Virtual Quantitative Marketing Seminar

Previous Talks:




Generative AI in Equilibrium: Evidence from a Platform Marketplace, Recording 


Generative AI (GenAI) promises to greatly change the means of production. These technologies will impact both the input goods markets (for example, labor) and the output product markets. In both cases, depending on what types of production GenAI advantages, market equilibrium will shift - leading to differing welfare and policy implications. We examine the introduction of GenAI to a major stock images platform in order to evaluate the equilibrium implications of these technologies. Our setting allows us to observe many markets with different ex-ante market structures and provides us with a control group of markets for which GenAI is not introduced. Using a difference-in-difference methodology, we provide causal evidence of how the introduction of GenAI products affect entry, product variety, sales and market concentration.


Title: Investigating the Impact of Advertising on Smoking Cessation: The Role of DTC Prescription Drug Advertising, Recording


This study documents the impact of different forms of advertising on the demand for cigarettes and their substitutes. We examine the effects of educational public service announcements (PSAs), e-cigarette advertising, direct-to-consumer (DTC) advertising for tobacco cessation prescription drugs, and advertising by nicotine replacement therapies (NRTs) such as patches and gums on the demand for cigarettes, e-cigarettes, NRTs, and smoking cessation prescription drugs. Among these different forms of advertising, we find evidence only for the effectiveness of DTC advertising of prescription drugs in reducing cigarette demand. DTC advertising increased new prescriptions for tobacco cessation, increased outpatient visits for mental health and substance abuse, and led to a significant spillover effect on the sales of over-the-counter alternatives. This spillover varies systematically with the level of insurance coverage for smoking cessation medications across different regions. Our results highlight the role of DTC advertising and insurance coverage in smoking cessation, providing valuable insights for policymakers. Importantly, our findings suggest that banning DTC advertising could have unintended consequences, resulting in increased cigarette sales and overall nicotine consumption.


Title: The Effects of Delay in Bargaining: Evidence from eBay, Recording


Delay in negotiations is common in many settings, but the effects of delay have rarely been studied empirically in the field. We measure the causal effect of delay using data from millions of negotiations on eBay. We find that for both buyers and sellers, the longer the bargaining party delays, the less likely the opponent is to continue the negotiation by countering. However, the downstream consequences vary. The more the seller delays, the more likely the negotiation will fail, but the more the buyer delays, the more likely the seller will accept the buyer's offer. The effects of delay are robust; they exist even under short amounts of delay (under 6 hours) and for negotiations for low-priced goods. We find that these effects are consistent with models of strategic delay, in which delay acts as a signal of bargaining power.



Title: Platform-Generated Quality Ratings: Theory, Empirics and Welfare Implications (joint with Jie Bai,   Daniel Xu, and Zhe Yuan), Slides


In order to address the issue of asymmetric information and promote market regulation, digital platforms have implemented platform-generated rating (PGR) systems to provide quality information to the market. Unlike user-generated rating (UGR) systems, PGRs rely on platform-generated data including product quality, customer service, and logistics. To better understand the impact of PGRs, we conducted a study in collaboration with a large E-commerce platform, exploring both empirical and theoretical implications on consumer beliefs, sellers' quality incentives, and market outcomes. Our empirical analysis utilized a regression-discontinuity method to examine customer response to high PGR sellers. Results indicate that customers tend to purchase more from high PGR sellers, suggesting that PGRs can be effective in signaling quality to consumers. We also found that the presence of a PGR system affects sellers' quality incentives and the distribution of equilibrium quality. These findings demonstrate that PGRs have the potential to promote quality competition among sellers in the market. Finally, we conducted a counterfactual experiment to evaluate the welfare consequences of PGR adoption. Our (tentative) results suggest that PGR adoption can have a positive impact on overall welfare. Overall, our study highlights the potential benefits of PGRs in promoting market efficiency and enhancing consumer welfare.



Title: Bad-drug Ads or Killer Ads: The Effects of Drug Injury Advertising on Viewers’ Health Outcomes


Consumers are frequently exposed to drug injury lawsuit advertisements, which highlight the potential dangers associated with a particular drug, with the intent to recruit potential lawsuit plaintiffs. Physicians have expressed concerns that drug injury ads might lead patients to misperceive the risks and benefits associated with their drugs, leading them to discontinue their medication. I analyze the effects of drug injury ads on prescriptions in the context of anticoagulants, which is a class of drugs primarily taken by elderly people to help prevent strokes. I observe a decline in the total number of filled prescriptions of anticoagulants, which lead to meaningful changes in health outcomes, such as increases in inpatient hospital visits and death rates for conditions treated with anticoagulants with no changes for placebo diagnoses.



Title: The Uneven Impact of Generative AI on Entrepreneurial Performance (with Nicholas G. Otis, Rowan Clarke, Solène Delecourt, and Rembrand Koning)


There is a growing belief that scalable and low-cost AI assistance can improve firm decision-making and economic performance. However, running a business involves a myriad of open-ended problems, making it hard to generalize from recent studies showing that generative AI improves performance on well-defined writing tasks. In our five-month field experiment with \n Kenyan entrepreneurs, we assessed the impact of AI-generated advice on small business revenues and profits. Participants were randomly assigned to a control group that received a standard business guide or to a treatment group that received a GPT-4 powered AI business mentor via WhatsApp. While we find no average treatment effect, this is because the causal effect of generative AI access varied with the baseline business performance of the entrepreneur: high performers benefited by just over 20% from AI advice, whereas low performers did roughly 10% worse with AI assistance. Exploratory analysis of the WhatsApp interaction logs shows that both groups sought the AI mentor's advice, but that low performers did worse because they sought help on much more challenging business tasks. These findings highlight how the tasks selected by firms and entrepreneurs for AI assistance fundamentally shape who will benefit from generative AI.



Title: Recommending for a Multi-Sided Marketplace: A Multi-Objective Hierarchical Approach


Recommender systems play a vital role in driving the long-term value for online platforms. However, developing recommender systems for multi-sided platforms faces two prominent challenges. First, different sides have different and possibly conflicting utilities. Recommending in this context entails jointly optimizing multiple objectives. Second, many platforms adopt hierarchical homepages, where items can either be individual products or groups of products. Off-the-shelf recommendation algorithms are not applicable in these settings.


To address these challenges, we propose MOHR, a novel multi-objective hierarchical recommender. By combining machine learning, probabilistic hierarchical aggregation, and multi-objective optimization, MOHR efficiently solves the multi-objective ranking problem in a hierarchical setting through an innovative formulation of probabilistic consumer behavior modeling and constrained optimization. We implemented MOHR on one of the world's largest food delivery platforms, and demonstrate that long-term profit maximization can be achieved through a multi-objective approach as we proposed, outperforming existing single-score based approaches. Moreover, the MOHR framework offers managers a mathematically principled tool to make quantifiable and interpretable trade-offs across multiple objectives for long-term profit optimization. Online experiments showed significant improvements in consumer conversion, retention, and gross bookings, resulting in a \$1.5 million weekly increase in revenue. As a result, MOHR has been deployed globally as the recommender system for the food delivery platform’s app homepage



Title: Deeply Personal Marketing


The emergence of machine learning, and in particular deep learning, has had a significant impact on science and society. Typically, deep learning models are used for various forms of prediction but they have the potential to be used in alternative ways in the context of the economics and marketing. One particular use case is the approximation of individual heterogeneity. The key idea being that with the abundance of unit-level data we should be able to construct measures that describe individual differences flexibly and accurately. As a natural next step this heterogeneity can then be used to inform and design personalized policies.


Deep learning models are particularly well suited to this task because of their flexibility, their compatibility with economic structure and the availability of scalable, easy to use software. In this talk I will introduce a framework to learn heterogeneous effects of economic decisions using deep learning. The proposed framework takes a standard economic model and recasts the parameters as flexible functions which afford the capture of heterogeneity across agents based on high dimensional or complex observable characteristics. These parameter functions retain the interpretability and economic meaning the original parameters and can be used in much the same way scalar parameters are to construct measures, conduct counterfactuals or design policies of interest. In addition, the framework also provides an automatic inference engine based on a computationally viable influence function approach. This automatic approach leverages recent developments in automatic differentiation engines to allow the researcher to conduct statistical inference without the need for any additional conceptual effort.


Following this introduction, I will demonstrate the implementation of this approach using a series of applications including optimal targeting, price personalization and the generation of text based policy interventions. The applications will aim to showcase the choices that need to be made in the implementation of this framework, the viability of this approach, it advantages, as well as inconveniences and limitations. The use of such ML-based personalization approaches also open up a set of concerns and questions about downstream consequences. I will conclude the talk by offering some discussion on these from the perspective of consumers, firms and policy makers. 


This presentation is based on a series of projects with various co-authors. Details on the framework are contained in https://arxiv.org/pdf/1809.09953.pdf and https://arxiv.org/pdf/2010.14694.pdf



The effect of First-Price vs. Second-Price auctions on Display Advertising Bidding and Revenues


In 2019, the $60 billion display advertising market was shaken when Google transitioned from a second-price to first-price auction. Many reported rationales underpinned the change, including the increased transparency of the auction (an advertiser pays what they bid rather than relying on the auction partner to report the closing price), the facilitation of bidding across partners, greater allocative efficiency, higher revenues for publishers, and criticisms of Google’s so-called last look algorithm which purportedly enabled its exchange users to outbid competitors from competing exchanges. Consistent with industry views, Despotakis et al. theorize that the change to the first-price auction, precipitated by header bidding, leads to higher clearing prices.


Despite these beliefs, little empirical evidence exists to ascertain how the change affected auction display outcomes. Bidding became more complex as advertisers turned to various bid shading tools from different vendors to manage the complex task of optimizing their bids. Left unanswered is whether this change did lead to higher bid advertiser CPMs and greater publisher revenue.


Using data collected during the duration of the switch, we find that i) advertisers adjust bids quickly to changes in format, ii) the FPA does not increase allocative efficiency, iii) advertisers bid too low, and iv) publisher revenues suffer under FPA. We consider two possible explanations for these deviations from theoretical predictions: a) advertisers have difficulty computing optimal bids and b) advertisers do not fully

account for competitive bidding.



Estimating the Value of Offsite Data to Firms: Evidence from 1.5 Studies


Browsing activity and purchase data are frequently tracked across applications and used to help target online advertising. These types of ‘offsite’ data are viewed as highly valuable for advertisers, but their usage faces increasing headwinds. In this presentation, I'll discuss results from two related projects, focusing mostly on the first.

