A Representative Sampling Method for Peer Encouragement Designs in Network Experiments

Network experiments using peer encouragement designs are widely used to estimate the direct and indirect effects of social marketing campaigns. To satisfy the stable unit treatment value assumption, contaminated nodes are typically excluded from analysis. This exclusion approach creates two problems: underrepresentation, where postexclusion samples are not representative of the population, and undersupply, where postexclusion samples are too small for reliable inference. We propose a representative ego network sampling method that directly addresses these challenges. The method embeds a distance constraint within a Metropolis-Hastings algorithm to generate contamination-free ego network samples whose joint distribution of key network attributes converges to that of the population. Using both simulated and real-world networks, we show that the proposed method consistently generates larger and more representative samples than the standard exclusion approach. These samples produce more accurate and reliable estimates of average and heterogeneous treatment effects and improve statistical inference. The advantages are especially pronounced when experiments require large samples, population networks are small or dense, or higher-order contamination must be controlled. Practically, the method can be adapted for various applications, providing a valuable tool for firms designing and evaluating social marketing campaigns.

History: Olivier Toubia served as the senior editor for this article.

Supplemental Material: The online appendices and data files are available at https://doi.org/10.1287/mksc.2022.0409.