Power and sample size calculations for evaluating spillover effects in networks with non-randomized interventions

Determining the appropriate sample size for desired statistical power is crucial for obtaining reliable research outcomes. While methods exist for multiple types of studies, the method for evaluating power of estimating spillover effects in sociometric network-based studies with non randomized interventions remain inadequately explored. We conducted a simulation study to assess how the design parameters (i.e., number of components, number of nodes, node degree, transitivity, and effect size) affects the statistical power for estimating spillover effects in non randomized, network-based studies. Both simulated networks and a real-world network from Transmission Reduction Intervention Project (TRIP) were used in this study. Our simulation results suggests that: (1) power increases with more nodes or a larger effect size, but not necessarily with more components when the number of nodes is fixed; (2) A higher node degree or greater transitivity results in reduced power; (3) Highly unbalanced networks (e.g., most of the nodes are in one component) can drastically reduce power. Furthermore, the power calculated using a closed-form expression developed in this work also shows that power remained the same or even decreases slightly with more components when the number of nodes are fixed, aligning with the simulation findings. All the results were specific to the inverse probability weighting estimator we employed in this study and assumptions it required. An alternative estimator or interference assumption may lead to different results.