MSOM Society Student Paper Competition: Abstracts of 2019 Winners

Dynamic Learning and Pricing with Model Misspecification

Milashini Nambiar

Massachusetts Institute of Technology

Current affiliation: Institute for Infocomm Research, A*STAR, [email protected]

Advisor: David Simchi-Levi, Massachusetts Institute of Technology

We study a multiperiod dynamic pricing problem with contextual information, where demand depends on time-varying features such as product characteristics, customer types, and underlying market economic conditions. At each time period, the seller sequentially observes past demand, updates the demand model parameters, and then chooses the price for the next period based on these features. We are especially interested in model misspecification in this context: in particular, we study the setting where the seller assumes an incorrect relationship between demand and the features, and we show that model misspecification can lead to the price and the per-period demand prediction error becoming correlated, which in turn leads to inconsistent price elasticity estimates and, hence, suboptimal pricing decisions. To address this issue, we propose a “Random Price Shock” (RPS) algorithm that dynamically generates randomized price shocks in order to simultaneously estimate price elasticity and maximize revenue. We show for the cases of independent identically distributed (IID) as well as non-IID features that the RPS algorithm has strong theoretical performance guarantees and that it is robust to model misspecification. Finally, we collaborate with Oracle Retail to validate the RPS algorithm in a real-world setting. We perform offline simulations on a large fashion retail data set and find that RPS is expected to earn 8%–20% more revenue on average than competing algorithms that do not account for price endogeneity.

Inferring Consideration Sets from Sales Transaction Data

Dmitry Mitrofanov

New York University, [email protected]

Advisors: Srikanth Jagabathula, New York University; Gustavo Vulcano, Universidad Torcuato Di Tella

Understanding consumer preferences is critical when optimizing prices and planning in retail operations, as well as when matching supply and demand in online platforms. In pursuing such an objective, the identification of the consideration set of the consumers (i.e., the set of products really accounted for by consumers prior to making a choice) is indeed a fundamental input. In this paper, we propose a methodology to identify consideration sets from sales transaction data in a data-driven way. We assume that customers are boundedly rational and make their purchases in a two-stage process. First, they sample their consideration set and then purchase the most preferred item therein. Our contribution to the literature is twofold. Theoretically, we address the problem of identifiability of consider-then-choose models from data. Because calibrating this class of choice models is a hard problem, we propose a framework to effectively estimate them and infer consideration sets. The methodology to model the consideration set formation is founded on machine learning techniques that can account for nonlinear-in-parameter utilities in a tractable way. Then we apply the proposed methodology to retail store data and data obtained from a car-sharing platform, and we observe that accounting for consideration sets can boost the predictive performance in comparison with classical choice-based demand benchmarks. Our findings suggest that consider-then-choose models tend to be rather robust to the degree of ambiguity in the consideration set definition, and their relative importance in prediction tasks increases with this noise. Moreover, we show that the consider-then-choose type of choice models can provide important managerial insights about the consideration set formation.

Multichannel Delivery in Healthcare: The Impact of Telemedicine Centers in Southern India

Kraig Delana

London Business School

Current affiliation: University of Oregon, [email protected]

Advisors: Kamalini Ramdas, London Business School; Nicos Savva, London Business School

Telemedicine is increasingly used across the developing world to expand access to healthcare, to improve outcomes, and to reduce costs. However, the impact of telemedicine in these settings—particularly on existing physical healthcare delivery channels—has not been thoroughly examined. We use data from one of the largest teleophthalmology implementations in the world to examine this issue. Using a quasi-experimental difference-in-differences approach on a data set of more than 4.8 million visits over almost a decade (January 2006–May 2015), we find that telemedicine centers generate a 31% increase in the overall network visit rate, of which 62% comprise new patients, suggesting a substantial increase in access. We also find a 5.1% reduction in hospital visit rates, suggesting that some patients substitute hospital visits with telemedicine center visits. We find substantial heterogeneity in the impact of treatment depending on the clinical complexity of patient needs. The rate of eyeglasses prescriptions to correct for simple refractive errors increases by 18.5%, but the rate of cataract surgery to replace the natural lens in a patient’s eye with an artificial lens remains unchanged. The increase in access and treatment rates does not significantly impact the direct costs incurred by patients but reduces their indirect costs (measured as the average distance travelled to care) by 30% (12 km). Finally, we find significant spatial heterogeneity in these effects depending on the relative distance of a patient’s location from the telemedicine center as well as from the hospital. These results have important implications for the design of telemedicine networks and the portfolio of healthcare services provided through them.

