2025 ASA Statistics in Marketing Doctoral Research Award Finalists Presentation

Encouragement design is widely used in randomized controlled trials when noncompliance in the treatment group, control group, or both is non-negligible. The standard identification strategy uses the randomized group assignment as an instrumental variable to estimate the local average treatment effect (LATE). In many experiments, however, this instrument may violate the exclusion restriction condition, because the encouragement can directly impact the outcome variable of interest. We develop a new root-n-consistent estimator using the randomized group assignment to construct an instrument that relies on the heteroskedasticity of treatment intensities between groups. Our identification strategy can recover not only LATE but also the direct impact of the encouragement on outcomes. We further propose a min-max estimator for consistent nonparametric estimation of heterogeneous treatment effects. Finally, we conducted a large-scale field experiment with a social media platform to study how expanding users' social networks influences their platform usage. While ordinary least squares and standard two-stage least squares estimators report a positive effect, our estimator suggests that the effect comes solely from the encouragement. We find evidence supporting the null effect of network expansion, indicating that firms may waste resources on false positives when the exclusion restriction is violated in their field experiments. 

Firms are increasingly interested in conducting network experiments through peer encouragement designs to causally quantify the potentially heterogeneous effects of social marketing programs. For example, researchers randomly sample ego networks (where each ego network consists of one ego and the alters who are directly connected to the ego) for assignment to treatment and control conditions, and then estimate the direct treatment effects based on responses of egos and the indirect treatment effects based on responses of alters. To satisfy the Stable Unit Treatment Value Assumption (SUTVA), researchers have typically adopted the excluding approach by removing those contaminated ego networks defined by some criteria. However, this practice often results in two problems documented in the literature: underrepresentation (i.e., the post-exclusion samples are not representative of the population on key network attributes such as degree and clustering coefficient), and undersupply (i.e., the post-exclusion samples consist of a limited number of ego networks). We propose a method that directly addresses the underrepresentation and undersupply problems, and efficiently generates proper and representative treatment and control samples satisfying SUTVA. We employ the Metropolis-Hastings algorithm to obtain optimal samples that minimize the distance between the samples and the population network based on the joint distribution of some key network attributes that may significantly influence the general properties of networks and the magnitudes of treatment effects. Our method comprises three key input modules: the selection of key network attributes upon which the sample and population distributions will be compared, the definition of the distance between the sample and population distributions, and the specification of desired sample sizes. By adjusting these three modules, researchers can flexibly generate proper and representative samples across a variety of network conditions as needed to facilitate causal inferences in network experiments employing peer encouragement designs. Through extensive simulations using both simulated and real population network data, our results collectively demonstrate that underrepresentation and undersupply become more pronounced for the post-exclusion samples when the required sample size of ego networks is large, the population network is not of large scale, the population network is dense, and when both first- and second-degree contamination are considered. We demonstrate the usefulness and boundary conditions of the proposed method in generating larger and more representative samples. We also demonstrate that the representative samples generated by our proposed method effectively improve the efficacy of the estimation and statistical inference. The proposed representative sampling method can be adapted and incorporated into many applications to help firms improve designs of social marketing programs. 

Online sellers have been adopting AI learning algorithms to automatically make product pricing and advertising decisions on e-commerce platforms. When sellers compete using such algorithms, one concern is that of tacit collusion—the algorithms learn to coordinate on higher than competitive prices which increase sellers' profits, but hurt consumers. This concern, however, was raised primarily when sellers use algorithms to decide on prices. We empirically investigate whether these concerns are valid when sellers make pricing and advertising decisions together, i.e., two-dimensional decisions. Our empirical strategy is to analyze competition with multi-agent reinforcement learning, which we calibrate to a large-scale dataset collected from Amazon.com products.
Our first contribution is to find conditions under which learning algorithms can facilitate win-win-win outcomes that are beneficial for consumers, sellers, and even the platform, when consumers have high search costs. In these cases the algorithms learn to coordinate on prices that are lower than competitive prices. The intuition is that the algorithms learn to coordinate on lower advertising bids, which lower advertising costs, leading to lower prices for consumers and enlarging the demand on the platform.
Our second contribution is an analysis of a large-scale, high-frequency keyword-product dataset for more than 2 million products on Amazon.com. Our estimates of consumer search costs show a wide range of costs for different product keywords. We generate an algorithm usage index based on the correlation patterns in prices and find a negative interaction between the estimated consumer search costs and the algorithm usage index, providing empirical evidence of beneficial collusion. We predict that in more than 50% of the product markets, consumers benefit from tacit collusion facilitated by algorithms.
We also provide a proof that our results do not depend on the specific reinforcement learning algorithm that we analyzed. They would generalize to any learning algorithm that uses price and advertising bid exploration.
Finally, we analyze the platform's strategic response through adjusting the ad auction reserve price or the sales commission rate. We find that reserve price adjustments will not increase profits for the platform, but commission adjustments will, while maintaining the beneficial outcomes for both sellers and consumers.
Our analyses help alleviate some worries about the potentially harmful effects of competing learning algorithms, and can help sellers, platforms and policymakers to decide on whether to adopt or regulate such algorithms.
 