 

In the first project, we study how much firms benefit from using offsite data in their ad delivery. With this goal in mind, we conduct a large-scale, randomized experiment that includes more than 70,000 advertisers on Facebook and Instagram. We first estimate advertising effectiveness at baseline across our broad sample. We then estimate how much less effective the same campaigns would be were advertisers to lose the ability to optimize ad delivery with offsite data. In each of these cases, we use recently developed deconvolution techniques to flexibly estimate the underlying distribution of effects. We find a median cost per incremental customer at baseline of $42.04 that under the median loss in effectiveness would rise to $56.77, a 35% increase. Further, we find ads targeted using offsite data generate more long-term customers per dollar than those without, and losing offsite data disproportionately hurts small scale advertisers. [with Anna Tuchman, Bradley Shapiro, and Robert Moakler, see paper]

 

In the second (early stage) project, we aim to better understand the effects of a major policy change in the offsite data ecosystem -- the introduction of Apple's App Tracking Transparency (ATT) framework in iOS 14.5 -- on firms in the US. To that end, we have conducted preliminary analyses on several online and offline firm outcomes I'll discuss during the talk. [with Daniel Deisenroth, Utsav Manjeer, Zarak Soheil, and Steve Tadelis]



Using Domain Knowledge to Enhance Deep Learning for Emotional Intelligence


Emotion identification provides information to managers and more granular emotions provide more specific information. Whereas most existing emotion classifiers focus on a coarse set of emotions (e.g., sadness, joy), we focus on a larger set of 24 granular emotions (e.g., disappointment, neglect, sympathy). Granular classification is challenging because it aims to distinguish subordinate-level categories, which often have small inter-class variation but large intra-class variation. We propose a hierarchical classification architecture for identifying granular emotions in unstructured text data. The proposed classifier takes advantage of a semantic network of emotions from psychology which maps out how individuals categorize emotions. In the first stage of classification, input text is classified as being one or more coarse emotions. In the second stage, a more granular emotion is identified based on the coarse classification of the previous stage. Using self-labeled Twitter data, we find that our proposed hierarchical classifier outperforms a single-stage flat classifier in terms of F1 by increasing recall at the cost of precision. In addition, the hierarchical structure increases the explainability of the model by enabling interpretation at multiple levels, providing additional insight to end users.



Behavioral Skimming: Theory and Evidence from Resale Markets


Lack of information distorts markets, and communicating product value to potential consumers is a crucial ingredient of marketing strategy. However, a large body of behavioral research has suggested that even when information is easily accessible, consumers often fail to attend to it. Evidence of consumer inattention has been studied in various settings, both inside and outside the laboratory. How an intermediary should react when communication fails as a result of consumers’ failure to use the provided information is unclear. Can or should firms profit from asymmetric information caused by consumer inattention? If so, by how much? Does competition alleviate the effect? We consider these questions in the context of resale markets, both theoretically and empirically. The theoretical model demonstrates that a centralized intermediary can extract surplus from serving consumers who are less attentive and, as a result, overestimate the product value. We test the theory using a detailed dataset of millions of automobile transactions from a seven-year period. First, we find clear evidence of a specific type of inattention: Buyers exhibit left-digit bias and systematically underestimate the depreciation of vehicles that have odometer readings immediately below round cutoffs. Second, the estimated level of inattention is twice as high in dealership transactions than in consumer transactions, so that dealers make a significantly higher margin on such vehicles. Third, we estimate the supply-side response to consumer inattention and find 2.53% additional transactions, compared to the no-inattention counterfactual. As a result, the average margin is 1.8% higher, leading to an aggregate increase in operating profits of 4.37%, or about $422 million, within the seven-year sample period. The surplus obtained by the product owners who sell in the market increases by about 2.77%. Back-of-the-envelope calculations imply that U.S. used vehicle dealers’ annual profits attributable to consumer inattention are about $700 million.



Designing Effective Music Excerpts


Excerpts are widely used to preview and promote musical works. Effective excerpts induce consumption of the source musical work and thus generate revenue. Yet, what makes an excerpt effective remains unexplored. We leverage a policy change by Apple that generates quasi-exogenous variation in the excerpts of songs in the iTunes Music Store to estimate that having a 60 second longer excerpt increases songs' unique monthly listeners by 5.4% on average, by 9.7% for lesser known songs, and by 11.1% for lesser known artists. This is comparable to the impact of being featured on the Spotify Global Top 50 playlist. We develop measures of musical repetition and unpredictability to examine information provision as a mechanism, and find that the demand-enhancing effect of longer excerpts is suppressed when they are repetitive, too predictable, or too unpredictable. Our findings support platforms' adoption of longer excerpts to improve content discovery and our measures can help inform excerpt selection in practice.



Demand Analysis under Latent Choice Constraints


Consumer choices are constrained in many markets due to either supply-side rationing or information frictions. Examples include matching markets for schools and colleges; entry-level labor markets; limited brand awareness and inattention in consumer markets; and selective admissions to healthcare services. Accounting for these choice constraints is essential for estimating consumer demand. We use a general random utility model for consumer preferences that allows for endogenous characteristics and a reduced-form choice-set formation rule that can be derived from models of the examples described above. The choice-sets can be arbitrarily correlated with preferences. We study non-parametric identification of this model, propose an estimator, and apply these methods to study admissions in the market for kidney dialysis in California. Our results establish identification of the model using two sets of instruments, one that only affects consumer preferences and the other that only affects choice sets. Moreover, these instruments are necessary for identification – our model is not identified without further restrictions if either set of instruments does not vary. These results also suggest tests of choice-set constraints, which we apply to the dialysis market. We find that dialysis facilities are less likely to admit new patients when they have higher than normal caseload and that patients are more likely to travel further when nearby facilities have high caseloads. Finally, we estimate consumers’ preferences and facilities’ rationing rules using a Gibbs sampler.



Platform Vertical Integration and Consumer Welfare: Evidence from a Field Experiment


Many firms, from retailers to investment management companies, offer their own products alongside products sold by competitors. This vertical integration, although common across the economy, is particularly controversial in a digital setting. In this work, we study the effects on consumer choice of vertical integration practices by Amazon, which offers products from its own brands (Amazon Basics and others) that compete with other brands. To study this question, we run a field experiment in which some users are randomized to not be shown Amazon owned brands as they browse Amazon. Our experiment is made possible by a new piece of software, Webmunk, which is a browser extension that can change the look and feel of websites as people browse them. Our key preliminary finding is that shoppers substitute towards cheaper products when products owned by Amazon are not available.



Are Menthol Cigarettes More Addictive? A Cross-Category Comparison of Habit Formation, recording, slides


Menthol cigarettes have been banned in parts of the U.S. based on the premise that they are more addictive than non-menthol cigarettes. In this paper, I propose a framework and a novel identification strategy to compare the extent to which consumption is driven by addiction across different categories based on consumer panel data. Using variation in the length of temporary breaks in consumption, I compare the effect of past consumption to that of static preferences in driving consumption levels for each cigarette type. I find that demand for menthol cigarettes does not depend more on past consumption and therefore smokers of menthol cigarettes are no more addicted than smokers of non-menthol cigarettes. However, menthol cigarettes compare unfavorably to non-menthol cigarettes on other dimensions of addictive behavior: they are harder to quit successfully and more attractive to first-time users.



Addiction and Alcohol Tax: Evidence from Japanese Beer Industry (with Kohei Hayashida, Masakazu Ishihara, and Makoto Mizuno), Recording link


This paper studies the effects of taxation and regulation on addictive alcohol consumption. Exploiting the changes in tax policies and sales regulation in the Japanese beer market, we first show some descriptive evidence that consumers (i) are addicted to alcohol, (ii) are forward-looking and stockpile, but potentially present-biased, and (iii) substitute across categories in response to policy changes. To quantify the impacts of policy changes, we then estimate a dynamic structural model of alcohol purchase and consumption where consumers can be present-biased. A series of counterfactual simulations show that the current Japanese alcohol tax system is suboptimal in that alternative policies can increase tax revenues while keeping alcohol addiction lower. Finally, we derive the optimal alcohol tax policy, taking both externalities and internalities into account.



On the Design of Food Labeling Policies (with Cristóbal Otero, Sebastián Otero, and Joshua Kim), Recording link


We study a regulation in Chile that mandates front-of-package warning labels on products whose sugar or caloric concentration exceeds certain thresholds. We find that after the introduction of the regulation, consumers reduced their overall sugar and caloric intake by 9% and 6%, respectively. This change is explained by consumers buying healthier products and firms reformulating their products. On the demand side, labels induce consumers to substitute within categories rather than between categories. On the supply side, we document bunching at regulatory thresholds, with substantial heterogeneity across categories. We provide insights to inform the design of effective food labeling policies.



Simultaneous versus sequential: product release strategies for trial design in online education (with Paulo Albuquerque) 


We study the impact of content release strategies on product usage and trial conversion. We examine two strategies for releasing product features during trial periods: simultaneously at the first interaction with the product, or sequentially over the length of the trial period. While introducing multiple features simultaneously may encourage consumer exploration and increase utility at the start of the trial, delaying the release of some features may increase engagement in later usage sessions when novel content is released. Using data from a randomized control trial on an online educational platform, we find that 1) compared to simultaneous release, sequential release of features leads to higher usage, lower attrition, and eventually higher subscription rates; 2) enjoyment of novel content throughout the trial is an important driver of the observed effect; 3) retention dynamics of learners vary due to their heterogenous preference for different types of content. To predict how customized sequences of features release would impact usage and conversion by different groups of learners, we build a model of usage decisions that accounts for product feature satiation. 


Managing Information Disclosure: An Empirical Analysis of a Search Advertising Market with Non-Strategic Behaviors (with Mingxi Zhu)

Bidding in search advertising is commonplace today. However, determining a bid can be challenging in light of the complexity of the auction process. By aggregating the information of many bidders and managing information disclosure, the advertiser platform can assist less sophisticated advertisers. We analyze the data from a platform that initiated a bid recommendation system based on historical data. We find that less sophisticated advertisers simply adopt the platform’s suggestions instead of constructing their own bids. We characterize an equilibrium model of bidding in the Generalized Second Price (GSP) auction and show that following the platform’s bid suggestion is theoretically sub-optimal. We construct a novel structural model of bidding in GSP to identify advertisers’ private values using observed bids. Counterfactual results suggest that the market suffers from such sub-optimal behaviors. The ad platform can increase revenue and the total surplus by sharing bid recommendations based on values rather than historical bids. These results shed light on the importance of carefully managing information disclosure algorithms in the search advertising market, where non-strategic and strategic bidders interact.

