Date of Award
Master of Science in Statistics
Computer Science and Statistics
Determining the appropriate sample size that is needed to achieve a desired level of statistical power is an integral aspect of study design, as it ensures that studies are adequately powered and that their results are reliable and meaningful. Adequate statistical power can also address some practical and ethical considerations. Various methodologies for power and sample size calculations have been developed for distinct types of studies. However, the implications for power when estimating spillover effects in sociometric network-based studies with non-randomized interventions (SNS-NRIs) remain inadequately explored. In this study, we first conducted a series of simulations to evaluate how the statistical power of estimating spillover effect in SNS-NRIs was affected by factors such as network size (i.e., number of components, number of nodes), node degree, transitivity, and effect size. Subsequently, we developed methods for power and sample size calculations specifically tailored for estimating spillover effect in SNS-NRIs. The findings indicated that power typically increased with an increase in the number of nodes or a larger effect size regardless of the number of components, but did not necessarily change with more components when the number of nodes was fixed. In addition, power was observed to decrease with higher node degree or transitivity. One hypothesis for this phenomenon is that high node degree or transitivity may lead to complex dependencies between individuals, which could actually make it harder to discern clear spillover effects. Further research is needed to investigate the underlying causes. Also, the results suggested that a highly unbalanced network, for example, a network that contains more than one component but has most nodes concentrated in one component, significantly impacts statistical power, rendering it near zero irrespective of the effect size. To address the issue related to a highly unbalanced network, collecting data from multiple sites that are geographically isolated may be a potential solution. Furthermore, the power calculated using closed-form expression also showed that power remains the same or even decreased slightly with more components when the number of nodes are fixed, which is consistent to the simulation implications.
Zhang, Ke, "POWER AND SAMPLE SIZE CALCULATIONS FOR EVALUATING SPILLOVER EFFECTS IN NETWORKS WITH NON-RANDOMIZED INTERVENTIONS" (2023). Open Access Master's Theses. Paper 2347.
Available for download on Friday, September 05, 2025