Social Advertising Effectiveness Across Products: A Large-Scale Field Experiment
 (with Sinan Aral, Jeffrey (Yu) Hu and Erik Brynjolfsson)
Under revision, Marketing Science, Special issue on field experiment

Almost all of the empirical evidence of a lift from social advertising focuses on a single product at a time. As a result, we know little about how social advertising effectiveness varies across product categories or how product characteristics impact social advertising effectiveness. We therefore collaborated with WeChat to conduct a randomized field experiment measuring social ad effectiveness across 71 products in 25 categories among a random sample of more than 37 million users of WeChat Moments Ads. We found some product categories, like food, clothes, and cars, experienced significantly stronger social advertising effectiveness than other categories like financial services and electrical appliances. More generally, we found that status goods, which rely on normative social influence, displayed strong social advertising effectiveness, while social ads for experience goods, which rely on informational social influence, did not perform any better or worse than their theoretical counterpart search goods. The status and expertise of the user displayed in the ad also moderated these effects differently across different products. Understanding the heterogeneous effects of social advertising across products will help marketers differentiate their social advertising strategies and lead researchers to a more general theory of social influence in product adoption.

Rational Herding in Social Advertising: A Large-Scale Randomized Field Experiment 

This study uses data from a large-scale field experiment to investigate how social cues (i.e., friends' likes) impact users' public (i.e., liking) and private (i.e., clicking) responses to social ads. The public responses will become social cues, which are broadcast with ads in social networks and are the source of social influence. The private responses are the main measure for ad engagement. In the experiment, I randomly manipulated the presence and the number of social cues (i.e., friends' likes) shown in ads among 37 million users of WeChat Moments ads. The results demonstrate that, on average, showing the first social cue significantly enhances users' liking and clicking propensity, but showing the additional social cues only increases users' tendency to like but does not affect their tendency to click an ad. Although users will always herd in publicly responding to (i.e., liking) an ad, I find the evidence of rational herding in users' private engagement with (i.e., clicking) an ad. It indicates that users infer the trustworthiness of social cues (i.e., likes) by observing the process of generating (i.e., liking) them. The first like, generated independent of social conformity, always exhibits significantly positive effects on ad clicking. The unpopularity of brands enables the herding momentum in clicking, as users infer the superior trustworthiness of social cues associated with small brands to justify the herd. Social influence in social advertising may fail if users attribute the herd of social cues to external factors, such as social conformity and popularity of brands.

Identifying Subgroups with Enhanced Peer Effects using High-Dimensional Data  
(with Tong Wang and Haojun Wu)

We develop a machine learning model that identifies user-friend pairs where social cues (i.e., friends’ likes) have enhanced effects on users' ad engagement, using high-dimensional data. The impact of social cues in social ads is jointly determined by the characteristics of influencers (influence), those influenced (susceptibility) and their relationships. Our model searches for the user-friend pairs exhibiting enhanced effects of social cues. Such pairs are characterized by the interpretable rules constructed from the demographic, behavioral and network characteristics of users, their friends shown in ads and the tie strength and social embeddedness between the users and the friends. We present a Bayesian framework for learning a rule set. The Bayesian model consists of a prior to control the size and shape of a rule set and a Bayesian logistic regression to characterize interactions of features, treatment and subgroup membership. We also develop an efficient inference method for learning a MAP model. Our results on ad engagements for different product categories demonstrate that the enhanced effects of social cues associated with the identified pairs can be much larger than that on the entire population. As a result, targeting the contagious user-friend pairs is an effective strategy to improve social ads’ effectiveness.

Does Monetary Incentive Lead to Better Stock Recommendations on Social Media?  
(with Hailiang Chen and Jeffrey (Yu) Hu)

Draft available upon request

Social media not only is a new channel to obtain financial market information but also becomes the venue for investors to share and exchange investment ideas. We examine the performance consequences of providing monetary incentive to amateur analysts on social media and its implications for crowd-sourced equity research. We find that monetary incentive is effective in increasing the amount of content outputs but does not lead to better stock recommendations. Additional analysis suggests that monetary incentive results in wider stock and industry coverage, a sign of increased content diversity. This study contributes to the understanding of incentive mechanisms for social media communities in the financial context.