I study consumer choices and market outcomes using unstructured, large-scale mobile location data that I combine with business and marketing information. I also examine the tradeoff between the marketing value of mobile data and consumer privacy. I employ methods from econometrics, causal inference, and machine learning.
My PhD advisor is Dina Mayzlin.
Please scroll down to find more about my research.
Departments
RESEARCH + PUBLICATIONS
Do professional critics' opinions still matter in the age of online user reviews? We study both the economic and the social impact of critics' reviews in the restaurant industry. We ask: (1) To what extent do food critics' reviews impact restaurant foot traffic? (2) Are there spillover effects on neighboring restaurants? (3) Do food critics' reviews increase the racial diversity of the clientele? To address these questions, we assemble a unique dataset of food critics' reviews featured in major US newspapers and a popular restaurant recommendation website, The Infatuation (2019-2020, pre-pandemic). We combine this dataset with granular consumer mobility data collected from smartphones to measure restaurant foot traffic and to infer visitors' home location and demographics. We exploit quasi-random variation in both restaurant selection and review timing and use a matching estimator to causally identify the effect of critics' reviews on foot traffic to the reviewed as well as neighboring restaurants. We find an average increase in visits of about 7% to the reviewed restaurants in the weeks following the publication of the review. The effect is larger for restaurants with more available capacity and gradually dissipates over time. Interestingly, reviews have a broader economic impact: they lead to an average increase of about 3% in foot traffic to neighboring restaurants. Finally, we demonstrate that critics' reviews contribute to interracial exposure by drawing a more racially diverse clientele to restaurants that were previously homogeneous.
We investigate the impact of brand sociopolitical involvement on customer store visits. We assemble a unique dataset of news articles to identify brand sociopolitical involvement during 2018-2022 and combine it with weekly foot traffic data from about 150,000 stores of 47 highly visited US brands. To measure the effect of each identified event, we employ coarsened exact matching and the synthetic difference-in-differences model. Our results show that brand involvement in sociopolitical issues leads to an average modest decline of 0.8% in foot traffic, with effects varying substantially across events. Through detailed case studies—such as Walmart’s firearm sales restrictions, a Starbucks racial profiling incident, and Chick-fil-A’s change in LGBTQ+ donation policy—we highlight how consumer responses depend on the initiator of the involvement, store neighborhood characteristics, and the brand’s response. Events initiated by corporate leadership tend to produce more variable and less negative outcomes compared to those triggered by employees which can be seen as service failure. Company response to non-corporate incidents—such as issuing public apologies or conducting employee training—can mitigate negative consumer reactions. Finally, service-oriented brands are more susceptible to negative impacts, likely due to the salience of customer-employee interactions.
Mobile location data offers granular detail into consumers' mobility patterns that are indicative of preferences and behavior. Such data are a veritable goldmine for contextual marketing. At the same time, however, the granularity of the data can reveal sensitive information about consumers based on the places that they visit such as healthcare facilities or religious centers. Can the desires of firms seeking actionable insights and users preferring privacy be accommodated? Using device-level mobile location data, we examine the extent to which predictive performance is affected by aggregating individuals into homogeneous clusters that afford increased privacy. Our analyses reveal that predictive performance is minimally affected by forming relatively small homogeneous clusters, while the risk to privacy is reduced substantially. We also find that the use of locations of commercial activity can be used in lieu of home locations, affording consumers increased privacy. Moreover, intentionally avoiding data collection at locations deemed sensitive does not adversely affect business performance while further reducing risks to privacy. Our findings offer guidance to data providers who must balance service to their clients with consumers' growing privacy expectations, as well as providing regulators with insight into the data granularity that firms require for their marketing operations.
Retailers’ efforts to monetize consumer location data remain dominated by inefficient protocols (e.g., geofencing) that customize marketing interactions based solely on app users’ current location. Although extant trajectory mining techniques can remedy these shortcomings, they require high-frequency location data, which poses severe risks to consumers’ privacy. The authors present a novel method to extract marketing value from low-granularity urban mobility data and demonstrate its use in analyzing gas station choice to value customers. The data, also used to infer gas station visits, contain 1.06 million location records on nearly 27,000 devices observed near selected retailers including gas stations during a six-month period in Staten Island, New York. The authors pool consumers’ mobility trajectories from several days to dynamically calculate the distance of stores from consumers’ anticipated trajectories. They then supplement the data with station-level daily fuel prices and estimate a conditional logit model to assess how consumers trade off gas prices versus store distance. In addition to a generally high station loyalty, the authors find that consumers strongly prefer not to deviate far from their common trajectories for fueling trips. Applying their methods in a predictive context, the authors infer the value of newly acquired customers to the studied gas stations to be between $3.00 and $7.59.
In this paper, we show how retailers can use consumer mobility data to assess the relative performance of each store within their network. We use mobile location data from over 5M devices in Manhattan, NY to construct a weighted network of Starbucks stores as nodes, with the edge weights between any two stores reflecting both the overlap between the customers of as well as the distance between the stores. We then compute network centrality measures to capture consumption dynamics in the network. Finally, we employ these variables to train machine learning models predicting whether or not each store closed down during the 20 months following our observation period. Our findings indicate that including network centrality measures derived from urban mobility data using our methods can lead to a better identification of underperforming stores in a retailer’s network, revealed by subsequent store closure decisions.