Are Paid Subscriptions on Music Social Networks Contagious ? A Randomized Field Experiment

Demonstrating compelling causal evidence of the existence and strength of peer to peer influence in online social networks has become the holy grail of the modern research in online social networks. While it has been consistently demonstrated that user characteristics and behavior tend to cluster inside social networks both in space and in time, there are several different mechanisms that can cause this observed clustering. Among the most frequently cited explanations for this effect are two rival mechanisms: peer influence and homophily. While both peer influence and homophily mechanisms can each lead to very similar observational data, the policy implications of each of these mechanisms are significantly different: under peer influence an effective policy might be to identify “influential” people and induce the desired behavior among them, while under homophily mechanism this policy may have no effect and a different type of action is needed. Considering the massiveness and continuing growth of online social networks, it is therefore of critical importance both for research and for businesses to reliably identify presence of each of these mechanisms in the general population of online social networks and quantify the strength of them. Traditionally, the econometric identification of peer influence from purely observational data has proven to be a hard challenge. In this paper, we present a novel randomized experiment that tests the existence of causal peer influence in the general population of a particular large-scale online social network. We present our findings starting with insights learned from observational data, followed by a quasi-experiment based on observational data and concluding with the randomized field trial. Both quasi-experiment and randomized experiment demonstrated that new adoptions were significantly higher in the treatment group vs. control group. The quasi-experiment allowed us to determine the adequate sample size for our randomized trial as well as provided an interesting baseline to compare our results against. Both simple t-test and logistic regression indicate that user’s adoption of a product causes her online friends to pay for it and adopt it as well. Our point estimates show that, for a median social network user, the odds of adopting the paid subscription increase by 116% due to peer influence when her friend adopts it. In addition, we find that peer influence is significantly stronger for users with smaller number of friends as compared to the ones with large number of friends. Finally, we find that the quasi-experiment tends to produce the results similar to randomized trial, somewhat over-estimating the effect on users with larger number of friends and under-estimating it for the users with smaller number of friends, thus providing the first insights about the nature of bias in estimating peer-effects by the models with self-selected populations.