Operations and Analytics Seminars and Workshops

Abstract: Problem Definition: Lack of patient adherence to treatment protocols is a main barrier to reducing the global disease burden of tuberculosis (TB). Using data from a completed RCT, we study the operational design of a treatment adherence support platform that requires patients to verify their treatment adherence on a daily basis, with a focus on improving personalization.
Methods: We first focus on personalized enrollment. We trained a causal forest model to answer three research questions: (1) Was the effect of the intervention heterogeneous across individuals? (2) Was the intervention less effective for high-risk patients? (3) Can differentiated care improve program effectiveness and equity in treatment outcomes? We then focus on personalized outreach. Inspired by reinforcement learning, we provide a model-free approach to solving the problem of optimizing personalized interventions for patients to maximize some long-term outcome, in a setting where interventions are costly and capacity-constrained. Results: For personalized enrollment, we find that individual intervention effects—the percentage point reduction in the likelihood of an unsuccessful treatment outcome—ranged from 4.2 to 12.4, with an average of 8.2. The intervention was beneficial for 76% of patients, and most beneficial for high-risk patients. Differentiated enrolment policies, targeted at high-risk patients, have the potential to (1) increase the average intervention effect of DAT services by up to 28.5% and (2) decrease the population average and standard deviation (across patients) of the probability of an unsuccessful treatment outcome by up to 8.5% and 31.5%, respectively. For personalized outreach, we show that under a natural set of structural assumptions on patient dynamics, our novel approach recovers at least 1/2 of the improvement possible between a naive baseline policy and the optimal policy. At the same time, our policy is both robust to estimation errors and interpretable. Numerically, we find that our policy can provide the same efficacy as the status quo with approximately half the capacity for interventions.

Abstract: The combination of machine learning (for prediction) and optimization (for decision-making) is increasingly used in practice. However, a key challenge is the need to align the loss function used to train the machine learning model with the decision loss associated with the downstream optimization problem. Traditional solutions have limited flexibility in the model architecture and/or scale poorly to large datasets. We propose a principled decision-aware learning algorithm that uses a novel Taylor expansion of the optimal decision loss to derive the machine learning loss. Importantly, our approach only requires a simple re-weighting of the training data, allowing it to flexibly and scalably be incorporated into complex modern data science pipelines, yet producing sizable efficiency gains. We apply our framework to optimize the distribution of essential medicines in collaboration with policymakers at the Sierra Leone National Medical Supplies Agency; highly uncertain demand and limited budgets currently result in excessive unmet demand. We leverage random forests with meta-learning to learn complex cross-correlations across facilities, and apply our decision-aware learning approach to align the prediction loss with the objective of minimizing unmet demand. Out-of-sample results demonstrate that our end-to-end approach significantly reduces unmet demand across 1000+ health facilities throughout Sierra Leone. Joint work with O. Bastani, T.-H. Chung and V. Rostami.

Abstract: The majority of online sales worldwide take place in online marketplaces that connect many sellers and buyers. The presence of numerous third-party sellers leads to a proliferation of options for each product, making it difficult for customers to choose between the available options. Online marketplaces adopt algorithmic tools to curate how the different options in an assortment are presented to customers. This talk will discuss some of the key differences between retailers and marketplaces, and different mechanisms for algorithmic assortment curation in marketplaces. We will focus on one such tool, a buybox, that algorithmically chooses one option to be presented prominently to customers. We leveraged the staggered introduction of buybox within a prominent product category in a leading online marketplace to study how introducing buybox impacts marketplace dynamics, exploring consequences for customers, sellers, and the marketplace operator.

Abstract: Despite recent advances in language processing algorithms, chatbot technology continues to face adoption hurdles. We survey chatbot users about their experiences and use their testimonies to construct a decision model of customer choice between the chatbot service channel and the live agent service channel. The fundamentals of this choice are the time spent in line and in service, the chatbot’s success rate, and the qualitative differences in the service experience provided by the chatbot and by the live agent. We then conduct experiments in which participants choose, and then experience, the chatbot or the live agent channel as we vary operational (i.e., times spent and chatbot success rates) and qualitative features of the chatbot. We find that users respond positively to improvements in chatbot operational performance; however, the chatbot channel remains underutilized relative to what expected time minimization would predict. Additional experiments show that this underutilization is caused by two separate mechanisms: algorithm aversion (aversion to an algorithmic service provider), and gatekeeper aversion (aversion to any service format that may involve multiple stages). Examining potential remedies, we find that algorithm aversion can be mitigated by making salient the expected time savings offered by the chatbot. However, gatekeeper aversion is more persistent and harder to overcome. We conclude by building and estimating a structural model of channel demand and by proposing a behavior-aware service design that reduces the firm’s staffing costs by up to 22%.
Authors: Evgeny Kagan (JHU Carey), Maqbool Dada (JHU Carey), Brett Hathaway (BYU Marriott)

