This year the Wisconsin School of Business' Marketing Department is inviting doctoral candidates to come and present their research to our school.

...

Synopsis

Mobile in-app advertising has grown exponentially in the last few years. In-app ads are often shown in a sequence of short-lived exposures for the duration of a user’s stay in an app. The current state of both research and practice ignores the dynamics of ad sequencing and instead adopts a myopic framework to serve ads. In this paper, we propose a unified dynamic framework for adaptive ad sequencing that optimizes user engagement in the session, e.g., the number of clicks or length of stay. Our framework comprises of two components – (1) a Markov Decision Process
that captures the domain structure and incorporates inter-temporal trade-offs in ad interventions, and (2) an empirical framework that combines machine learning methods such as Extreme Gradient Boosting (XGBoost) with ideas from the causal inference literature to obtain counterfactual estimates of user behavior. We apply our framework to large-scale data from the leading in-app ad-network of an Asian country. We document significant gains in user engagement from adopting a dynamic framework. We show that our forward-looking ad sequencing policy outperforms all the existing methods by comparing it to a series of benchmark policies often used in research and practice. Further, we demonstrate that these gains are heterogeneous across sessions: adaptive forward-looking ad sequencing is most effective when users are new to the platform. Finally, we use a descriptive approach to explain the gains from adopting the dynamic framework.

Revenue-Optimal Dynamic Auctions for Adaptive Ad Sequencing 

Synopsis

Digital publishers often use real-time auctions to allocate their advertising inventory. These auctions are designed with the assumption that advertising exposures within a user’s browsing or app-usage session are independent. Rafieian (2019) empirically documents the interdependence in the sequence of ads in mobile in-app advertising, and shows that dynamic sequencing of ads can improve the match between users and ads. In this paper, we examine the revenue gains from adopting a revenue-optimal dynamic auction to sequence ads. We propose a unified framework with two components – (1) a theoretical framework to derive the revenue-optimal dynamic auction that captures both advertisers’ strategic bidding and users’ ad response and app usage, and (2) an empirical framework that involves the structural estimation of advertisers’ click valuations as well as personalized estimation of users’ behavior using machine learning techniques. We apply our framework to large-scale data from the leading in-app ad-network of an Asian country. We document significant revenue gains from using the revenue-optimal dynamic auction compared to the revenue-optimal static auction. These gains stem from the improvement in the match between users and ads in the dynamic auction. The revenue-optimal dynamic auction also improves all key market outcomes, such as the total surplus, average advertisers’ surplus, and market concentration.

Tesary Lin, Doctoral Student, University of Chicago

...

Synopsis

In this paper, I propose a framework for understanding why and to what extent people value their privacy. In particular, I distinguish between two motives for protecting privacy: the intrinsic motive, that is, a “taste” for privacy; and the instrumental motive, which reflects the expected economic loss from revealing one’s “type” specific to the transactional environment. Distinguishing between the two preference components not only improves the measurement of privacy preferences across contexts, but also plays a crucial role in developing inferences based on data voluntarily shared by consumers. Combining a two-stage experiment and a structural model, I measure the dollar value of revealed preference corresponding to each motive, and examine how these two motives codetermine the composition of consumers choosing to protect their personal data. The compositional differences between consumers who withhold and who share their data strongly influence the quality of firms’ inference on consumers and their subsequent managerial decisions. Counterfactual analysis investigates strategies firms can adopt to improve their inference: Ex ante, firms can allocate resources to collect personal data where their marginal value is the highest. Ex post, a consumer’s data-sharing decision per se contains information that reflects how consumers self-select into data sharing, and improves aggregate-level managerial decisions. Firms can leverage this information instead of imposing arbitrary assumptions on consumers not in their dataset.

Matt McGranaghan, Doctoral Student,  Cornell University

...

