Research

Research Interests

Substantive: Digital / Mobile Marketing, Promotions, and Advertising

Methodological: Bayesian Statistics, Field Experiments, and Natural Language Processing

Research Projects

Wait For Free: A Consumption-Decelerating Promotion for Serialized Digital Media

Abstract: Promotions for digital goods have typically focused on enticing users to hasten their consumption. Here, we investigate a novel deceleration-incentivizing promotional policy, “Wait For Free” (WFF), applied to serialized digital content – sequences of interconnected episodes – monetized via episode-level paywalls. Specifically, customers can sample early episodes of promoted series for free, and can continue to do so by waiting a pre-specified time; or for those unwilling to wait, by paying. WFF can draw users to start viewing a promoted series, and generate revenue through two sources: impatient users opting to pay to consume the next episode immediately; and additional users continuing through to paid-only episodes at the end. We analyze large-scale viewership data from a platform that enacted WFF for digital comics. A comic-level Difference-in-Differences (DiD) analysis provides robust evidence that WFF boosts paid viewership for the promoted series, and the degree of lift varies by user type and over time. A more granular episode-level analysis incorporating inter-comic spillovers and promotional lift heterogeneity suggests that WFF can boost net-of-cannibalization revenue at the platform level: specifically, the model-optimized set of promoted comics performs roughly 18% better than the firm-enacted one and 25% better than the no-promotion baseline; furthermore, WFF and paid-only episodes each receive nontrivial degrees of lift, 70% and 59% respectively.

Keywords: serialized digital content, Wait For Free, promotional spillovers, digital content monetization, promotion optimization  

Politically Biased Moderation Drives Echo Chamber Formation: An Analysis of User-driven Content Removals on Reddit

Abstract: Echo chambers-online spaces where individuals are met with reinforcing viewpoints and insulation from opposing viewpoints-are of increasing interest to social media platforms and policymakers amidst rising political polarization. Selective participation and algorithmic recommendations are commonly cited as drivers of echo chamber formation. In this work, we document a novel mechanism for the formation of echo chambers on social media sites: the politically biased content removal decisions of user moderators. Applying a combination of natural language processing and network analysis to characterize political leanings on a dataset of removed comments from 1.2 million users on Reddit, we document political bias in content removal: moderators are significantly more likely to remove content that differs from their political orientation. Further, using a matching approach, we show that content removals have an indirect chilling effect on censored users' subsequent political speech on the same channel. Finally, we conduct a counterfactual simulation and demonstrate that politically biased content removal increases echo chamber intensity. These findings are of broad interest to users, platforms and regulators who weigh the costs and benefits of freedom of expression with the logistics of moderating massive online spaces.

Comparing the Effectiveness of Retargeting and Acquisition Online Banner Ads: A Flexible Approach to Estimating Ad Stock

Abstract: One of the earliest and most extensive literatures in quantitative marketing concerns the measurement of ad effectiveness. Because ads do not “work” immediately, and often require multiple exposures, econometric approaches to measuring the cumulative impact of advertising typically rely on the concept of a (latent) “ad stock” or “goodwill”. Owing to the influential work of Nerlove and Arrow (1962), it has often been assumed that the contribution of each ad to the ad stock decays exponentially over time at a constant rate, captured by a single parameter common across all users and ad types. Here we examine how two different types of online advertising campaigns – for acquisition and retargeting – differentially affect online users’ behaviors by proposing flexible parametric and nonparametric Bayesian approaches for regularizing over past ad weights. Specifically, we make use of an online panel of individual-level ad impression data for a French financial services firm and online activity data of internet users who were shown its banner ads to demonstrate that relaxing the restrictive “single parameter decay” assumption allows us to more flexibly capture the differential impacts of acquisition and retargeting ads on website visits, both in terms of their initial impact on the day of exposure and their lingering effects over time. Our results also suggest that a constant decay rate over time may be an over-regularization, potentially entailing substantive artifacts dictated by the common Nerlove-Arrow assumption.

Keywords: online banner ads, retargeting ads, ad stock, Nerlove-Arrow, Gaussian processes

To Whom, When, and What to Ask?: Mitigating Unhealthy Behaviors and Detecting Relapse with Customized Real-Time Mobile Interventions

Abstract: Ecological Momentary Assessments (EMAs), a type of repeated real-time self-measurements, have been used extensively in health-related studies, especially those regarding substance usage, due to their capabilities both as a form of intervention and as an inobtrusive way of collecting granular data for predicting individual-level outcomes. However, one often-cited practical concern to using EMAs is response fatigue, leading to participant inattentiveness and even attrition, which may limit researchers’ ability to gather diagnostic information and hinder participants’ abatement progress. To assuage such concerns, we propose a Bayesian dynamic factor model that allows researchers to curtail EMA activations for individuals more likely to experience response fatigue and whose responses are expected to vary little over time, as well as to omit participant-specific “temporally redundant” items (i.e., for which responses are expected remain stable) from EMAs. We will be applying the framework to a rich panel dataset from a smoking cessation program where we expect the balance between participants’ response burdens and researchers’ ability to gather timely and relevant data to result in a decline in EMA noncompliance and an improvement in health outcomes.

Optimizing One-shot Promotional Inducements in a Two-sided Choice Setting: An Application to Scholarship Offerings

Abstract: Each year, university admissions administrators face the complex task of distributing scholarship funds across admitted students with the objective of luring the most desirable cohort (in expected gender parity, standardized test scores, etc.) to accept their offers, subject to various stochastic constraints (e.g., expected total scholarship expended; cohort size). This constitutes a unique two-sided choice setting in which schools, when determining the optimal scholarship amounts, need to incorporate their beliefs about each student’s offers from competing ones. We use a dataset from a university graduate program to make inferences on each student’s unobservable choice (admissions-scholarship) set, and then evaluate the prospects of converting them under various scholarship amounts. To accurately assess the impact of scholarships on students’ propensity to accept an offer, we collaborate with the financial aid directors to conduct two rounds of efficient field experiments on portions of admitted students one year apart. Specifically, scholarship amounts are orthogonally adjusted within each stratum of estimated offer acceptance propensities to induce maximal variations in treatment.