State Dependence in Brand Choice
In a recent paper with Julia Levine (published in Marketing Science) we analyze whether consumers that start purchasing new brands persist in their choices. We leverage stock-outs due to hurricanes that force to consumers to switch to a different brand because their preferred product is not available. Interestingly, we find that switches due to hurricane stock-outs do not exhibit any persistence and consumers switch back to their pre-hurricane purchases immediately. Therefore, we conclude that consumers do not exhibit structural state dependence in their brand choices. We rule out that our null result is driven by unusual purchases during a hurricane or context-specific purchase behavior when preparing for a hurricane. Our research is nicely summarized in this VoxEU article.
Machine Learning and Targeted Marketing
Machine learning methods are increasingly used to design personalized marketing strategies such as targeted coupons and personalized advertising messages. In a talk at the 2023 Katia Campo symposium, I talked about what I see as three important ingredients to designing targeting strategies:
(1) the importance of incremental targeting (i.e. based on treatment effect heterogeneity)
(2) how to derive policies from machine learning estimates by using a decision-theoretic framework
(3) how to compare the performance of ML methods based on the profits that policies derived from a given model generate
Flexible Demand Estimation & Optimal Price Setting in Online Retail
Many online retailers sell hundreds of different products, which makes optimal price setting a difficult task that requires the retailer to understand substitution patterns between similar products in their assortment. In a recent working paper (joint with Tomomichi Amano and Andrew Rhodes) we propose a demand model that leverages information on which products consumers tend to search (i.e. browse) together before making a purchase. Such data on search behavior is often collected by online retailers, more abundant than purchase data, and highly informative about product similarity and hence substitutability.
The presentation posted here (prepared for the TSE Digital Economics conference) provides an overview of our approach.
Optimal Price Targeting
In a recent paper with Adam Smith and Ishant Aggarwal (published in Marketing Science), we compare different approaches to estimating demand with consumer panel data with the goal of deriving personalized pricing policies. We provide a unified framework that lets us compare different modeling approaches with different data inputs in terms of the profits they generate when a targeted pricing policy is derived from each model.
We find that purchase histories are useful inputs in any model of demand, whereas demographics information has a small impact on the profitability of pricing policies. Model performance is variable with a Bayesian hierarchical model generating the largest profit gain followed by regularized regression and a neural network. We also find that measures of model fit are almost uncorrelated with predicted profits from a targeted pricing policy and hence statistical fit does not provide useful guidance to model selection.
More detail are in this twitter thread about the paper.
Preference Heterogeneity, Consumer Search, & Targeted Marketing
It is a well-known stylized fact in quantitative marketing that consumers are highly persistent in their product choices. Such persistence has typically been rationalized through strong heterogeneity in preferences across consumers. In a recent paper with Ilya Morozov, Xiaojing Dong, and Liwen Hou, we show that even mild preferences for specific products can lead to persistent choices if consumers only evaluate a limited set of options. We show that when data on search and purchase behavior is available, we can disentangle the role of preferences and search frictions and, taking search costs into account leads to smaller estimates of preference heterogeneity. As a consequence there is less scope for targeted marketing.
More details in this twitter thread: "Consumer Search & Heterogeneity"
The Impact of Soda Taxes
What do we find?
The tax leads to a 34% price increase and a 46% reduction in sales in Philadelphia. A large amount of cross-shopping to stores outside of Philadelphia offsets more than half of the reduction in sales in the city and decreases the net reduction in sales of taxed beverages to only 22%. We find no significant substitution to bottled water and modest substitution to untaxed natural juices.
What does this mean for tax policy design?
The current tax of 1.5 cents/oz is close to revenue-maximizing, but a slightly higher tax could lower sales of taxed beverages with only a small loss in revenue. A tax of 3 cents/oz (which was contemplated in Philadelphia) would have shrunk revenue by 70%. We also show that tax avoidance through cross-shopping severely constrains revenue generation and nutritional improvement. Widening geographic coverage will likely generate both more revenue and a larger decrease in the sales of taxed beverages.
The video on the right provides more details on our study ...
Runner-up for Dick Wittink Best Paper Award
My paper “The Impact of Advertising Along the Conversion Funnel” with Song Yao was the runner-up for the Dick Wittink best paper award at the QME journal. In the paper, we analyze advertising effectiveness along the conversion funnel in brick-and-mortar supermarkets using novel “path-tracking” data on consumers’ movement in the store. The interview posted here provides a short summary of our findings.
Harnessing the Power of Social Media
Firms are increasingly trying to tab into social media and have users promote products on their behalf. But how effective is such a marketing strategy? In recent research of mine (with Song Yao and Wenbo Wang), we address this question. The video interview posted on the right provide a summary of our findings.