Synopsis

A challenge to measuring TV viewer attention is that instant access to social media, news, and work has raised the opportunity cost of engaging with TV ads.  The result may be a significant difference between traditional engagement measures, e.g., tuning, and measures which can capture more nuanced avoidance behaviors.  This paper asks two questions relating to viewer behavior in the context of TV advertising.  First, how do traditional TV tuning metrics relate to a novel set of viewer measures that may be more aligned with broadcasters’ and advertisers’ interests?  Second,what is the relationship between these new measures and ad content?  To answer these questions,we leverage novel, in-situ, audience measurement data that use facial and body recognition technology to track tuning, presence (in room behavior), and attention for a panel of 6,291 viewers and8,465,513 ad impressions,  as well as consider four different classifications of advertising content based on human and machine-coded features.  We find meaningful differences in the absolute levels and dynamics of these behaviors, and can identify ad content for which viewers are systematically more likely to change the channel, leave the room, and stop paying attention.  Such ads reduce the pool of attention to subsequent advertisers as well as the platform itself, a negative externality.  We quantify these spillover effects for the publisher by conducting a series of counterfactual simulations, and find that requiring advertisers to improve their content can result in significant increases in the cumulative levels of viewer tuning, in-room presence, and attention.

Cheng He, Doctoral Student, Georgia Institute of Technology

...

Synopsis

In response to growing environmental concerns, governments have promoted products that are less harmful to the environment—green products—through various incentives. We empirically study the impact of a commonly used non-monetary incentive, namely the single-occupancy permission to high-occupancy vehicle (HOV) lanes, on green and non-green product demand in the U.S. automobile industry. The HOV incentive could increase unit sales of green vehicles by enhancing their functional value through time-saving. On the other hand, the incentive may prove counterproductive if it reduces the symbolic value (i.e., signaling a pro-environmental image) consumers derive from green vehicles. Assessing the effectiveness of green-product incentives is challenging given the endogenous nature of governments' incentive provisions. To identify the effect of the HOV incentive on unit sales of green and non-green vehicles, we take advantage of incentive changes at the county level, and we employ a multitude of quasi-experimental methods, including difference-in-differences with Coarsened Exact Matching, border strategy, and regression discontinuity in time. Unlike previous studies that only examine the launch of the HOV incentive and find an insignificant association between incentive launch and green vehicle demand, we concentrate on its termination. We find that the termination of the HOV incentive decreases unit sales of vehicles covered by the incentive by 14.4%. We provide suggestive evidence that this significant negative effect of HOV incentive termination is due to the elimination of the functional value the incentive provides: time-saving. Specifically, we find that the negative effect is more pronounced in counties where consumers value time-saving more (i.e., counties with a longer commute to work and higher income). Additionally, in line with prior literature, the launch of the HOV incentive is not found to have a significant effect on green vehicle sales. Combined, our findings reveal that the effect of termination is not simply the opposite of that of launch, implying that governments' green product incentives could backfire.

Keywords: sustainability, green products, public policy, government incentives, climate change, technology adoption, policy evaluation, quasi-experiments, difference-in-differences, coarsened exact matching

Alex Burnap, Doctoral Student, MIT Sloan Schooasl of Management

...

Synopsis

Aesthetics are critically important to market acceptance in many product categories. In the automotive industry in particular, an improved aesthetic design can boost sales by 30% or more. Firms invest heavily in designing and testing new product aesthetics. A single automotive "theme clinic" costs between $100,000 and $1,000,000, and hundreds are conducted annually. We use machine learning to augment human judgment when designing and testing new product aesthetics. The model combines a probabilistic variational autoencoder (VAE) and adversarial components from generative adversarial networks (GAN), along with modeling assumptions that address managerial requirements for firm adoption. We train our model with data from an automotive partner — 7,000 images evaluated by targeted consumers and 180,000 high-quality unrated images. Our model predicts well the appeal of new aesthetic designs — 38% improvement relative to a baseline and substantial improvement over both conventional machine learning models and pretrained deep learning models. New automotive designs are generated in a controllable manner for the design team to consider, which we also empirically verify are appealing to consumers. These results, combining human and machine inputs for practical managerial usage, suggest that machine learning offers significant opportunity to augment aesthetic design.

Tommaso Bondi, Doctoral Student, New York Stern School of Business

...