Abstract: We examine a large transaction-level data set of an Italian omnichannel furniture retailer to study channel-specific effects of fulfillment lead time on demand. This omnichannel retailer sells the same products and has the same product fulfillment across three channels – showroom, online, and catalog. A showroom channel carries no inventory but allows customers to touch and feel the products. An online channel provides a website for consumers to browse and order products. A catalog channel sends a product catalog to all the households in Italy for them to place an order over the phone. We find that the showroom channel makes consumers less sensitive to fulfillment lead time than both online and catalog channels. In particular, a 10% increase in lead time (1.83 days from the sample mean of 18.26 days) causes a 0.20% reduction in the daily sales volume (approximately 20 units from the sample mean of 10,140 units per day) at the showroom, less than the reduction of 0.87% and 0.56% in the online and the catalog channels, respectively. This finding contradicts the common practical and theoretical assumption about homogeneous lead time-sensitivity across channels. In addition, we find that niche products and experience goods accentuate the difference of lead time-sensitivity between the showroom and non-physical channels. Our study highlights the previously-ignored fulfillment time-sensitivity aspect of the physical store’s value.

Abstract: Problem definition. Generic drugs provide life-saving, affordable treatment to millions of people each day. A key assumption that providers and consumers make is that all generics are equally safe, regardless of where the drug is manufactured. However, due to the opaqueness of the generic drug supply chain, few empirical studies have tested this pivotal assumption. This study addresses this gap. To do so, we identify a novel source of drug manufacturing location data and conduct exact matching analysis to compare serious recalls for generic drugs made in emerging economies, such as India and China, to serious recalls for interchangeable generic drugs made in advanced economies, such as the U.S, Canada, and Europe. Methodology/results. We exactly match 4,076 generic drugs from 2009 to 2018 across advanced and emerging economies based on Food and Drug Administration (FDA) drug equivalency criteria. We operationalize drug safety as serious drug recalls. We find that generic drugs manufactured in emerging economies experience significantly higher recall rates, which translates to 33.2% more frequent serious recalls, than supposedly interchangeable generic drugs made in advanced economies. Numerous robustness checks and additional analyses, including a propensity score matching model, a two-stage least squares analysis, and a between-within regression model, demonstrate consistent results. Managerial implications. Our study finding leads to implications for how the FDA should regulate firms that make generic drugs in emerging economies but sell them in the U.S. Particularly, we call for increased transparency of the drug supply chain and increased unannounced inspections in generic drug manufacturing plants in emerging economies.

Abstract: In order to improve the welfare of smallholder farmers, multiple countries (e.g., Ethiopia and India) have launched digital agricultural platforms to transform traditional markets. However, there is still mixed evidence regarding the impact of these platforms and more generally how they can be leveraged to enable more efficient agricultural supply chains and markets. In this talk, we describe a body of work that provides the first rigorous impact analysis of such a platform and highlights several important supply chain and logistics parameters that can inform its design and optimization. The work is focused on the Unified Market Platform (UMP) that connects all the agriculture markets in the state of Karnataka, India. Leveraging both public data and detailed bidding data from the platform, a difference-in-differences analysis demonstrates that the launch of the UMP has significantly increased the modal prices of certain commodities (2.6%-6.5%), while prices for other commodities have not changed. Furthermore, the analysis provides evidence that logistical challenges, bidding efficiency, market concentration, and price discovery processes are important factors explaining the variable impact of UMP on prices. These insights led to the design and field implementation of a new two-stage auction mechanism. The auction design aims to intensify the anticipated regret of the traders to increase the farmers’ revenue. To ensure implementability and protect farmers’ revenue, the design process is guided by theory-informed, semi-structured interviews with a majority of the traders in the field and carefully accounts for operational constraints. The interviews suggest that both anticipated regret and anchoring would likely affect the traders’ bidding strategies in a two-stage auction. A new behavioral auction model is thus developed to capture these factors and determine when the two-stage auction can generate higher revenue for farmers than the traditional single-stage, first-price, sealed-bid auction. The new auction mechanism was implemented on the UMP for a major market of lentils in February 2019. By the end of May 2019, commodities worth more than $6 million (USD) had been traded under the new auction. A difference-in-differences analysis demonstrates that the implementation has yielded a significant 4.7% price increase with an impact on farmer profitability ranging from 60%–158%, affecting over 10,000 farmers who traded in the treatment market. This talk is based on joint work with Retsef Levi (MIT), Somya Singhvi (USC), Manoj Rajan (ReMS), and his team in Karnataka, India.

Abstract: The unobservability of the borrower’s action is a major reason that smaller enterprises face high financing costs and credit rationing. We show that bank loans fully monitored by blockchain allow poor firms with low working capital to eliminate this agency cost. Interestingly, this is achieved by financing all production using fully monitored debt and leaving all internal capital unused; because the use of private, unmonitored internal capital creates unobservability whereas all-debt financing provides full transparency of operations. In contrast, rich firms find it costly to eliminate moral hazard via transparency and they prefer a mix of internal capital and unmonitored debt to finance production. We identify the working capital level at which a firm is indifferent between using all-debt or mixed financing for production. A poor firm with working capital below this indifference level strictly prefers all-debt financing and a rich firm above the level strictly prefers a debt-equity mix. We extend our results to a supply chain and show that the entire supply chain benefits from the use of monitored debt by an individual firm. The “inefficient” bankruptcy cost can create value under blockchain by mitigating the deadweight loss due to decentralization.