Knowledge that Transforms

Problem definition: The dedicated docket was introduced by the Biden Administration to reform the immigration system. It creates a separate queue for immigration proceedings where judges are supposed to issue a decision for each case within a target timeframe. The administration announced the docket with the goals of speed, accuracy, and fairness. Though the program meets its first goal, legal advocacy groups report that this comes at the expense of the last. Referring to it as a “Denial of justice”, they find that cases on the dedicated docket routinely fail to access legal representation, and have a much lower asylum grant rate. Against this backdrop, we aim to understand the operational implication of the dedicated docket. Methodology/results: We develop a stylized queueing model wherein a policy maker (PM) routes asylees to either the regular or the dedicated docket, and sets a delay target for the dedicated one. Constrained by the target, the court allocates its limited capacity to minimize the average delay. Immigration lawyers schedule their time between dockets to maximize the rate of successful asylum cases. Compared to a single docket, we show that the dedicated-docket system can Pareto-improve both the speed and accuracy. However, if the PM sets its decision poorly, the system may be dominated by a single docket. Moreover, we prove that the dedicated docket system can satisfy three natural fairness rules only if it is dominated by the single docket. Managerial implications: Our analysis can inform both policy makers and legal advocacy groups. We prove that the lack of fairness is a fundamental design flaw of the new program. Nonetheless, through delay differentiation, the program also enables a surprising efficiency gain. Policy makers and legal advocacy groups should be aware of such tradeoffs: the same delay differentiation that enables greater efficiency also creates unfairness between dockets.

Problem definition: In online-marketplace bargaining, how fast to respond to a buyer’s offer is a question many sellers face. Whereas a fast response signals the seller’s eagerness to sell, a slow response signals the seller’s insincerity or irresponsibility. In this paper, we use a unique dataset from eBay’s best-offer platform to study the effect of the seller’s response time on the buyer’s concession. Methodology/results: We first use a multiple instrumental variable (IV) regression to study the average effect of the seller’s response time. We find that the effect of the seller’s response time on the buyer’s concession first increases and then decreases (i.e., following an inverted-U shape). We then develop a new multi-treatment IV forest approach, which incorporates a multiple IV regression into a forest, to study the heterogeneous and nonlinear effects of the seller’s response time. We find that the effect of the seller’s response time varies widely across different items. Finally, we use the results from our empirical analyses to derive the best response time for the seller. We find that using the best response time leads to a sizeable increase in the buyer’s concession. Managerial implications: Our results are helpful to online marketplaces, because we find that the effect of the seller’s response time is nonlinear and heterogeneous across different items. Our results are also helpful to the seller, because we find the best response time that increases the buyer’s concession. Finally, our results are helpful to academia, because we develop a new causal machine learning method for analyzing the heterogeneous effects of multiple treatments with endogeneity issues.

Problem definition: This paper examines the influence of an artificial intelligence (AI) application – facial recognition technology at airports – on flight on-time performance. While facial recognition at airports has the potential to save time during check-in and boarding procedures, flight departures could be delayed due to recognition errors and system inaccuracies of this immature technology. Therefore, the impacts of facial recognition on flight on-time performance remain uncertain and require rigorous empirical investigation. Methodology/results: In this study, we exploit the first terminal-wide implementation of facial recognition in the U.S. and examine its impact on flight on-time performance. Our analysis of flight-level data reveals approximately a 16% reduction in departure delays and a 6% reduction in arrival delays but no increase in early departures or early arrivals. Interestingly, the improvement in on-time performance is smaller for flights to destinations in Asia and Africa, which tend to have a higher proportion of non-Caucasian passengers, but more pronounced for flights with larger seat capacity. Managerial implications: These findings demonstrate that implementing AI tools such as facial recognition can enhance operational efficiency and reduce flight delays on average. However, the magnitude of benefits varies across flight destinations and aircraft sizes. These findings offer valuable insights for firms that are considering the deployment of AI technologies for operational efficiencies.