Synopsis

Consumer ratings have become a prevalent driver of choice. I develop a model of social learning in which ratings can inform consumers about both product quality and their idiosyncratic taste for them. Depending on consumers’ prior knowledge, I show that ratings relatively advantage lower quality and more polarizing products. The reason lies in the stronger positive consumer self-selection these products generate: to buy them despite their deficiencies, their buyers must have a strong taste for them. Relatedly, consumer ratings should not be used to infer product design: what is polarizing ex-ante needs not be so among its buyers. I test these predictions using Goodreads book ratings data, and find strong evidence for them. Moreover, social learning appears to serve mostly a matching purpose: tracking the behaviour of Goodreads users over time shows that they specialize as they gather experience on the platform: they rate books with a lower average and number of ratings, while focusing on fewer genres. Thus, they become less similar to their average peer. Taken together, the findings suggest that consumer ratings contribute to both the long tail and, relatedly, consumption segregation. For managers, this illustrates, counterintuitively, the reputational benefits of polarizing products, particularly early in a firm’s lifecycle, but only when paired with the ability to match with the right consumers.

The good, The Bad and The Picky: Consumer Heterogeneity and The Reversal of Movie Ratings

We explore the consequences of consumer heterogeneity on online word of mouth. Consumers differ in their experience, which has two effects. First, experience is instrumental to choice: experts purchase and review better products than non-experts. Second, because of their superior choices, experts endogenously form higher expectations, and thus post more stringent ratings given quality. Combined, these two forces imply that the better the product, the higher the standard it is held to, the more stringent its rating. Thus, relative ratings are biased: low quality products enjoy unfairly high ratings compared to their superior alternatives. When this bias gets large, reputation needs not be increasing in quality. The bias needs not disappear, and can worsen, over time: products with unfairly high ratings mostly attract unexperienced consumers, reinforcing their advantage. We test our theory by scraping data from a well known movie ratings website. We find strong evidence for both of our hypotheses, and that this bias is quantitatively important. We then debias the ratings, and find that the new ones better correlate with the opinions of external critics.

Sam Maglio, Professor, University of Toronto Scarborough 

...

Synopsis

...

Image AddedConspicuous Self-Control: When Status Motives Lead Consumers to Signal Restraint

Synopsis

Extant research suggests that consumers associate high status with wealth, which leads them to behave indulgently by purchasing expensive (vs. less expensive) products when status-signaling motives are activated. We propose that consumers also associate high status with being goal-oriented, which leads them to conspicuously engage in self-control (vs. indulgence). A series of six studies finds that status motives lead consumer to choose products that signal self-control (e.g., healthy vs. unhealthy food, educational vs. entertainment programming) because status motives increase the desire to appear goal-oriented. The effect of status motives on conspicuous self-control is particularly apparent in contexts where the opportunity to signal wealth is absent. For example, the effect of status motives on conspicuous self-control is moderated by the variance in price among product options. Additional findings demonstrate that consumers with active status motives are more willing to conspicuously save (vs. spend) their money when they are reminded that saving behavior is a signal of self-control. This research has important implications for marketers and consumers regarding how and when status motivations can be harnessed to enhance self-control.

Esther Uduehi, Doctoral Student, University of Pennsylvania

Image AddedI am a Person, But They Are a Condition: The Role of Language Choices on Consumer Behavior for Stigmatized Groups

Synopsis

Although millions of consumers deal with various stigmatized identities such as obesity, homelessness, and substance use disorders, little is known about 1) whether stigmatized identity language within the marketplace matches consumer preferences and 2) the psychological factors that impact the use of these language choices towards stigmatized groups. While there is some evidence that people dealing with stigmatized conditions prefer person-first language (e.g., person with obesity) instead of identity-first language (e.g., obese person), the use of person-first language is not universal by brands. This talk first addresses research that uses health ads and voice-over lab studies. We find that people dealing with weight issues are more interested in engaging with nutrition brands that use person-first language to describe their weight identity. However, data scraping weight and nutrition brand websites reveals that brands are more likely to use identity-first language to describe weight identities. This talk then explores what drives the use of identity vs. person-first language for stigmatized groups. Using textual analysis of 1326 nonprofit organizations and academic literature as well as follow-up lab experiments, we find that for conditions perceived to be more changeable, organizations and people are more likely to use identity-first language. The relationship between people’s use of identity-first language for conditions viewed as changeable is mediated by perceptions of onset or personal responsibility for stigmatized others. Our results suggest that if person-first language is helpful in empowering stigmatized groups, it will be necessary to point out the multi-faceted nature of the systems that support the stigmatized condition, such that individuals are not saddled with the type of responsibility that hides their personhood behind their condition.