Time Allocation and Multi-Category Search (joint with Yufeng Huang and Ilya Morozov)

Using detailed clickstream data, we study how consumers allocate time to different retailers and product categories when shopping online. We find that most consumers search in multiple categories in a given week, but within each category they allocate all shopping time to one retailer. When the opportunity cost of time decreases, consumers expand their search to more categories and examine more products per category. These findings imply that the opportunity cost of time affects how consumers search online, as conjectured by Stigler (1961). To formalize this idea, we develop a model in which consumers optimally allocate their limited time to searching in different product categories. In this model, search costs arise endogenously because time spent on searching in a given category could be otherwise spent on searching in other categories or on non-shopping activities. Estimating the model, we show that reducing search frictions may have unequal effects on consumers who face different time constraints. We also show that facilitating search in one product category might have unexpected spillover effects on search and choice in other, seemingly unrelated categories.


Valuing Solar Subsidies (with Kenneth Gillingham A. Justin Kirkpatrick), recording link


Individuals trade future for present consumption across a range of economic behaviors, and this tradeoff may differ across demographics. This study employs unique data on rooftop solar adoption and the expected returns from such adoption to estimate heterogeneous discount rates by wealth. We develop a dynamic discrete choice model of optimal system sizing and adoption, and base identification on plausibly exogenous variation in the future savings from installing solar and electricity rates. We estimate implied discount rates of 19.8%, 10.3%, and 10.8% for low-, medium-, and high-wealth households in California. Counterfactual simulations demonstrate opportunities to reduce the regressivity of solar adoption and improve policy cost-effectiveness.



Voluntary Provision of Sustainability Claims: Evidence from Consumer Packaged Goods (with Kristina Brecko), recording link


We study firms' voluntary provision of sustainability claims in the consumer packaged goods (CPG) market using detailed information on product labels. Analyzing product-level information and sales data across 60 product markets in the health and beauty care, detergents, and household cleaners categories from 2012 to 2020, we find that firms' decisions to provide sustainability claims are systematically correlated with the nature of competition within a product market, while not much correlated with the market's average demand for sustainability claims. The pattern is robust even within a single brand that covers multiple product markets, suggesting that brands make strategic choices when providing claims based on market-specific competition.  Our findings have implications on market interventions for sustainability including anti-greenwashing regulations.  


tipping by 40% for new buyers and 10% for repeat buyers, relative to the status quo message. The effect was largest upon first exposure for satisfied buyers. The two reciprocity treatments did not significantly change tipping behavior. The Norms-driven tipping increase did not significantly change buyers’ subsequent platform usage, seller pricing or seller effort. Two post-experiment platform design changes show that reverting to the status-quo message causally reduced tipping by Norms-treated buyers, and then platform-wide adoption of the Norms message corresponded with meaningfully higher tipping rates. Collectively, the results indicate the importance of digital platform design in establishing market norms, the greater impressionability of new platform users, and also the inherent uncertainty in measuring platform design effects on user populations.



Title: Taxes, Passthrough & Market Structure



*** Note the late start time 12:15pm ET ***


Title: Does Model Understanding Improve Human Decision Making?, Recording


Abstract: As machine learning (ML) models are increasingly being employed to make consequential decisions in high-stakes settings such as finance, healthcare, and hiring, it becomes important to ensure that these models are actually beneficial to human decision makers. To this end, recent research in ML has focused on developing techniques which aim to explain complex models to domain experts/decision makers so that they can determine if, when, and how much to rely on the predictions of these models. In this talk, I will give a brief overview of the state-of-the-art in explaining ML models, and then present some of our recent research on understanding the impact of explaining the rationale behind model predictions to decision makers. More specifically, I will discuss two user studies that we carried out with domain experts in healthcare (e.g., doctors) and hiring (e.g., recruiters) settings where we analyzed the impact of explaining the rationale behind model predictions on the accuracy and the discriminatory biases in the decision making process.



Title: Unstructured Data, Econometric Models, and Estimation Bias (with Nikhil Malik), Recording


Abstract: This article examines the powerful combination of machine learning and econometric models to examine unstructured data. Researchers estimate an econometric model (e.g., logit regression, structural model) that relates an outcome of interest (e.g., sales) to a focal feature in unstructured data (e.g., presence of pets in images), with the feature extracted using machine learning algorithms. We focus on potential estimation bias due to prediction errors by the machine learning algorithm. Unfolding the causes of bias, we point out important differences from classical measurement errors. Particularly, bias of either direction is possible. We derive a strategy to alleviate the bias, under the typical setting that the feature is correctly labeled in a fraction of the sample. The strategy extends and improves the few pioneering works in this area, by covering general nonlinear econometric models and relaxing the assumption that unstructured data affect outcome only via the focal feature.



*** midterm elections seminar ***


Title: Small Campaign Donors (with Laurent Bouton, Edgard Dewitte, and Vincent Pons), Recording


Abstract: We study the characteristics and behavior of small campaign donors and compare them to large donors by building a dataset including all the 340 million individual contributions reported to the U.S. Federal Election Commission between 2005 and 2020. Thanks to the reporting requirements of online fundraising platforms first used by Democrats (ActBlue) and now Republicans (WinRed), we observe contribution-level information on the vast majority of small donations. We first show that the number of small donors (donors who do not give more than $200 to any committee during a two-year electoral cycle) and their total contributions have been growing rapidly. Second, small donors include more women and more ethnic minorities than large donors, but their geographical distribution does not differ much. Third, using a saturated fixed effects model, we find that race closeness, candidate ideological extremeness, whether candidates and donors live in the same district or state, and whether they have the same ethnicity increase contributions, with lower effects for small donors. Finally, we show that campaign TV ads affect the number and size of contributions to congressional candidates, particularly for small donors, indicating that pull factors are relevant to explain their behavior.



Title: The Economics of Recommender Systems: Evidence from a Field Experiment on MovieLens (with Duarte Goncalves, Ruoyan Kong, Daniel Kluver, Joseph Konstan), Recording


Abstract:

We conduct a field experiment on a movie recommendation platform to identify if and how recommendations affect consumption. Using within-subject randomization at the item level, we find recommendations significantly increase consumption beyond mere exposure. We test and provide support for an informational mechanism: recommendations affect consumers' beliefs, which in turn explains consumption. Recommendations reduce uncertainty about goods that consumers are most uncertain about and induce information acquisition. Using rich data we collect, we show beliefs about goods are spatially correlated and demonstrate there are subsequent informational spillovers from good consumption, highlighting the importance of dynamic considerations in evaluating recommender systems.



Title: Bargaining and International Reference Pricing in the Pharmaceutical Industry (with Pierre Dubois and Ashvin Gandhi)


Abstract: The United States spends twice as much per person on pharmaceuticals as European countries, in large part because prices are much higher in the US. This fact has led policymakers to consider legislation for price controls. This paper assesses the effects of a US international reference pricing policy that would cap prices in US markets by those offered in reference countries. We estimate a structural model of demand and supply for pharmaceuticals in the US and reference countries like Canada where prices are set through a negotiation process between pharmaceutical companies and the government. We then simulate the counterfactual equilibrium under such international reference pricing rules, allowing firms to internalize the cross-country externalities introduced by these policies. We find that in general, these policies would result in much smaller price decreases in the US than price increases in reference countries. The magnitude of these effects depends on the number, size and market structure of references countries. We compare these policies with a direct bargaining on prices in the US.



Title: Algorithmic pricing and consumer sensitivity to price variability (with Diego Aparicio and Madhav Kumar), Recording


Abstract:

Algorithmic pricing can be broadly defined as a formula to set prices by a computer. It is typically associated with a lower cost of changing prices and a greater frequency of price changes. While commonly observed in ride-sharing, lodging, and airline tickets, there has been recent evidence of its implementation in pharmaceutical drugs, gasoline, and online retail. However, little is known about how consumers respond to encountering frequently changing prices for goods, which algorithmic pricing can increase. Here we use detailed clickstream data from an online retailer that varied pricing methods to examine how exposure to the frequently-changing prices feature of algorithmic pricing affects purchase behavior — particularly price sensitivity. We have evidence from multiple identification strategies and additional lab experiments that exposure to more price changes increases price sensitivity. This talk will focus on analyses exploiting haphazard timing of user visits, whereby if users viewed products at slightly different times they would have seen different individual prices and different numbers of price changes. This will facilitate drawing connections to other related methodological work.



Abstract:
We develop a flexible yet tractable model of consumer search and choice, and apply it to the problem of product rankings optimization by online retail platforms. In the model, products are characterized by a search index, which governs what consumers search; and a utility index, which governs which of the searched options is purchased. We show that this framework can be microfounded based on the Weitzman (1979) model and, in its most general form,  subsumes several other commonly used search models. We then consider how platforms should assign products to search ranks. To optimize consumer surplus, platforms should facilitate product discovery by promoting “diamonds in the rough,” products whose utility index exceeds their search index. By contrast, to maximize static revenues, the platform should favor high-margin products, creating a tension between the two objectives. We develop computationally tractable algorithms for estimating consumer preferences and optimizing rankings, and we provide approximate optimality guarantees in the latter case. When we apply our approach to data from Expedia, our suggested ranking achieves both higher consumer surplus and higher revenues than is achieved by the Expedia ranking, and also dominates ranking the products in order of utility.


Title: A Neuro-Autopilot Theory of Habit: Evidence from Canned Tuna

Abstract:
In economics and marketing, habits are typically modeled as a preference complementarity, whereby a consumer’s utility for a good increases after consuming it. While this approach captures persistence in consumption over time, it ignores the primary benefit of a habit as conceptualized in psychology and neuroscience --- that seemingly complex behaviours can be automated at little cognitive cost. Here, we integrate a neuroeconomic concept of habit into a structural consumer choice model. We propose that habit represents one of two distinct decision-making modes: a habit mode which automatically repeats past choices and a ``model-based" mode of decision-making in which consumer maximize a random utility. The transition between these decision modes is governed by the reliability of a reinforcement learning algorithm, such that habits arise when the consumption environment is stable and predictable. We estimate and test this model on product choice in the canned tuna product category between 2006 and 2009, a period which underwent considerable price and product variation. Our results suggest that a considerable proportion of choice persistence is due to a habitual automaticity in consumption, in addition to a degree of state-dependent utility.