Problem definition: This paper examines whether widely used online learning algorithms used in pricing can independently reach competitive outcomes or whether they may instead foster tacit collusion. This issue has drawn considerable attention from competition regulators, because algorithmic pricing is increasingly common in digital markets. Understanding when such algorithms lead to equilibrium prices or to supra-competitive prices is critical for buyers, sellers, and policymakers. Methodology/results: We study the behavior of multiarmed bandit algorithms in repeated price competition. These algorithms only observe profits from the prices actually chosen, making them realistic models of automated pricing. Using formal analysis, we show that an important class of online learning algorithms, called mean-based algorithms, reliably converges to the Nash equilibrium in Bertrand competition. This finding is notable because, in general, online learning algorithms do not guarantee convergence to equilibrium. In addition, we run extensive numerical experiments with different widely used bandit algorithms. The experiments confirm that most of them, including those that are not mean based, also converge to equilibrium. We observe supra-competitive prices only in special cases where all sellers implement the same symmetric version of certain algorithms, such as upper confidence bound. Even then, supra-competitive pricing vanishes as the number of competing sellers increases. Managerial implications: Our results highlight that the risk of algorithmic collusion in competitive pricing markets is often overstated. For most practical implementations of bandit algorithms, sellers’ prices converge to competitive levels. Only under very specific and symmetric setups do prices remain above competitive benchmarks, and this effect diminishes with more competitors. These insights provide reassurance to regulators concerned with consumer welfare, as well as to managers considering algorithmic pricing tools. They suggest that, although vigilance is warranted, fears of widespread algorithm-driven collusion may be exaggerated.

Problem Definition: Startup firms, often small in size, face the challenge of making cross-functional decisions due to the absence of distinct departments like marketing and operations. These interdependent decisions are further complicated by the lack of historical customer data. As a result, these firms must learn about customer preferences while making marketing decisions (such as pricing and advertising) and operational decisions (such as inventory levels). This paper examines a scenario where a firm jointly determines pricing, advertising, and inventory decisions over T periods while learning about demand and advertising response models. Methodology/Results: We first characterize a set of sufficient conditions to achieve exponentially fast learning rates with cross-functional decisions. Based on these sufficient conditions, we propose an easy-to-implement policy that is asymptotically optimal, showing that the gap between the profit of this policy and that of a clairvoyant with perfect information on the demand and advertising models is of order log T. Numerical experiments reveal that joint learning about the advertising and demand models is crucial for good profit performance. Managerial Implications: Our findings emphasize the need to jointly consider marketing and operational decisions for better business outcomes, providing valuable insights for entrepreneurs and managers. Our learning conditions highlight that the connections between cross-functional decisions play a key role in learning the impact of these decisions on consumers. Moreover, our numerical studies indicate that deviating from our learning conditions to avoid experimentation results in poor profit performance. Overall, our integrated approach can lead to better resource allocation, improved customer understanding, and increased profitability for firms.

Problem definition: We study how adding new vendors to the Massachusetts Healthy Incentives Program (HIP), a food subsidy program, affects program utilization. This is an instance of a core problem in causal inference: estimating heterogeneous treatment effects (HTEs) using panel data. Existing methods either do not utilize the underlying structure in the panel data or do not accommodate the general vendor addition patterns in the HIP data. Methodology/results: We propose the panel clustering estimator (PaCE), a novel method that first partitions observations into disjoint clusters with similar treatment effects using a regression tree and then, leverages the (assumed) low-rank structure of the panel data to estimate the average treatment effect (ATE) for each cluster. Our theoretical results establish the convergence of the resulting estimates to the true treatment effects. Computational experiments with semisynthetic data show that PaCE achieves superior accuracy for ATE and HTE estimation compared with existing approaches. This performance was achieved using a regression tree with no more than 40 leaves, making PaCE both more accurate and interpretable. Managerial implications: Applying PaCE to HIP data, we discern the heterogeneous impacts of vendor additions on program utilization across different regions of Massachusetts. These findings provide valuable insights for future budget planning and for identifying which Massachusetts zip codes to target with vendor additions. History: This paper was selected as part of the 1RR initiative between the M&SOM Journal and the MSOM Society. This paper was part of the 2025 MSOM Sustainable Operations SIG Conference.