Prashant Rajaram, Doctoral Student,  Ross School of Business

Image AddedVideo Influencers: Unboxing the Mystique

Synopsis

Influencer marketing is being increasingly used as a tool to reach customers. This is because of the increasing popularity of social media stars who primarily reach their audience(s) via custom videos published on a variety of social media platforms (e.g., YouTube, Instagram, Twitter and TikTok). Despite the rapid growth in influencer marketing, there has been little research on the design and effectiveness of influencer videos. Using publicly available data on YouTube influencer videos, we implement novel interpretable deep learning architectures, supported by transfer learning, to identify significant relationships between advertising content in influencer videos (across text, audio, and images) and video views, interaction rates and sentiment. This is followed up with a second study to investigate whether influencers “learn” these relationships over time. By avoiding ex-ante feature engineering, and instead using ex-post interpretation, our approach avoids making a trade-off between interpretability and predictive ability. We filter out relationships that are affected by confounding factors unassociated with an increase in attention to video elements, thus facilitating the generation of plausible causal relationships between video elements and marketing outcomes which can be tested in the field. A key finding is that brand mentions in the first 30 seconds of a video are on average associated with a significant increase in attention to the brand but a significant decrease in sentiment expressed towards the video. We illustrate the learnings from our approach for both influencers and brands.

Remi Daviet, Doctoral Student, University of Pennsylvania

Image AddedBayesian Deep Learning for Small Datasets: Leveraging Information from Product Pictures

Synopsis

Marketers are often confronted with datasets that contain many variables but are limited in the number of observations, leading to a“large P, small N” problem. With unstructured data, such as product pictures, commonly used deep learning models require the estimation of a large number of parameters, also resulting in a “large K” problem. In this research, we propose a pipeline to process and exploit such unstructured data. We apply it to a novel dataset aggregating all retail sales of distilled spirits in Pennsylvania. We first reduce the high pixel-based dimensionality of the product pictures using a Conditional Generative Adversarial Variational Auto-Encoder (CGAVAE). We then use the result in a deep learning model to predict sales volumes, using Bayesian estimation to mitigate overfitting issues. We show that using the product pictures’ information, in addition to traditional variables such as price and product characteristics, increases the out-of-sample prediction performance for sales volumes by nearly half its base value (R2 increasing from 0.24 to 0.35). We also propose a method to interpret the results and identify relevant product features, potentially allowing for the creation of new theories. Lastly, we use our model in a design optimization exercise, where we identify classes of bottle designs that are predicted to maximize expected revenue.

Christopher Bechler, Doctoral Student, Stanford Graduate School of Business

Image AddedPerceiving attitude change: How qualitative shifts augment change perception

Synopsis

Attitude change and persuasion are among the most studied topics in social psychology. Surprisingly, though, as a field we have virtually zero insight into perceived attitude change—that is, how people assess the magnitude of a shift in someone's attitude or opinion. The current research provides an initial investigation of this issue. Across 6 primary experiments and a series of supplemental studies (total N = 2880), we find consistent support for a qualitative change hypothesis, whereby qualitative attitude change (change of valence; e.g., from negative to positive) is perceived as greater than otherwise equivalent non-qualitative attitude change (change within valence; e.g., from negative to less negative or from positive to more positive). This effect is mediated by ease of processing: Qualitative attitude change is easier for people to detect and understand than non-qualitative attitude change, and this ease amplifies the degree of perceived change. We examine downstream consequences of this effect and discuss theoretical, methodological, and practical implications.

Choosing persuasion targets: How expectations of qualitative change increase advocacy intentions

Synopsis

Advocacy is a topic of increasing import in the attitudes literature, but researchers know little to nothing about how people (i.e., persuaders) choose their targets (i.e., the recipients of their advocacy). Four main experiments and six supplemental studies (total N = 3684) demonstrate that people prefer to direct persuasion efforts toward individuals who seem poised to shift their attitudes qualitatively (e.g., from negative to positive) rather than non-qualitatively (e.g., from positive to more positive). This preference stems from the fact that qualitative attitude change is perceived as greater in magnitude and expected to have a larger impact on behavior. These findings provide initial insight into the factors that drive persuasion target selection, and are inconsistent with what past persuasion research, conventional marketing wisdom, and our empirical evidence suggests persuaders should do. People tend to select persuasion targets they believe they can change qualitatively, but at least sometimes can have greater persuasive impact by targeting individuals who are already leaning in their direction.