Timely After-Sales Service and Technology Adoption: Evidence from the Off-Grid Solar Market in Uganda

Amrita Kundu

London Business School, [email protected]

Advisor: Kamalini Ramdas, London Business School

Adoption and continued use of novel technologies has the potential to significantly accelerate social and economic development in emerging markets. In this paper, we examine to what extent timely after-sales service—that is, fast resolution of repair tasks—impacts technology adoption in emerging markets. We address this question using detailed customer-level sales and service data from a leading solar distribution company operating in Uganda. We develop a fixed effects base specification and two instrumental variables specifications that leverage different sources of geospatial variation—in service task locations, weather, and road quality. We find that timely after-sales service experienced by existing customers is a strong driver of adoption by first-time users. A one-week increase in average wait time for service decreases adoption up to 32.2%. The relationship between wait times and adoptions is highly heterogeneous and depends on the type of pending service cases. We also find that the number of customers acquired through referrals from an existing customer depends on the referring customer's service wait time. This provides evidence of a strong word-of-mouth channel of information sharing. Our findings have direct implications for the customer acquisition strategies of technology firms and for technology investors in emerging markets. Our results are also relevant for policy makers who aim to harness technology to improve the socioeconomic lives of people living in these regions.

Can Big Data Cure Risk Selection in Healthcare Capitation Programs?

Zhaowei She

Georgia Institute of Technology, [email protected]

Advisors: Turgay Ayer, Georgia Institute of Technology; Daniel Montanera, Georgia State University

Early empirical evidence indicates that Medicare Advantage (MA), the largest capitation payment program in the U.S. healthcare market, unintentionally incentivizes health plans to cherry-pick profitable patient types, which is referred to as risk selection. Motivated by this observation, we study the root causes of risk selection in the MA market design and potential strategies to eliminate risk selection. The existing literature primarily attributes the observed risk selection in the MA market to data limitations and low explanatory power (e.g., low ) of the current risk adjustment design in the MA market. With the availability of big data and advancements in machine learning (ML) techniques, risk selection as a result of imperfect risk adjustment is expected to gradually disappear from the MA market. However, our study shows that big data and ML alone cannot cure risk selection in the MA capitation program. More specifically, we show that even if the current MA risk adjustment design becomes informationally perfect (e.g., ) through the availability of big data and advanced ML algorithms, health plans still have incentives to conduct risk selection through strategically subsidizing some subgroups of patients using capitation payments collected from other subgroups, which we call risk selection induced by cross subsidization. Furthermore, we develop and present selection-proof capitation mechanisms to eliminate this type of risk selection behavior from the MA market. Our findings further indicate that through some small modifications to the existing medical loss ratio mechanism, risk selection of this kind could be eliminated from the MA market.

Matching in Labor Marketplaces: The Role of Experiential Information

Jiding Zhang

University of Pennsylvania, [email protected]

Advisors: Elena Belavina, Cornell University; Karan Girotra, Cornell University; Ken Moon, University of Pennsylvania

Online labor marketplaces assign workers to short-term jobs. For some jobs, the choice of the best worker is based on ex ante observable information (e.g., driver assignment based on location in ride-hailing). In others, the assignment is driven by experiential information—that is, information obtained privately only through the worker performing the job (e.g., the fit of a childcare provider with a family). This study develops an empirical framework to impute the relative importance of each kind of information from participants' past hiring choices. Our moment inequality approach accommodates high worker turnover, varying choice sets, and limited observations of a very large number of market participants—all key characteristics of online labor markets. We apply our framework to two markets, exploiting a natural experiment that changed marketplace commissions. On the basis of over 1.2 million hiring decisions, we estimate that experiential information is a key driver of hiring choices, whereas ex ante observable fit is relevant only for the simplest jobs. Using our estimates, we propose and evaluate alternative assignment policies. The best-performing policies prioritize repeat work and, surprisingly, ignore ex ante observable information to instead experiment with new workers and generate experiential information. Such policies can increase buyer welfare by as much as 45.3% (47.1%) of gross revenue in the data entry (web development) market compared with the current practice of skills-based matching. Policies exploiting buyers' past revealed preferences (in repeat work) without incorporating exploration still underperform by 18.9% in data entry and 8.7% in web development.