employees was randomized to a healthy eating intervention or control group from 2016-2019. We constructed longitudinal social networks from data about employees' cafeteria purchases and applied stochastic actor-oriented models (SAOMs) to determine whether RCT participants and non-participants occupied significantly different structural positions. Then, we performed a series of computational knockout experiments to assess whether the elimination of specific network-related phenomena impacted estimates of the intervention's effect. Results suggest that RCT participants made cafeteria co-purchases with more of their colleagues than non-participants did. These differences downwardly biased estimates of the intervention's impact, both with respect to the trial's efficacy among participants and its expected effectiveness in the larger population of employees.