MISDIRECTING PERSUASIVE EFFORTS DURING THE COVID-19 PANDEMIC: THE TARGETS PEOPLE CHOOSE MAY NOT BE THE MOST LIKELY TO CHANGE

Synopsis

Persuading people to engage in specific health behaviors is critical to prevent the spread of and mitigate the harm caused by COVID-19. Most of the research and practice around this issue focuses on developing effective message content. Importantly, though, persuasion is often critically dependent on choosing appropriate targets—that is, on selecting the best audience for one’s message. Three experiments conducted during the COVID-19 pandemic explore this target selection process and demonstrate misalignment between who persuaders target and who will display the greatest attitude and behavior change. Although people prefer to send messages encouraging COVID-19 prevention behaviors to targets with slightly negative attitudes toward the behaviors in question, their messages can often have more impact when sent to targets whose attitudes are slightly favorable. Recent insights in categorical perception and message positioning effects in persuasion help explain this misalignment.

David Holtz (Dave), Doctoral Student, MIT Sloan School of Management

Image AddedThe Engagement-Diversity Connection: Evidence from a Field Experiment on Spotify

Synopsis

We present results from a randomized field experiment on approximately 900,000 Spotify users across seventeen countries, testing the effect of personalized recommendations on consumption diversity. In the experiment, users were given podcast recommendations, with the sole aim of increasing podcast consumption. However, the recommendations provided to treatment users were personalized based on their music listening history, whereas control users were recommended the most popular podcasts among their demographic group.We find that the treatment increased podcast streaming, decreased individual-level podcast streaming diversity, and increased aggregate podcast streaming diversity. These results indicate that personalized recommendations have the potential to create consumption patterns that are homogeneous within and diverse across users, and provide evidence of an "engagement-diversity trade-off" when optimizing solely for consumption: while personalized  recommendations increased user engagement, they also affected the diversity of consumed content. This shift in consumption diversity can affect user retention and lifetime value, and impact the optimal strategy for content producers. Additional analyses suggest that exposure to personalized recommendations can also affect the content that users consume organically. We believe these findings highlight the need for both academics and practitioners to continue investing in personalization techniques that explicitly take into account the diversity of content recommendations.

Anchor
mengxia-zhang mengxia-zhang

Mengxia Zhang, Doctoral Student, USC Marshall School of Business

Image AddedCan User-Posted Photos Serve as a Leading Indicator of Restaurant Survival? Evidence from Yelp

Synopsis

Despite the substantial economic impact of the restaurant industry, large-scale empirical research on restaurant survival has been sparse. We investigate whether consumer-posted photos can serve as a leading indicator of restaurant survival above and beyond reviews, firm characteristics, competitive landscape, and macro conditions. We employ machine learning techniques to analyze 755,758 photos and 1,121,069 reviews posted on Yelp between 2004 and 2015 for 17,719 U.S. restaurants. We also collect data on these restaurants’ characteristics (e.g., cuisine type; price level), competitive landscape, and their entry and exit (if applicable) time based on each restaurant’s Yelp/Facebook page, own website, or the Google search engine. Using a predictive XGBoost model, we find that photos are more predictive of restaurant survival than are reviews. Interestingly, the information content (e.g., number of photos with food items served) and helpful votes received by these photos relate more to restaurant survival than do photographic attributes (e.g., composition or brightness). Additionally, photos carry more predictive power for independent, mid-aged, and medium-priced restaurants. Assuming that restaurant owners do not possess any knowledge about future photos and reviews for both themselves and their competitors, photos can predict restaurant survival for up to three years, while reviews are only informative for one year. We further employ causal forests to facilitate interpretation of our predictive results. Our analysis suggests that, among others, the total volume of user-generated content (including photos and reviews) and helpful votes of photos are both positively related to restaurant survival.

Ishita Chakraborty, Doctoral Student, Yale School of Management

Image Added

Attribute Sentiment Scoring with Online Text Reviews: Accounting for Language Structure and Missing Attributes

Synopsis

...