Design of egocentric network-based studies to estimate causal effects under interference

TL;DR

This paper develops methods for designing egocentric network-based studies to accurately estimate direct, spillover, and overall causal effects of interventions in interconnected populations, using potential outcomes and regression models.

Contribution

It introduces a novel study design framework and sample size formulas for estimating multiple causal effects in network settings with interference.

Findings

Derived sample size formulas considering intra-class correlation and treatment probability

Clarified assumptions for causal effect identification in egocentric networks

Provided regression approaches for causal inference under interference

Abstract

Many public health interventions are conducted in settings where individuals are connected to one another and the intervention assigned to randomly selected individuals may spill over to other individuals they are connected to. In these spillover settings, the effects of such interventions can be quantified in several ways. The average individual effect measures the intervention effect among those directly treated, while the spillover effect measures the effect among those connected to those directly treated. In addition, the overall effect measures the average intervention effect across the study population, over those directly treated along with those to whom the intervention spills over but who are not directly treated. Here, we develop methods for study design with the aim of estimating individual, spillover, and overall effects. In particular, we consider an egocentric network-based randomized design in which a set of index participants is recruited from the population and randomly assigned to treatment, while data are also collected from their untreated network members. We use the potential outcomes framework to define two clustered regression modeling approaches and clarify the underlying assumptions required to identify and estimate causal effects. We then develop sample size formulas for detecting individual, spillover, and overall effects. We investigate the roles of the intra-class correlation coefficient and the probability of treatment allocation on the required number of egocentric networks with a fixed number of network members for each egocentric network and vice-versa.