Title: Algorithmic collusion: Implications of competing multi-product firms using single product pricing algorithms 


Abstract: 

We revisit the nascent literature on algorithmic collusion (Calvano et al, Hansen et al) which considers settings where single-product firms compete by setting prices via algorithm, and establishes that supra-competitive prices may arise in such settings. Our key point of departure is that we consider multi-product firms. We show evidence that despite selling multiple products, in practice, firms often price each item via independent algorithms to mitigate the curse of dimensionality. In other words the algorithms in use optimize each product individually rather than jointly optimizing over the entire product assortment. We show that in such settings, the risk of supra-competitive outcomes is reduced and can even result in sub-competitive prices. Conversely, we show that if firms were able to solve the dimensionality and use algorithms that priced jointly, this may increase the mechanisms by which collusive prices are reached, including multi-market contact. 


Gender-Based Pricing in Consumer Packaged Goods: A Pink Tax? with Sarah Moshary (UChicago – Booth) and Natasha Bhatia (Cornerstone Research)


This paper studies a controversial application of a textbook pricing practice: gender-based price segmentation in CPG, which allegedly has created a pink tax whereby products targeted at women are more expensive than their counterparts marketed toward men. This paper sheds light on the form and magnitude of gender-based pricing for personal care products. We first find that gender segmentation is ubiquitous, as more than 80% of products sold are gendered, and segmentation involves product differentiation; there is little overlap in the ingredients of men’s and women’s products made by the same manufacturer within the same category. Using a national dataset of grocery, convenience, drugstore, and mass merchandiser sales, we show that this differentiation sustains meaningful price differences for men’s and women’s products made by the same manufacturer. However, in an apples-to-apples comparison of women’s and men’s products with similar ingredients, we do not find evidence of a systematic price premium for women’s goods: the women’s variant is less expensive in three out of five categories and what price differences do exist are economically small. Our findings are consistent with the ease of arbitrage in posted price markets where CPG are sold. These results call into question the need for and efficacy of recently proposed and enacted legislation, which mandates price parity across substantially similar gendered products.



Efficient and targeted COVID-19 border testing via reinforcement learning 

Throughout the COVID-19 pandemic, countries relied on a variety of ad-hoc border control protocols to allow for non-essential travel while safeguarding public health: from quarantining all travellers to restricting entry from select nations based on population-level epidemiological metrics such as cases, deaths or testing positivity rates. Here we report the design and performance of a reinforcement learning system, nicknamed ‘Eva’. In the summer of 2020, Eva was deployed across all Greek borders to limit the influx of asymptomatic travellers infected with SARS-CoV-2, and to inform border policies through real-time estimates of COVID-19 prevalence. In contrast to country-wide protocols, Eva allocated Greece’s limited testing resources based upon incoming travellers’ demographic information and testing results from previous travellers. By comparing Eva’s performance against modelled counterfactual scenarios, we show that Eva identified 1.85 times as many asymptomatic, infected travellers as random surveillance testing, with up to 2-4 times as many during peak travel, and 1.25-1.45 times as many asymptomatic, infected travellers as testing policies that only utilize epidemiological metrics. We demonstrate that this latter benefit arises, at least partially, because population-level epidemiological metrics had limited predictive value for the actual prevalence of SARS-CoV-2 among asymptomatic travellers and exhibited strong country-specific idiosyncrasies in the summer of 2020. Our results raise serious concerns on the effectiveness of country-agnostic internationally proposed border control policies that are based on population-level epidemiological metrics. Instead, our work represents a successful example of the potential of reinforcement learning and real-time data for safeguarding public health.



Non-parametric Estimation of Habitual Brand Loyalty (with Xinyao Kong and Oeystein Daljord)

Brand loyalty constitutes a key sources of intangible marketing capital as a moderator of long-tern ROI from marketing expenditures. We test for and measure the state-dependent effect of habitual brand loyalty, a classic form of brand loyalty studied in quantitative marketing that creates psychological switching costs. To test for habitual brand loyalty, we use a nonparametric ``dynamic potential outcomes'' approach that circumvents the classic identification challenge associated with the decoupling of state dependence and unobserved heterogeneity. We then propose a semi-parametric test for forward-looking behavior to assess whether consumers plan their future loyalty. Through several case studies of consumer packaged goods categories, we implement these tests and compare our results to the estimates from a parametric dynamic discrete-choice model of demand. We find non-parametric evidence for HBL and semi-parametric evidence for rational HBL. Counter-factual simulations indicate that the long-run price-elasticity of demand is considerably higher when consumers plan their future brand habits.



The Economic Value of Norm Conformity and Menu Opt-Out Costs, Slides

This paper theoretically and empirically analyzes trade-offs between consumption versus norm-adherence and choosing from a menu of default options versus computing a non-default choice. In the theoretical model, peoples' choices depend on consumption, norm conformity, and menu-opt-out costs. Using passengers' tips sampled from a billion NYC taxi rides, I empirically estimate the model parameters. I find that the cost of deviating from the norm tip and opting out of the default tip menu are both high relative to the taxi fare. I then examine the welfare implications of norm conformity and the positive and normative effects of default menu design.



Title: NFT Marketplace Design and Market Intelligence

Nonfungible tokens (NFTs) have exploded in popularity in 2021, generating billions of dollars in transaction volume. In tandem, market intelligence platforms have emerged to track summary statistics about pricing and sales activity across different NFT collections. We demonstrate that marketplace design can significantly influence market intelligence, focusing specifically on the costs of bidding which can differ across marketplaces depending on transaction fees, the prevalence of bidding bots, or the user interface for placing bids. We use data from the CryptoPunks marketplace and build an empirical model of the strategic interaction between sellers and bidders. Counterfactual simulations show that a reduction in bidding costs does not change the quantity of sales, but increases the share of sales that result from bids. Listing prices increase as sellers expect to accept more bids, making assets appear more valuable. The listing and realized sale price ratios between rare and common assets shrink, making the market appear more homogeneous. Collections that are offered by two different marketplaces can exhibit significantly different market statistics because of differences in bidding costs rather than differences in inherent value. The results have implications for the interpretation of NFT market intelligence.


Title: A Neural-Autopilot Analysis of Social Media Engagement


This paper describes and estimates a novel ``neural autopilot" model of habit formation using a large sample of individual-level daily data from Chinese social media postings around the 2020 lockdown period. The model produces interpretable parameter estimates consistent with neuropsychological understanding of habit and captures the time series features of the actual posting frequencies. Our counterfactual results suggest that reducing reward volatility increases habitual posting on social media platforms, due to the importance of reward reliability---not just reward---in creating habit. However, forced experimentation has minimal impact on habituation in the long run.


The Identity Fragmentation Bias (with Sanjog Misra)


Consumers interact with firms across multiple devices, browsers, and machines; these interactions are often recorded with different identifiers for the same consumer. The failure to correctly match different identities leads to a fragmented view of exposures and behaviors. This paper studies the identity fragmentation bias, referring to the estimation bias resulted from using fragmented data. Using a formal framework, we decompose the contributing factors of the estimation bias caused by data fragmentation and discuss the direction of bias. Contrary to conventional wisdom, this bias cannot be signed or bounded under standard assumptions. Instead, upward biases and sign reversals can occur even in experimental settings. We then compare several corrective measures, and discuss their respective advantages and caveats.


Buy Baits and Consumer Sophistication: Theory and Field Evidence from Large-Scale Rebate Promotions


Can firms exploit behavioral biases to increase profits? Does consumer sophistication about these biases limit the scope of exploitation? To answer these questions, I run a series of natural field experiments with over 600,000 consumers and estimate novel sufficient statistics of consumer sophistication. The empirical application is a ubiquitous and widely regulated form of price discrimination: rebates that need to be actively claimed by consumers. These promotions are suspected of boosting sales even though many consumers eventually fail to claim the rebate—a phenomenon marketers refer to as “slippage.” I show theoretically that consumers’ subjective redemption probabilities can be inferred from how demand responds to rebates as opposed to simple price reductions. I identify these elasticities in three natural field experiments with a major online retailer, in which I randomize prices, redemption requirements, and reminders. Results reveal that claimable rebates in fact increase sales substantially even though 47% of consumers do not redeem the rebate. However, consumers exhibit a remarkable degree of sophistication: the demand response to a rebate is only 76% of the demand response to an equivalent price reduction. Structural estimates imply that consumers are almost perfectly aware of their inattention but vastly underestimate the hassle of redemption by 20 EUR per consumer. Exploiting this misperception increases the profitability of rebates by up to 260%.


Going Backward to Move Forward? The Effects of Backward Compatibility on the Sales of Previous and New Generation Console Video Games (with Venky Shankar)


In several product categories, such as electronics, video games, computer hardware and software, and other hi-tech products, backward compatibility–the property of a current generation of hardware to allow previous generation of software or accessory to work with it–is an important strategic decision for firms introducing hardware upgrades. We empirically investigate the effect of Microsoft Xbox’s decision to make its new generation console (NGC, Xbox One) backward compatible with selected games for its previous generation console (PGC, Xbox 360) on the sales of video games for both PGC and NGC. We assemble a unique dataset containing aggregate and individual data from a large proprietary game retailer and data scraped from gaming websites during 2013-2017. We analyze the effects of backward compatibility using a difference-in-differences approach and appropriate synthetic control methods. Our results show that when a video game console firm makes its NGC compatible with some PGC games, the average unit sales of backward compatible PGC games decrease relative to non-backward compatible PGC games. However, the dollar sales of backward compatible PGC games increase relative to non-backward compatible PGC games due to a relative price increase effect. The effects are heterogeneous across games; unlike low-selling games, high-selling games experience disproportionately large revenue lifts. Interestingly, the results also show that the sales (units and dollars) of NGC games increase due to a spillover effect, driven by console upgrades and enhanced budget for new games for the previous owners of backward compatible games. Based on our results, managers should consider high-selling games, games for general audience, action games, and those with high user ratings as the primary candidates for backward compatibility.


Designing Dealer Compensation in the Auto Loan Market: Implications from a Policy Change 

(with Yanaho 'Max' Wei, Tat Chan, and Naser Hamdi)


We study dealer compensation in the indirect auto lending market, where most lenders give dealers the discretion to mark up interest rates and the markup constitutes a dealer's compensation. To protect consumers from potential discrimination by this dealer discretion, several banks adopted a policy that removes dealer discretion and compensates dealers by a fixed percentage of the loan amount. We document that this policy decreased (increased) the interest rates for low-credit (high-credit) consumers. However, the market share of these banks decreased (increased) for low-credit (high-credit) consumers — a reversal of the usual demand curve. This reversal highlights the influence of dealers on consumers when choosing which bank to finance a loan. Accordingly, we develop an empirical model that features dealer–consumer bargaining. Our estimation results show discrimination without the policy, as reflected in the systematically lower bargaining power of certain consumer groups. We use the model to explore alternative compensation schemes that remove dealer discretion. We find that a lump-sum compensation scheme obtains the most market share. In particular, this holds true for low-credit consumers, where the lump-sum scheme allows more of them to access lower interest rates. Our study highlights the importance of accounting for the incentives and bargaining power of middlemen.


