Working Papers under Review and Revision

Discovering and measuring the causal effects of the purchase of one product on the purchases of other products in a shopping basket is an important problem in retailing. It is also extremely challenging because the number of possible interactions among all the products in a retail store can be exponentially large. We present a graph-theoretic causal product network (CPN) representation of consumer basket-shopping behavior and a causal structure learning-based approach to learn the network and recover causal effects among product purchases from basket-shopping data. To address the unique challenge of data sparsity in such datasets, we propose a Euclidean distance-based independence test instead of the partial correlation-based test in the PC algorithm for causal learning. We validate our approach and demonstrate its value by utilizing a large-scale basket-shopping dataset from Numerator. Our results show that CPNs more accurately represents interactions among products than other machine-learning or theory-based models, require significantly fewer parameters, and reveal new insights into consumer behavior such as that the brick-and-mortar channel exhibits denser causal connections among product purchases in a basket than the online channel. Finally, we propose a two-stage graph learning algorithm to improve computational efficiency and demonstrate that the constructed CPN, when applied to a multi-category assortment optimization problem, increases total sales by 8.26%–20.62% compared to a benchmark within-category model, viz., the multinomial logit model. These results make a significant advancement towards utilizing large-scale basket-shopping data for generating new managerial insights and optimizing decisions.

Corporate operational misconduct—also known as operational misconduct—refers to illegal, unethical, or intentionally improper behavior by managers that affects a corporation’s operational systems, which can influence service quality. However, fully understanding the causal relationship and mechanisms is challenging due to (1) much of operational misconduct remaining undetected internally or being concealed from the public, and (2) complex interactions among operational misconduct, service quality, and other factors (e.g., financial and inventory measures), which makes it costly to investigate these pairwise relationships through controlled experimental studies. To address these challenges, we introduce a text-mining approach to detect unobservable misconduct and a causal discovery approach to learn a causal Directed Acyclic Graph (DAG) that reveals the influence pathways from operational misconduct to service performance. We find that misconduct in workforce scheduling can both directly affect service quality and indirectly affect it through inventory levels. In contrast, misconduct in supply chain management has no direct impact but influences service quality indirectly via inventory levels. Additionally, financial measures act as mediators in the influence pathways. We conduct a counterfactual analysis showing that our DAG-based framework can be effectively utilized to enhance service quality. Overall, our study provides new managerial insights through novel measures of operational misconduct and a DAG-based causal framework, enabling companies to effectively enhance service performance by managing misconduct.

Many service firms attract customers and deliver experiences through firm-managed physical assets, such as plants in botanical gardens, animals in aquariums, and artworks in museums. Yet customers typically encounter these assets only as part of a broader service experience, without forming an ongoing connection to a particular asset. We examine whether symbolic adoption, an emerging service practice that links customers to existing service assets while firms retain ownership and operational responsibility, can increase customer demand. We conduct a randomized field experiment in collaboration with an ecological park. The sample includes 533 prior customers who voluntarily registered for a symbolic adoption program. We randomly assigned 266 applicants to receive the program and 267 to a control condition. Treated customers were linked to specific strawberry-plant units, whereas control customers did not receive access. Using a difference-in-differences analysis of weekly customer-level data, we estimate the effect of receiving the program relative to not receiving it among registered applicants. Assignment to the program increases weekly visits by 0.1242 visits per customer-week and log weekly spending by 0.1437 during active implementation. The spending effects extend beyond the asset-linked category: assignment to the program increases spending on strawberries and other product categories. These effects persist during the observed six-week post-program period after the active program and its associated touchpoints ended. Symbolic adoption can increase repeat visits and broaden on-site demand by linking customers to existing firm-managed assets. Firms can use this approach to engage customers and increase on-site spending without substantial investment in new physical assets, while retaining ownership and operational responsibility. An accounting analysis further suggests potential positive net value under the stated assumptions. 

As obesity remains one of the most prevalent health concerns in the United States, the government and other public sectors have increasingly invested in pharmaceutical R&D to address it. Beyond the goal of maximizing R&D output, the government has broader responsibilities, such as collecting accurate market information (e.g., private R&D cost) to support the design of effective subsidies and policy interventions for obesity treatment. Our study addresses this challenge by proposing a mechanism that incentivizes firms to truthfully reveal market information (as reflected in their private development costs), while preserving the government's payoff to the greatest extent. We design a funding mechanism in a static setting with multiple firms, in which truthful reporting is theoretically optimal for risk-neutral firms. However, firms do not always behave fully rationally in practice. To account for potential behavioral biases, we conduct laboratory experiments comparing our mechanism to two benchmarks: a threshold-based mechanism and a first-price auction-based mechanism. The experimental results show that firms report private information more truthfully under our proposed mechanism. We also provide behavioral explanations for these findings, shedding new light on behavioral biases in pharmaceutical R&D funding allocation. Overall, our paper proposes an effective mechanism for public sectors to elicit truthful information from pharmaceutical firms while preserving the public sector's expected payoff. The mechanism, supported by both theoretical and experimental evidence, also offers guidance for designing funding policies aligned with public health objectives.

Authorizing maintenance work orders requires information held by field evaluators about operational value and by suppliers about fulfillment costs. Suppliers may also offer inducements that encourage inflated classifications, linking the two agents’ reporting incentives. We develop a mechanism-design model in which an administrator jointly chooses authorization probabilities, documentation requirements, and supplier compensation subject to class-contingent cost coverage. We first solve a prescribedinducement auxiliary problem that provides an upper bound and characterizes when its optimum uses deterministic class differentiation, probabilistic authorization, or pooling. An auxiliary policy need not be implementable: class-contingent cost coverage can leave a lower-cost supplier with a profitable deviation that combines cost overstatement with induced classification inflation. We construct a feasible mechanism that addresses these deviations by reallocating authorization across classes and adjusting documentation while preserving, for every supplier cost type, both aggregate authorization and the supplier’s auxiliary payoff. Comparing this mechanism with alternative feasible constructions yields computable performance guarantees and sufficient conditions for attaining the auxiliary bound. The analysis further distinguishes documentation’s role in screening evaluator reports from its role in deterring supplier-financed inflation. Documentation can therefore be valuable for implementation even when the auxiliary optimum uses none. A proportional reduction in documentation and review costs, including the cost of supporting inflated reports, weakly reduces the administrator’s optimal value. Class-specific authorization can complement compensation and documentation by limiting the supplier’s incentive to induce classification inflation without increasing supplier information rents.