The Effect of Alter Ego Accounts on A/B Tests in Social Networks
Abstract: Social network users often maintain multiple active accounts, sometimes referred to as alter egos. Examples of alter egos include personal and professional accounts or named and anonymous accounts. If alter egos are common on a platform, they can affect the results of A/B testing because a user's alter egos can influence each other. For a single user, one account may be assigned treatment, while another is assigned control. Alter-ego bias is relevant when the treatment affects the individual user rather than the account. Through experimentation and theoretical analysis, we examine the worst and expected case bias for different numbers of alter egos and for a variety of network structures and peer effect strengths. We show that alter egos moderately bias the results of simulated A/B tests on several network structures, including a real-world Facebook subgraph and several types of synthetic networks: small world networks, forest fire networks, stochastic block models, and a worst-case structure. We also show that bias increases with the number of alter egos and that different network structures have different upper bounds on bias.
Loading