Choice Frictions in Large Assortments  


This paper studies how the growth and evolution of product assortments impact consumer adoption, churn, and long run consumption. Most economic theories of product variety and the value of platforms suggest consumers at least weakly prefer larger product assortments. In contrast, the psychological literature on the phenomenon of choice overload finds that larger assortments overwhelm consumers with decision costs or induce more regret. I provide empirical evidence of how the size and contents of product assortments impact consumers over their lifetime in an online platform market that provides restaurant delivery. I find that assortment expansion increases the acquisition of new consumers but reduces the frequency of consumption among consumers who remain on the platform. I rationalize these impacts via a model of costly attention and choice under limited information. Counterfactual exercises show that targeting choice set reductions can improve revenue among existing customers by up to 54%.


What drives demand for playlists on Spotify?  (with Hannes Datta)


We study the drivers of playlist demand on Spotify – the leading music streaming service. Our main analysis relies on a unique data set that combines daily data of about 12,000 popular playlists on Spotify with information on how these playlists are featured by music categories (e.g., mood, dinner, pop). We estimate a model of users' playlist choice and decompose three major drivers of playlist demand on Spotify, (i) persistent time-invariant user preferences for playlists, (ii) users' responsiveness to featured playlists in the Spotify app as well as (iii) preferences for the popularity of the playlist content that varies over time depending on compositions and external factors. Our results show strong persistent consumer tastes for Spotify-curated playlists. In contrast, users are less responsive to changes of the content popularity of playlists over time. We discuss implications for the power imbalances and dependencies between content producers and music streaming platforms.


Artificial intelligence for a reduction of false denials in refugee claims

(with Hilary Evans Cameron, and Leah Morris)


Deciding refugee claims is a paradigm case of an inherently uncertain judgment and prediction exercise. Yet refugee status decision-makers may underestimate the uncertainty inherent in their decisions. A feature of recent advances in artificial intelligence (AI) is the ability to make uncertainty visible. By making clear to refugee status decision-makers how uncertain their predictions are, AI and related statistical tools could help to reduce their confidence in their conclusions. Currently, this would only hurt claimants, since many countries around the world have designed their refugee status determination systems using inductive inference which distorts risk assessment. Increasing uncertainty would therefore contribute to mistaken rejections. If, however, international refugee law was to recognize an obligation under the UN Convention to resolve decision-making doubt in the claimant’s favour and use abductive inference, as Evans Cameron has advocated, then by making uncertainty visible, AI could help reduce the number of wrong denied claims.



Organizational Structure and Pricing: Evidence from a Large U.S. Airline 

(with Olivia Natan, Hayden Parsley, Timothy Schwieg, and Kevin Williams)


The availability of large amounts of data and improvements in computational technology have allowed firms to develop sophisticated pricing and allocation systems. However, decision rights within these systems are often allocated across different organizations/divisions within the firm. We study how organizational boundaries affect pricing decisions using comprehensive data provided by a large U.S. airline. Contrary to prevailing theories of the firm, we show that advanced pricing algorithms have multiple biases. These biases can be attributed to the various teams responsible for managing pricing algorithm inputs. We quantify the impacts of these biases by estimating a structural demand model that combines sales and search information. We recover the demand curves the firm believes it faces using detailed forecasting data. In counterfactuals, we show that correcting biases introduced by teams individually have little impact on market outcomes, but addressing all biases simultaneously leads to higher prices and increased dead-weight loss in the markets studied. Our results suggest that decentralized decision making can curtail a firm's ability to set optimal prices.


Search Gaps (with Raluca Ursu and Qianyun Zhang)


In the canonical sequential search model, consumers inspect options consecutively until they decide to stop searching, a decision which occurs only once before consumers determine whether and what to purchase. However, using data on consumers’ online browsing histories, we document that consumers frequently take breaks during their search (“search gaps”), that is, they obtain information on a number of options, pause, and later resume their search. Further, we provide model-free evidence that consumers take breaks from searching due to fatigue. To describe search processes that include gaps due to fatigue, we extend the Weitzman (1979) framework and develop a sequential search model that rationalizes search gaps by allowing consumers to additionally decide when to search an option: now or after a break. Fatigue enters the model through increasing search costs: the more a consumer searches, the higher her search costs per option; taking a break reduces these costs to a baseline and enables the consumer to resume her search at a later time. We estimate the proposed model using our data and quantify the effect of fatigue on consumer search and purchase decisions. We find the effect of fatigue to be larger than that of baseline search costs. Lastly, using counterfactuals, we demonstrate the managerial importance of consumers’ search fatigue.


Debunking Misinformation in Advertising (with Jessica Fong and Anita Rao)


Many brands differentiate themselves by highlighting the absence of certain ingredients (e.g. no GMO) with some going as far as to deceptively claim those ingredients are toxic. Because such claims can spread misinformation among consumers, various interested parties – regulators, the media and competitor brands - aim to debunk such misinformation. However little is known whether a) such misinformation in advertisements alters consumers’ willingness to buy a product and b) if so, whether debunking can effectively revert the effect of misinformation. This paper aims to empirically understand the impact of misinformation and that of debunking in the context of three ingredients in product categories in which misinformation is prevalent: aluminum in deodorants, fluoride in toothpastes, and GMOs in food. We find that an additional exposure to misinformation can further alter consumers’ willingness-to-pay, and debunking from a trustworthy source plays an important role in correcting consumers’ misbeliefs.


Sweet Lemons (with Alain Cohn and Collin Raymond)


Modern economic models of choice assume that people derive utility not only from the realization of outcomes, but also from the anticipation of future events. This can lead people to distort  beliefs optimistically about future events. In this paper, we examine the notion of such motivated beliefs in a natural high-stakes environment―the COVID-19 pandemic. We conducted a series of surveys that elicited people’s beliefs about the risk of getting infected with the coronavirus. Leveraging exogenous variation in the timing of having to return to work after the initial wave of lockdowns, we find that people hold increasingly optimistic beliefs about the risk of catching the virus as the return date approaches. This pattern suggests that people dynamically distort their beliefs in an attempt to reduce their anxiety about an imminent, potentially negative event. This dynamic belief distortion is more pronounced  for risk averse individuals and those who are more likely to get severely ill. We provide a tractable model that accounts for the dynamic and heterogeneous nature of belief distortion.


Adverse Selection and Moral Hazard in a Dynamic Model of Auto Insurance


In this paper we measure risk-related private information and investigate its importance in a setting where individuals are able to modify risk ex-ante through costly effort. In particular, we estimate the impact of moral hazard and adverse section on efficiency of the car insurance market. Beyond measuring the relative efficiency loss, we demonstrate the effectiveness of static contract menus and dynamic contracts in combating frictions caused by both moral hazard and adverse selection. Our analysis is based on a model of endogenous risk production and contract choice. It exploits data from multiple years of contract choices and claims by customers of a major Portuguese auto insurance company.


Mitigating the Cold-start Problem in Reputation Systems: Evidence from a Field Experiment (with Zekun Liu and Weiqing Zhang)


Reputation systems are typically used in markets with asymmetric information, but they can cause the cold-start problem for young sellers who lack historical sales. Exploiting a field experiment on eBay, we show that in the presence of a long-run quality signal, introducing a less history-dependent quality signal mitigates the cold-start problem: it increases demand for high-quality young sellers, incentivizes their quality provision, and increases their chance of obtaining the long-run quality signal. Moreover, it prompts established sellers to re-optimize their effort decision. Therefore, the net impact of introducing a less history-dependent signal on quality provision depends on underlying market fundamentals.


Millennials and the Take-Off of Craft Brands: Preference Formation in the U.S. Beer Industry (With Bart J. Bronnenberg and Jean-Pierre Dube)


We conduct an empirical case study of the U.S. beer industry to analyze the disruptive effects of locally-manufactured, craft brands on market structure, an increasingly common phenomenon in CPG industries typically attributed to the emerging generation of adult Millennial consumers. We document a generational share gap: Millennials buy more craft beer than earlier generations. We test between two competing mechanisms: (i) persistent generational differences in tastes and (ii) differences in past experiences, or, consumption capital. Our test exploits a novel database tracking the geographic differences in the diffusion of craft breweries across the U.S.. Using a structural model of demand with endogenous consumption capital stock formation, we find that heterogeneous consumption capital accounts for 85% of the generational share gap between Millennials and Baby Boomers, with the remainder explained by intrinsic generational differences in preferences. We predict the beer market structure will continue to fragment over the next decade, over-turning a nearly century-old structure dominated by a small number of national brands. The attribution of the share gap to consumption capital shaped through availability on the supply side of the market highlights how barriers to entry, such as regulation and high traditional marketing costs, sustained a concentrated market structure.


The Effects of Influencer Advertising Disclosure Regulations: Evidence from Instagram (with Matthew Michell)


We study the effects of advertising disclosure regulations in social media markets. Theory generates ambiguous predictions about the effects of regulations on the equilibrium amount of advertising content, user engagement and welfare. Using data from a large sample of Instagram influencers in Germany and Spain and a difference-in-differences approach, we empirically evaluate the effects of German disclosure regulations on post content and follower engagement. We measure whether posts include suggested disclosure terms and use text-based approaches (keywords, machine learning) to assess whether a post is sponsored. We show a substantial adoption of disclosure after regulations, but also an increase in sponsored content including undisclosed sponsored content. We also find reductions in engagement, suggesting that followers were likely negatively affected.


Latent Stratification for Advertising Experiments (with Ron Berman), Slides


Abstract:


Advertising incrementality experiments often suffer from noisy responses making precise estimation of the average treatment effect (ATE) and evaluation of ROI difficult. We develop a new estimator of the ATE that improves precision by estimating separate treatment effects for three latent strata -- customers who buy regardless of ad exposure, those who buy only if exposed to ads and those who do not buy regardless. The overall ATE computed by averaging the strata estimates has lower sampling variance than the widely-used difference-in-means ATE estimator. The variance is most reduced when the three strata have substantially different ATEs and are relatively equal in size. Estimating the latent stratified ATE for 5 catalog mailing experiments shows a reduction of 36-57% in the posterior variance of the estimate. We also show that customers who have made a purchase recently and have been responsive to similar advertising in the past are less likely to be in the "do not buy regardless" stratum.  