This paper studies how the attribution algorithms used in online ad auctions affect the strategic interactions between advertisers and publishers, and it investigates optimal attribution strategies for advertisers. Because online advertisers typically advertise with several publishers to increase their reach, users may be exposed to ads from multiple publishers before converting. The attribution challenge for an advertiser is to measure the contributions of each publisher's advertising on conversions. These attributed conversion measures are crucial because they serve as inputs into the algorithms that advertisers use to determine bids in future ad auctions. The attribution challenge is aggravated by the fact that publishers typically have access to more information than advertisers, such as user behavior on their sites. This information asymmetry can lead to a moral hazard problem: publishers can exploit their information advantage to target ads to users who are likely to result in attributed conversions, rather than to users with large incremental ad effects. To investigate this misalignment of interests between advertisers and publishers, I cast the attribution problem as an incentive design problem. Using a structural model, I first characterize the dynamic incentives created by standard attribution algorithms and derive the advertiser's optimal strategy. I find that the advertiser's optimal strategy takes the form of team incentives, where each publisher is compensated only when a conversion is preceded by an ad impression by only that publisher. Counterfactual analysis shows that the optimal strategy increases the advertiser's ROI on the order of 20-40% compared with standard attribution algorithms. The findings highlight the importance of considering the dynamic incentives that measurement tools generate. 

Thumbnails, reduced-size preview images or clips, serve as pivotal visual cues that help consumers navigate through video selection while "previewing" for what to expect in the video. This paper provides a scalable framework combining state-of-art computer vision techniques, novel video platform design and novel use of Bayesian learning model in a high-dimensional context to study: (i) how thumbnails, relative to video content, affect viewers' behavior, and (ii) how to optimize video thumbnail selection under different creator objectives. To achieve this, we first propose a video mining procedure that automatically decomposes high-dimensional video data into interpretable features using computer vision, deep learning, and LLMs. Motivated by behavioral theories such as expectation-disconfirmation theory and Loewenstein's theory of curiosity, we then construct theory-based measures to assess the role through which thumbnails affect video reactions. Using both secondary data from YouTube and a novel video platform called "CTube" that we build to exogenously randomize thumbnails across videos, we find that content disconfirmation between the thumbnail and the video leads to opposing effects. It leads to more views, higher watchtime but lower post-video engagement (e.g., likes and comments). To further investigate the underlying behavioral process, we build a Bayesian learning model in a high-dimensional context in which consumers' decisions to click on a video and continue watching the video are based on their priors (the thumbnail) and updated beliefs of the video content (the video's frames, characterized as multi-dimensional and correlated video topic proportions). We show that viewers overall prefer watching videos longer when there is a higher disconfirmation between their initial and updated content beliefs, suggesting one role of thumbnails as generating curiosity for what may come next in the video. In addition, viewers prefer less disconfirmation before observing the thumbnail, highlighting the role of disconfirmation may change before and after the thumbnail. Using the model estimates, we then run counterfactual analyses to propose optimal thumbnails and compare them with current practices of thumbnail recommendation to guide creators and platforms in thumbnail selection. Our framework provides a scalable and cost-effective way to optimize thumbnail selection and design for companies, with broad applicability to content summarization across various formats (e.g., book covers, movie posters).