Ziyao Tang

Working Papers

• Data Valuation in Marketing Collaborations, with Guang Zeng and Paul B. Ellickson (Under 2nd Round Review at Management Science, Wharton AI & Analytics for Business Data Grant)

Abstract: Better data can make a partnership worse. We show that when a legacy contract misaligns incentives, sharing data can benefit the party controlling the decision while harming the party providing the data. We study this mechanism in a retailer-bank co-branded credit-card program, using linked retail and banking records to examine how access to retailer histories changes bank-controlled approval decisions and each partner's payoff. Under the observed contract, the bank controls approval, yet the bank and retailer value the same applicants differently. We use a doubly robust difference-in-differences design to estimate the average and heterogeneous payoff effects of approval. Approval creates positive joint value among approved applicants on average, yet the heterogeneous payoff estimates reveal that the bank and retailer prefer opposite decisions for 45% of applicants. Building on these estimates, we use out-of-sample policy learning to compare approval policies learned with and without retailer data. Under the legacy contract, adding retailer data sharpens the bank's targeting, enabling it to reject applicants who are costly to the bank but valuable to the retailer. The bank gains, but the retailer loses more, so joint value falls. A moderate contractual realignment reduces this misalignment; under the realigned contract, introducing the same data benefits both parties. Data access, decision rights, and payoff allocation must therefore be designed together.

• Transparency and Hostility: The Unintended Effects of Geographic Disclosure on Online Identity Attacks, with Guang Zeng and Huaxia Rui (Draft Available upon Request)

Abstract: This paper examines the causal effect of mandatory location disclosure on identity attacks in online discussions. We analyze a natural experiment on Zhihu, China's largest question-and-answer platform, which implemented a policy in May 2022 requiring the display of users' geographic locations. Using causal inference methods, we find that location disclosure significantly increases the likelihood of identity attacks in user comments. While the policy was designed to enhance accountability through transparency, our results demonstrate that displaying location information  amplifies regional hostility. The policy's impact varies systematically with cultural distance between user pairs. Identity attacks become more frequent when commenters and recipients are from culturally distant regions, with interactions at the greatest cultural distance exhibiting approximately 30% more identity attacks compared to those at the closest distance. These findings reveal unintended consequences of transparency-focused design choices and provide critical insights for platform governance and online discourse moderation.

• When Linear IV Estimator Fails: Avoiding Pitfalls in Causal Effect Estimation in Targeted Marketing, with  Guang Zeng and Paul B. Ellickson

Abstract: Linear estimators may fail to recover causal effects in the presence of treatment effect heterogeneity, introducing bias. While prior literature recommends nonparametric approaches to eliminate this bias, these estimators often suffer from high variance. We show that in targeting applications, where the goal aligns more closely with predictive performance, linear estimators can outperform their nonparametric counterparts by accepting a small amount of bias in exchange for substantial variance reduction. This highlights an important bias-variance tradeoff in causal effect estimation for decision-making contexts.

Work in Progress

• Targeting as Exploration, with Guang Zeng and Paul B. Ellickson

Abstract: Many targeting problems rely on supervised policy learning algorithms. However, in marketing applications, interventions often take time to produce observable outcomes, which limits the ability to update targeting strategies promptly. This paper reframes the targeting problem as a contextual bandit problem. By integrating causal inference techniques with bandit algorithms, we propose a targeting approach that balances exploration and exploitation. Our results demonstrate that incorporating exploration improves efficiency relative to traditional supervised learning methods, particularly in environments with delayed feedback.