The Long Tail Effect of Personalized Rankings (with Robert Donnelly and Ayush Kanodia)


Abstract: 


We study to what extent personalization in online retail reduces the concentration of sales and contributes to the long tail effect. Using data from a large-scale randomized experiment conducted by Wayfair, a large online retailer of furniture, we show that personalized product rankings induce users to search and purchase a larger variety of items relative to non-personalized bestseller-based rankings. To study whether users benefit from this shift in demand, we propose a novel empirical framework that estimates heterogeneous users' tastes from both click histories and displayed personalized rankings. The framework combines a standard consideration set model with a latent factorization approach from the computer science literature, modeling user tastes as functions of latent attributes observed by users but not by the researcher. This approach makes our strategy robust to having limited data on item attributes and explicitly recognizes that user click and purchase histories are confounded by the prominence of displayed items. Having estimated the model, we show that personalized rankings increase consumer surplus of the average user by 30%. This effect arises primarily because personalization makes users with niche tastes more likely to discover relevant niche items. We also argue that absent personalization, Wayfair would only be willing to offer 25% of its current product assortment because most niche items would never find their audience.


Chasing Stars: Firms' Strategic Responses to Online Consumer Ratings Slides


Abstract: 


In this paper, I show that a common way that platforms display firms' quality ratings incentivizes firms to strategically take costly short-run actions that improve their ratings. Most review platforms display star ratings of goods and services rounded to a half star, rather than display the exact average rating. Since the true average rating is not shown, firms have an incentive to remain just above the rounding threshold in order to have a higher displayed rating. However, once a firm's rating passes the rounding threshold, the incentive to improve the ratings drops as their rating moves farther from the threshold. I study this phenomena in the context of auto repair. I find that there is an excessive amount of bunching around ratings thresholds. The firms' actions toward improving their ratings are typically unobserved, but due to my novel data and the discontinuity of displayed ratings, I can model and infer firm behavior. Specifically, I provide evidence that firms change the services they provide and exert extra effort when they are close to rounding thresholds. Finally, I provide a theoretical framework in order to quantify the actions and provide optimal policies for firm actions depending on their rating and number of reviews in a variety of counterfactual settings.


Naive Analytics (with Yuval Heller)

Slides


Abstract:

Are firms better off with more accurate analytics? We study interactions with uncertainty about demand sensitivity, where (1) firms choose seemingly-optimal strategies given the level of sophistication of their data analytics, and (2) the levels of analytics sophistication form best responses to one another. Under the ensuing equilibrium firms underestimate price elasticities and overestimate advertising effectiveness, which is often what we observe empirically. These misestimates cause firms to set prices too high and to over-advertise. In games with strategic complements (substitutes), profits Pareto dominate (are dominated by) those of the Nash equilibrium. Applying the model to team production games also explains the prevalence of overconfidence among entrepreneurs and salespeople.


"Form + Function: Real-Time Aesthetic Product Design via Geometrized Bi-Level Queries" (with Max Yi Ren, Namwoo Kang, and Panos Papalambros)


Video, Slides


Quantitative marketing has, over the last 50 years, refined methodologies for zeroing in on individuals' preferences using adaptive, heterogeneous discrete choice tasks. Yet visual design elements, due to their high‐dimensional, holistic, and interactive nature, are notoriously difficult to capture. 


We incorporate real‐time, interactive, 3D‐rendered configurations into such measurement frameworks, using rapid, scalable machine learning algorithms to adaptively query respondents and update visual designs. At the heart of the method is a parametric decomposition of an object's geometry, along with a novel, adaptive “bi‐level” query task that can estimate individuals’ preferences among visual designs.


We illustrate the method’s performance through simulation and a field test for the design of a mid‐priced sedan, using real‐time 3D rendering and an online panel. Training via ranking SVM and HB mixtures for radial basis function elements allow the analyst to (1) elicit trade-offs between design and more traditional elements (e.g., price or MPG); (2) pinpoint which design details differentially drive (heterogeneous) preferences; and (3) determine which set of designs serve the respondent base best. 


[The method itself can be previewed at vehiclechoicemodel.appspot.com.]


Title: "The Market for Fake Reviews" (with Sherry He and Davide Proserpio)


Abstract:

We study the market for fake product reviews on Amazon.com. These reviews are purchased in large private internet groups on Facebook and other sites. We hand-collect data on these markets to characterize the types of products that buy fake reviews and then collect large amounts of data on the ratings and reviews posted on Amazon for these products, as well as their sales rank, advertising, and pricing behavior. We use this data to assess the costs and benefits of fake reviews to sellers and evaluate the degree to which they harm consumers. The theoretical literature on review fraud shows that there exist conditions when they harm consumers and conditions where they function as simply another type of advertising. Using detailed data on product outcomes before and after they buy fake reviews we can directly determine if these are low-quality products using fake reviews to deceive and harm consumers or if they are possibly high-quality products who solicit reviews to establish reputations. We find that a wide array of products purchase fake reviews including products with many reviews and high average ratings. Soliciting fake reviews on Facebook leads to a significant increase in average rating and sales rank but the effect disappears after roughly one month. After firms stop buying fake reviews their average ratings fall significantly and the share of one-star reviews increases significantly, indicating fake reviews are mostly used by low quality products and are deceiving and harming consumers. We also observe that Amazon deletes large numbers of reviews and we document their deletion policy. 


Title: “Estimating the Causal Effect of A Digitally Native Retailer Opening a New Store: A New Two-Step Synthetic Control Method” (with Venkatesh Shankar)

With the rapid growth of omnichannel retailing, digitally native retailers are increasingly opening physical stores. A critical issue for many digitally native retailers is to estimate the causal effect of a new store opening on their online sales. To assess the causal effect, a randomized control field experiment is infeasible, so quasi-experiments offer the best hope. Often, due to the non-availability of a readily matched control group, the use of synthetic control (SC) groups to estimate the causal effect is becoming popular. A crucial identifying assumption for the SC method is the parallel trends assumption, which states that the treatment unit would have followed a path parallel to the synthetic control group unit in the absence of treatment. However, this assumption may not hold in real data, in particular, in the omnichannel context. If this assumption is violated, current methods may yield incorrect and misleading estimates of causal effects. Unfortunately, no formal test of this assumption exists. We propose a new two-step synthetic control (TSSC) method that comprises a new test for the parallel trends assumption in the first step, and the application of an appropriate synthetic control method in the second step. Thus, our approach unifies the synthetic control and the modified synthetic control (MSC) methods. We examine the finite sample performance of our testing procedure using simulation. We apply this method to estimate the cross-channel effect of a digitally native retailer opening a physical showroom on its sales at two locations: Columbus, OH and Austin, TX. We demonstrate the value of our TSSC method by revealing that the cross-channel effect from the TSSC method for Columbus is positive and significant, contrary to the incorrect and misleading result from the traditional SC method that shows a negative and significant effect.


Title: TV Advertising and Online Sales: The Role of Inter-Temporal Substitution (with Anja Lambrecht and Catherine Tucker)

Digital technologies lead consumers to instantaneously engage with companies online following TV advertising. As a result, companies increasingly aim to coordinate TV advertising with consumer online behavior. This has led companies and researchers to examine the effect of TV ads on online browsing and sales. As firms tend to show ads when consumers are most likely to respond, prior research has typically focused on a tight time window around TV ad exposure in order to identify a casual effect on online behavior. The downside of such an approach, however, is that it abstracts away from potential shifts in browsing or sales over time. In this paper, we use data from a field test by an online travel agent. In the test, the company ran TV advertising in one region of the country while shutting off TV advertising for the remainder of the country. This allows the untreated region to serve as a control group when analyzing online browsing and sales in the treated region. We find that TV advertising indeed leads to an instantaneous increase in online browsing and sales. However, we document that this instantaneous increase in online browsing sessions and sales comes at the cost of inter-temporal substitution leading to lower browsing and lower sales at times when no ad is aired. We document though a positive side to TV advertising which mitigates somewhat the negative effects of inter-temporal substitution. TV advertising does appear to manage to attract consumers without having to resort to price promotions - potentially leading to higher prices being paid on average by consumers.


Title: Leveraging the Power of Images in Managing Product Return Rates (with Siham El Kihal, John R. Hauser, and Marat Ibragimov)

In online channels, products are returned at high rates. Shipping, processing, and refurbishing are so costly that a retailer's profit is extremely sensitive to return rates. In many product categories, such as the $500 billion fashion industry, direct experiments are not feasible because the fashion season is over before sufficient data are observed. We show that predicting return rates prior to product launch enhances profit substantially. Using data from a large European retailer (over 1.5 million transactions for about 4,500 fashion items), we demonstrate that machine-learning methods applied to product images enhance predictive ability relative to the retailer’s benchmark (category, seasonality, price, and color labels). Custom image-processing features (RGB color histograms, Gabor filters) capture color and patterns to improve predictions, but deep-learning features improve predictions significantly more. Deep learning appears to capture color-pattern-shape and other intangibles associated with high return rates for apparel. We derive an optimal policy for launch decisions that takes prediction uncertainty into account. The optimal deep-learning-based policy improves profits, achieving 40% of the improvement that would be achievable with perfect information. We show that the retailer could further enhance predictive ability and profits if it could observe the discrepancy in online and offline sales.


Title: Clickbait

Abstract: Online media consumption patterns are constructed through thousands of micro-choices: clicks. The dynamic of competition between media firms seeking clicks, platforms aiming to enhance user experience by avoiding deception and regret, and consumer evolution in both savvy and taste has produced a rapid evolution in the form taken by what I call Clickbait Media. I develop a theory of the economics of Clickbait Media, premised on social proof as an alternative form of media credibility. This ecosystem has produced an alternative media aesthetic that itself serves as a powerful signal, as I demonstrate with an incentivized media choice experiment. I conclude with a discussion of the evolution of clickbait using the Upworthy Research Archive to provide insights into the organizational behavior of a pioneering Clickbait Media firm.


Title: Flexible Estimation of Discrete Choice Models Using Aggregate Data with Amit Gandhi (U. Penn) and Jing Tao (U. of Washington)

Abstract: We propose a method to estimate a flexible distribution of consumer heterogeneity in the well-known BLP model.  The key insight is to decouple the estimation of the demand function from the estimation of the distribution of consumer heterogeneity. Specifically, we first estimate the structural inverse demand function, without specifying an underlying model of consumer heterogeneity, by employing flexible functional forms and using linear IV techniques (both parametric and non-parametric). Linearity allows us to take advantage of recent advances in the machine learning of instrumental variables to automate the construction of strong instruments from exogenous variables and can be easily utilized in a data rich environment. We can recover many parameters, such as demand elasticities from this first stage. We can also use the results of this first stage to recover the distribution of heterogeneity as a standard mixture problem, which implies that a variety of flexible procedures can be utilized. We focus on fixed grid methods. The resulting approach is faster and more flexible than the standard parametric approach.  We demonstrate its flexibility and speed in a series of Monte Carlo simulations and in an application using scanner data.


