Abstract
This dissertation examines the interaction between personalized recommender systems and social media users' digital consumption and privacy protection through large-scale field experiments. Motivated by the causal identification challenges commonly faced in field experiments, the dissertation further proposes a new estimator for encouragement designs when the exclusion restriction is violated. The first chapter studies how modifying a personalized recommender system to promote content diversity affects both the amount and diversity of users’ digital consumption. A fundamental trade-off in designing personalized recommender systems lies between exploitation and exploration: deciding whether to recommend familiar content that users favor or to introduce new, diverse content that may interest them in the future. We empirically examined this trade-off in a real-world-scale recommender system through a partnership with a leading global music-streaming service platform. We conducted a large-scale field experiment where users were randomly assigned to receive video recommendations from either the platform’s standard algorithm or a modified version that recommends more diverse content. Contrary to industry expectations, increasing the diversity of the recommender algorithm does not enhance users’ consumption diversity; instead, it marginally reduces their click days on the platform. However, among active users—who account for most of the platform’s content usage—a 1% increase in recommendation diversity resulted in a 0.55% increase in their consumption diversity, without affecting overall consumption levels. The increase in consumption diversity corresponds with the more accurate prediction of their consumption preferences. The results suggest that the platform should tailor its algorithm to recommend more diverse content for active users. The second chapter investigates how privacy protection measures offered by social media platforms affect users' digital consumption, both intentionally and unintentionally. In this digital age, privacy concerns are escalating with the increased collection and use of personal information. Consequently, regulators have increasingly pushed companies to make the use of personal information more transparent and protect users with more privacy protection measures. However, the impacts of these policies on user behaviors and welfare remain unclear. We investigate this issue through a large-scale field experiment on a leading global social media platform. In the experiment, treated users were offered a privacy protection option to disable the "People You May Know" (PYMK) recommender algorithm, which could display their content to users whom the algorithm predicts are their friends elsewhere. Control users were neither informed about nor allowed to disable this function. Interestingly, we found that treated users, on average, decreased their video usage time by 0.78% compared with control users. However, the usage time of those treated users who chose to disable the function increased by 14.62% compared with a matched sample. We interpret these results as the privacy protection option having two consequences: On one hand, it raises users' concerns by reminding them that their personal information is being used, thereby unintentionally reducing their usage time. On the other hand, it allows users to disable the use of personal information, which eliminates such concerns and leads to an increase in usage time as intended. To evaluate the social welfare impact of our and alternative privacy protection measures, we estimate a structural model that describes users’ decisions regarding usage and disabling the PYMK function, and use the results to run counterfactuals. The results demonstrate that different policies could lead to drastically different social welfare outcomes, highlighting the importance of considering both intended and unintended consequences. In particular, we find that lowering the costs of adopting the PYMK protection option is crucial not only for increasing overall social welfare but also for aligning the incentives of consumers and the platform, potentially creating a win-win scenario for both. The third chapter proposes a new estimator for causal inference in field experiments when the exclusion restriction is violated. Encouragement design is widely used in field experiments (or randomized controlled trials) when noncompliance in the treatment group, control group, or both is non-negligible. The standard identification strategy is to use 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 interested outcome variable. 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.
Committee Chair
Tat Chan
Committee Members
Dennis Zhang; Meng Liu; Xiang Hui; Zhengling Qi
Degree
Doctor of Philosophy (PhD)
Author's Department
Marketing
Document Type
Dissertation
Date of Award
4-30-2025
Language
English (en)
Recommended Citation
Chen, Guangying, "Essays on Social Media Algorithms and Causal Inference Methods" (2025). Olin Business School Theses and Dissertations. 56.
The definitive version is available at https://doi.org/10.7936/mfxj-pj21