Title: Dynamic Coupon Targeting Using Batch Deep Reinforcement Learning: An Application to LiveStream Shopping

We present an empirical framework to create dynamic coupon targeting strategies using a batch deep reinforcement learning (BDRL) algorithm and apply it in a novel, multi-billion-dollar livestream shopping context. Prior solutions to the dynamic coupon targeting problem either ignore long term demand implications or suffer from model bias. Our BDRL algorithm has four comparative advantages over existing strategies. First, it can capture consumers' intertemporal tradeoffs associated with dynamic pricing, especially the reference price effect. Second, it is not prone to the model bias in dynamic structural models, because it is based on Q-learning, a model-free reinforcement learning solution. Third, it alleviates the curse of dimensionality problem by leveraging deep neural networks to represent the high-dimensional state space. Fourth, it requires only historical data rather than live experimentation, as both policy learning and policy evaluation operate in the batch mode. Using both a field experiment and an off-policy evaluation method, we show that our solution increases the livestream shopping platform's revenue by 60%, twice as effective as static targeting policies. The dynamic targeting strategy recommends increasing the coupon discount level over time because of the reference price effect, and the rate of increase is faster for low spenders.


Title: Aiming for the Goal: Contribution Dynamics of Crowdfunding (with Joyee Deb and Kevin Williams )

We study reward-based crowdfunding, a new class of dynamic contribution games where a private good is produced only if the funding goal is reached by a deadline. Buyers face a coordination rather than a free-riding problem. A long-lived donor may alleviate this coordination risk, signaling his wealth through dynamic contributions. We characterize platform-, donor-, and buyer-optimal outcomes which are attained by stationary equilibria with simple donation strategies. We test the model's prediction using high-frequency data collected from the largest crowdfunding website, Kickstarter. The model fits the data well, especially for predictions concerning comparative statistics, donor dynamics, and properties of successful campaign.


Monday, October 26, Noon ET - Kristina Brecko (U. Rochester)

Place-Making or Consumer-Making: The Role of the Neighborhood in Shaping Consumption Patterns (with Tomomichi Amano)

The past decade has seen renewed interest in neighborhoods' location-specific characteristics such as amenity levels and mix and spatial and social integration. In this paper, we study the underpinnings a key policy goal of this effort: local economic development. We examine whether location-specific goods play a role in shaping individual consumption patterns and by which channel. In a series of empirical exercises, we provide suggestive evidence for three main observations. (1) Location-specific characteristics, unlike demographic characteristics, are not systematically associated with particular consumer product popularity. (2) Consumer goods firms are more responsive to changes in regional demographic characteristics than location-specific goods. (3) Location-specific characteristics play a role in the pass-through of regional consumption trends to the individual household. Taken together, these observations suggest that location-specific goods are more likely to affect consumption patterns through changes in individual consumer shopping patterns than changes in the supply conditions. Our findings have implications for policy-makers as well as marketers looking to hone the targeting strategy of their consumer goods.


Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement (with  Rob Moakler and Florian Zettelmeyer)

Abstract: Advertisers often rely on observational techniques to estimate a campaign’s effectiveness when they are unable to utilize randomized controlled trials (RCTs). This paper compares ad effects obtained using observational techniques to their RCT counterparts for the same set of ad campaigns to determine under which conditions observational approaches deliver accurate estimates of causal advertising effects. The analysis uses a unique data set of over 700 ad campaigns on Facebook with at least one million users, all from the second half of 2019 in the United States. The campaigns represent several hundred distinct advertisers across about 15 industry verticals, with the average campaign having eight million users and a budget exceeding $100K. By making a comparison at this scale of studies, we aim to provide generalizable insights to help advertisers understand the conditions under which observational methods are most reliable. We characterize the results across three types of campaign attributes: (1) those advertisers can change (e.g., targeting rules, campaign budget), (2) those advertisers cannot change (e.g., industry vertical, baseline conversion rate), and (3) non advertiser-specific attributes (e.g., observational model fit, exposure rate). A preliminary analysis indicates that the baseline conversion rate---the probability of conversion in the randomized control group---is an important characteristic that determines the ability of observational methods to recover an accurate causal estimate. We explore additional attributes and disentangle their relative importance to provide additional findings.


Network Structures of Collective Intelligence:  The Contingent Benefits of Group Discussion (with  Abdullah Almaatouq and Agnes Horvat)

Video

Abstract: Accurate numeric estimates such as forecasts are critical to strategic decisions such as market entry as well as operational decisions such as designing a community engagement campaign. Despite the increasing availability of formal predictive models, subjective judgements or "gut" estimates remain an important part of decision-making.  A common strategy to improve accuracy is to use the average of multiple forecasters, also known as the "wisdom of crowds principle," and a common expectation is that groups produce the most accurate estimates when their contributors are statistically independent. However, empirical evidence on whether social influence helps or harms accuracy has produced conflicting results. Some evidence suggests that interacting groups can be accurate, but only if they interact via carefully mediated processes such as the "Delphi method." Other evidence suggests that informal discussion can not only improve estimates, but can even outperform mediated proceses. Still other researchers continue to argue that any social influence will undermine belief accuracy. This talk will present novel experimental data and theoretical results resolving these apparently contradictory results and identifying when social influence improves belief accuracy. The primary finding is that the effect of social influence is task and context dependent. One interesting discussion point is how a group in practice might calibrate their decision process to the task and context at hand.


Automating the B2B Salesperson Pricing Decisions: A Human-Machine Hybrid Approach (with Yael Karlinsky-Shichor)

In a world advancing towards automation, we ask whether salespeople making pricing decisions in a high human interaction environment such as business-to-business (B2B) retail, can be automated, and when it would be most beneficial. Using sales transactions data from a B2B aluminum retailer, we create an automated version of each salesperson, that learns and automatically reapplies the salesperson’s pricing policy. We conduct a field experiment with the B2B retailer, providing salespeople with their own model’s price recommendations in real-time through the retailer’s CRM system, and allowing them to adjust their original pricing accordingly. We find that despite the loss of non-codeable information available to the salesperson but not to the model, providing the model’s price to the salesperson increases profits for treated quotes by 10% relatively to a control condition. Using a counterfactual analysis, we show that while in most of the cases the model’s pricing leads to higher profitability, the salesperson generates higher profits when pricing for quotes or clients with unique or complex characteristics. Accordingly, we propose a machine learning Random Forest hybrid pricing strategy, that automatically combines the model and the human expert and generates profits significantly higher than either the model or the salespeople.


The Role of Consumer Inattention in Perishable Purchases (with Karsten Hansen and Kanishka Misra)

Grocery chains manage perishable products' waste by using strategies such as stock rotation and dynamic pricing. The effectiveness of these strategies depends on consumer preferences for expiry dates. A key empirical challenge is observing how consumers make decisions across expiry dates within a SKU. We study consumer choices for perishable products using the first dataset tracking expiry dates for each unit on the self and consumer choices at the expiry date level. Our data also include a field study where the oldest expiry date items on the shelf received discounts.  We find that consumers are inattentive to expiry dates: when all items are priced identically, the most dominated expiry date items' choice share is 37 points.  Observed choices are consistent with consumers' employing a heuristic of choosing front-most facing items: rotating the oldest stock forward increases the purchase share of the oldest expiry date by approximately 12 points. Importantly, despite this inattention, we find that expiry date based discounts cause consumers to purchase older items. Discounts on items with a long shelf-life cause an immediate share increase on these items by approximately 11 points, with each additional  ten percentage point discount increasing the choice share by approximately 1.6 points. Contrastingly, discounts on items with a short shelf-life lead to no immediate level shift, but have a more price sensitive response of approximately 3.4 points per 10 percentage point discount. Our results are consistent with consumers having diminishing marginal utility of freshness in perishables. We discuss the implications of our findings for waste/shelf management.


Parallel Experimentation in a Competitive Advertising Marketplace (with Xiliang Lin, Harikesh S. Nair, and Navdeep S. Sahni)

The modern digital era represents a golden age of experimentation. Dramatic improvements in the ease of experimentation – such as better tools, data tracking and access to wider and broader audiences – has increased the propensity to experiment on web-enabled platforms such as online marketplaces. As a consequence, it is now common for many firms to be experimenting simultaneously on digital platforms, and for many consumers to be simultaneously in multiple experiments. This raises the question of how the treatment effects that are measured in an environment of such “parallel experimentation” should be interpreted and how the treatment effects measured for one firm may depend on the experimentation policies of others. To do this, we study the interpretation of treatment effects in a world where experiments are running in parallel. We characterize the treatment effects measured as equilibrium constructs that incorporate competition. This analysis shows that the typical experimental estimate – the difference in mean outcomes between treatment and control groups – represents an unstable object of unclear relevance for decision in making in the post-experiment environment. Motivated by this, we present a new set of causal estimands that are more transportable and policy-relevant. We show the causal estimands we suggest are nonparametrically identified by a simple experimental design that we develop for parallel experimentation. We also present an estimator for policy-relevant treatment effects that handle typical, high-dimensional situations. We implement the design and estimator on the advertising platform of JD.com, an eCommerce company which is also a publisher of digital ads in China. We discuss how this design is engineered within the platform’s auction-drive ad-allocation system, which is typical of modern digital advertising marketplaces. Finally, we present results from a parallel experiment involving 16 advertisers and millions of JD.com users. These results showcase the importance of accommodating a role for interactions and illustrate the main ideas.


Gaming or Gambling? An Empirical Investigation of the Role of Loot Boxes in Video Games (with Andrey Simonov)

We examine the role of “loot boxes” — a common in-game monetization model in which players purchase a “black box” with a randomized reward of virtual items — in mobile games. We first build a stylized model that separates out two alternative sources of loot box preferences — a functional value of loot boxes, stemming from the complementarity of loot boxes and game play, and a persistent taste from gambling, which is opening a loot box with an uncertain reward. These two alternative views are at the heart of an on-going policy debate in multiple jurisdictions. We separate out these alternative explanations by estimating a dynamic discrete choice model using detailed activity records from a popular Japanese mobile game. While we find that consumers are more likely to purchase loot boxes when they are functionally beneficial, our preliminary findings show that direct utility from loot boxes is a more important driver of their consumption.


How Recommendation System Feedback Loops Disproportionately Hurt Users with Minority Preferences (with Brandon Stewart)

Video

Algorithmic recommendation systems impact the choices of millions of consumers daily; these systems exist for a wide variety of markets, including both consumable and durable goods, as well as digital and physical goods. After a recommendation system is in place, it will need to be periodically updated to incorporate new users, new items, and new observed interactions between users and items. These observed data, however, are algorithmically confounded: they are the result of a feedback loop between human choices and the existing algorithmic recommendation system. Using simulations, we explore the roles of offline evaluation and A/B testing to perform model selection and update a recommendation system. We find that these choices have the greatest impact on users belonging to minority preference segments.


Automation, Career Values, and Political Preferences (with Maria Petrova, Gregor Schubert, Bledi Taska)

Recently, there has been much evidence linking economic shocks in the form of automation to employment and wage outcomes, as well as political outcomes. In this paper, we try to understand the mechanisms through which economic and political effects of automation are linked. In particular, we go beyond current worker outcomes by introducing a new measure of future career prospects. We show that automation does not only affect current wages, but that occupations also differ in the degree to which workers' career is affected by automation, as automation affects both wages in jobs that workers might aspire to move into, and the likelihood of different career moves. Moreover, the labor market effects of automation differ by demographic group and local area characteristics. We then demonstrate that these patterns of heterogeneity in the impact of automation align with shifts in voter preference towards Donald Trump in the 2016 election -- with negative impacts predicting a shift in preference towards Trump.


Privacy & Market Concentration: Intended & Unintended Consequences of the GDPR (with Samuel Goldberg and Scott Shriver)

Video

We show that the European Union's General Data Protection Regulation (GDPR) reduced data sharing online, but had the unintended consequence of increasing market concentration among technology vendors that provide support services to websites. We collect panel data on the web technology vendors selected by more than 27,000 top websites internationally. The week after the GDPR's enforcement, website use of web technology vendors for EU users falls by 15%. Websites that would face greater penalties under the GDPR drop more vendors. Websites are more likely to drop smaller vendors, which increases the relative concentration of the vendor market by 17%. Increased concentration predominantly arises among vendors that use personal data such as cookies, and from the increased relative shares of Facebook and Google-owned vendors, but not from website consent requests. This suggests that increases in concentration are driven by website vendor choices rather than changes in user behavior.


Government Advertising in Market-Based Public Programs:Evidence from the Health Insurance Marketplace (joint with You Suk Kim, Federal Reserve Board)

This paper studies government and private marketing activities in the context of the Affordable Care Act health insurance marketplace. Using detailed TV advertising data, we present evidence that government advertising and private advertising are targeted to different geographical areas and provide different messaging content. Then, by exploiting discontinuities in advertising along the borders of local TV markets, we estimate the impacts of government advertising and private advertising on enrollment. We find that government advertising has a market-expansion effect, whereas private advertising tends to steal consumers from other insurers, which may cause excessive private advertising spending. Using the equilibrium model of marketplaces, we explore the impact of changing government advertising spending. We find that government advertising increases the total program enrollment and reduces inefficient rent-seeking advertising competition among private insurers.


Sophisticated Consumers with Inertia: Evidence from a Large Scale Field Experiment (with Klaus Miller and Navdeep Sahni)

Consumer inertia - the tendency to remain passive - is well-documented and predicts choices in many contexts. Yet little is known about consumers expectations of their future inertia. Even if consumers are inert, will they avoid products or contracts that exploit inertia? Due to lack of good data, the total effect of inertia (initial take-up plus long-run behavior), coupled with its source, has not been studied. We overcome these challenges with a large-scale field experiment spanning a total of 2M readers of a large European newspaper. We vary between subjects the terms of promotional contracts. Most importantly if a contract renews automatically or cancels by default after a promo period. We observe the take-up, long-term subscription up to two years, and usage of the entire experimental sample. We find evidence in support of “sophisticated inertia” - consumers are indeed inert but appear to be forward looking and account for future inertia, thus mitigating the effects of exploitative contracts.


Political Advertising Online and Offline (with Erika Franklin Fowler, Michael M. Franz, Zachary Peskowitz, and Travis N. Ridout)

Despite the rapid growth of online political advertising, the vast majority of scholarship on political advertising relies exclusively on evidence from candidates’ television advertisements. The relatively low cost of creating and deploying online advertisements and the ability to target online advertisements more precisely may broaden the set of candidates who advertise and allow candidates to craft messages to more narrow audiences than on television. Drawing on data from the newly-released Facebook Ad Library API and television data from the Wesleyan Media Project, we find that a much broader set of candidates advertise on Facebook than television, particularly in down-ballot races. We then examine within-candidate variation in the strategic use and content of advertising on television relative to Facebook for all federal, gubernatorial, and state legislative candidates in the 2018 election. Among candidates who use both advertising media, Facebook advertising occurs earlier in the campaign, is less negative, less issue focused, and more partisan than television advertising.


Impact of Lottery Play Decisions on Consumer and Retailer Earnings (with Paul Parker and Yakov Bart)

The objective of this paper is to quantify the extent to which lottery spending, game choice, and number selection vary with players’ level of income, and how such variations affect their winnings and ticket retailers’ revenues. Previous literature has demonstrated that lotteries can be considered a regressive tax as poorer consumers spend a greater proportion of their income on tickets, but has ignored variations in the choice of game and of numbers made by different income groups, and the heterogeneous consequences for ticket sellers’ earnings. Using both market and transactional lottery data, we find that consumers in low-income areas have a preference for games with a lower payouts and are less likely to switch between games with changes in jackpot size. Further, we find that players in low-income areas are more likely to choose numbers manually and select popular number combinations; both of these preferences result in lower expected earnings. Overall, these differences in preferences over lottery play lead to players in low-income areas losing 10% more per ticket than those in high-income areas. Lottery retailers’ revenues in poorer areas are 45% lower than for those in higher income areas, and they earn 0.19% less per ticket, because low income players prefer games with lower payouts. Our results underscore the need to consider all stages of the consumer decision process when evaluating the economic impact of lotteries on consumers and retailers alike.


Shrinkage Priors for High-Dimensional Demand Estimation (with Jim Griffin)

Estimating demand for wide assortments of differentiated goods requires the specification of a demand system that is sufficiently flexible. However, flexible models contain many parameters and will require regularization in high dimensions. For example, log-linear models suffer from a curse of dimensionality as the number of price elasticity parameters grows quadratically in the number of goods. In this paper, we study the specification of Bayesian shrinkage priors for price elasticity parameters within a log-linear demand system. Traditional regularized estimators assume fixed shrinkage points set to zero which can be at odds with many economic properties of cross-price effects. We propose a hierarchical extension of the class of global-local priors to allow the direction and rate of shrinkage to depend on a product classification tree. We use both simulated data and retail scanner data to show that, in the absence of a strong signal in the data, estimates of cross-price elasticities and demand predictions can be improved by imposing shrinkage to higher-level group effects rather than zero.


Gains from Convenience and the Value of E-commerce (with Bart Bronnenberg)

Why do consumers value shopping online? We measure and decompose the value of e-commerce to individual consumers, allowing for a role of convenience in the form of avoiding transportation costs. Complementing panel data of household purchases in the apparel industry with precise locations of consumers and stores, we show that travel distance is a strong driver of consumer store choice, and of substitution across local chains and to the online channel. Using a structural model of retailer and channel choice, we report that around 2016-2018 the total value from e-commerce to consumers is equivalent to a 22% discount on all prices. About a quarter of this value comes from convenience in the form of lower transportation costs, a quarter from intensified price competition, and the remaining half from the addition of new online retailers and online channels of existing offline retailers. Further, we find that consumer gains are heterogeneous: consumers far from offline stores or experienced in online shopping will benefit more from e-commerce, whereas consumers who likely do not shop online still benefit indirectly from the lower prices. Finally, our results show that, as consumers gain more experience in shopping online, substantial additional gains from e-commerce are still to materialize in the future.


Asymmetric Consequences of Cyber-Vulnerability on Health Services (with Anja Lambrecht and Catherine Tucker)

Healthcare IT is crucial to both the healthcare industry and patients. However, at the same time, it leaves health systems vulnerable to cyber attacks. This paper explores the resilience of health services to cyber attacks. Specifically, we look at the effect of the WannaCry cyber attack in May 2017, on the National Health Service England. We document three important empirical facts: First, the disruption was far more short-lived than portrayed in official reports. Second, that hospitals managed to in general pursue strategies which minimized disruptions to the most ill patients. Third, that there is some evidence that the disruption did affect those with lower incomes and who were non-white more negatively.


Reference Dependence and Price Negotiations - The Role of Advertised Reference Prices

Paper, Slides

We investigate how the presence of two prices - a posted or sale price, and a higher regular or an advertised reference price (ARP) affects the discount a consumers negotiates. Utilizing data from a large durable goods retailer, we find that a $1 increase in ARP results in a 8.8 cents decrease in the realized discount. Our analysis accounts for potential issues due to consumer selection into products, sample selection bias from bargaining, endogeneity stemming from unobserved demand and quality shocks, and measurement error due to salespeople's compensation structure. We supplement these data with a laboratory experiment to explore the underlying mechanism, and study how much of the effect of ARP on revenue is driven by its impact on likelihood to purchase (as has been previously studied) versus the negotiated discount, independent of the purchase likelihood. Based on the experimental data, we find that a $1 increase in ARP decreases the realized discount by 4.3 cents. Further, the data reveal that (i) an increase in ARP lowers a subject's likelihood to initiate a negotiation, (ii) 11% of the total effect of ARP on revenue can be attributed to an increase in realized price after controlling for the purchase likelihood, and (iii) there exist gender differences in the effect of ARP on negotiated discount. We discuss the implications of these findings for retailers and policy makers.


BMI, Food Purchase, and Price Sensitivity [with Ying Bao (U. Toronto), Edward Jaenicke (Penn State), and Emily Wang (U. Mass) ]

Abstract: We examine the relationship between obesity and food purchase behavior using a novel and unique dataset that links individual-level scanner data on food purchases to survey data containing questions about an individual's obesity status. We find that BMI is positively related to higher purchase shares of vice goods such as ice cream or potato chips, and higher BMI individuals are more price sensitive in those product categories.  In contrast, we do not find any statistically significant relationship between BMI and purchase likelihood or price sensitivity for comparable non-vice product categories.  Interestingly, we do not find evidence that higher BMI individuals are more deal sensitive, or stockpile more in response to deals.  Our estimated relationship between BMI and price sensitivity is not moderated by individual habits, obesity related disease conditions, or factors related to an individual's concern about weight.


Overcoming the Cold Start Problem of CRM using a Probabilistic Machine Learning Approach