Problem definition: In many real-world applications, such as cloud computing, there is a growing trend to supplement long-term reserved capacity with short-term on-demand capacity. We study a queueing system that employs both reserved and (relatively more expensive) on-demand servers. The number of reserved servers is decided at the beginning of the time horizon, while the number of on-demand servers is decided dynamically in real time. The objective is to minimize the costs incurred in hiring servers and in job waiting. Methodology/results: We establish that the optimal on-demand control is a threshold-based bang-bang policy: If the number of jobs in the system is below a threshold, no on-demand servers are employed. Otherwise, the number of on-demand servers is chosen such that no jobs wait. We present algorithms for obtaining an ε-optimal solution for the discounted-cost problem and an optimal solution for the average-cost problem. Furthermore, to address operational frictions, we extend our analysis to incorporate fixed switching costs and provisioning latency. We propose a periodic tracking policy and prove its asymptotic optimality. Managerial implications: Our analysis provides practical guidance for managing capacity and waiting costs in queueing systems that have access to both long-term and short-term server capacity. Our numerical experiments provide prescriptive insights on tailoring reserved and on-demand capacity to job urgency and volume.

Problem definition: Pivoting is a significant shift from the current product under development to an alternative opportunity. Pivoting is strongly advocated in the popular press and commonly used in entrepreneurial product development. However, a high fraction of startups that pivot still fail subsequently, since many entrepreneurs chase pivots that are ultimately not fruitful. We explore how an entrepreneur’s level of experience influences their decision to pivot. Experienced entrepreneurs, possessing a clearer perception of potential obstacles and opportunities, tend to exhibit reduced overconfidence. Methodology/results: We develop a decision-theoretic model to analyze the entrepreneurial choice regarding pivoting, while accounting for the overconfidence, and market and pivot opportunity uncertainties. We show that lack of experience, i.e., a higher level of overconfidence, can systematically influence the pivot decisions. The effect of overconfidence on pivot decision depends on the market uncertainty of the pivot opportunity relative to that of the already initiated product. More specifically, we show that novice (more overconfident) entrepreneurs are likely to make radical risky pivots, whereas experienced (less overconfident) entrepreneurs are more likely to make safe pivots. We find that the dominant development errors for inexperienced founders relate to pivoting and persisting when faced with risky and safe pivots, respectively. Furthermore, we show that if pivoting can be postponed until after the product launch, novice founders are much more likely to postpone safe pivots than experienced founders. We also explore the effects of different types of experience on the entrepreneur’s pivoting decisions. Managerial implications: Our results can benefit entrepreneurs in deciding how to set the duration of their supply chain contracts, choose the types of employees to hire, and select appropriate mentors, depending on the experience level of the founders and their likely pivot behavior. For startup incubators and accelerators, our results indicate what type of training programs to provide for entrepreneurs when accounting for their experience level.

Problem definition: Ride-hailing poses significant challenges to Original Equipment Manufacturers (OEMs) as it offers an affordable mobility option, and thus may lead to lower new car ownership. Yet, its more intense competition with the sales of used cars in secondary markets (another low-cost mobility option) may reduce the cannibalization of new car sales to consumers, and therefore benefit OEMs. Moreover, ride-hailing creates additional new car demand for OEMs from drivers providing service on ride-hailing platforms. With these complex interactions, the effects of ride-hailing on the car sales markets and associated environmental impacts are not clear. Recent practices also indicate that ride-hailing platforms may impose vehicle age limits to increase the quality of cars providing service. A vehicle age limit may increase the total new car sales as it increases the drivers’ new car purchase frequency; however, as it requires drivers to replace their used cars with new ones, it also creates another source of used car supply to secondary markets, and thus may increase the cannibalization of OEMs’ new car sales to consumers. Methodology: We establish a game-theoretic model to endogenize these interactions by accounting for the durable nature of cars and the effect of secondary market. Managerial implications: We show that ride-hailing can in fact lead to higher new car ownership among consumers and total new car sales for the OEM; it can also increase the total environmental impact, despite its perceived environmental benefit from usage pooling. Moreover, despite increasing the quality of ride-hailing cars, vehicle age limits can lead to higher profit and total new car sales for the OEM, but they can also reduce the total usage impact of cars.

Problem definition: Enabling market access is widely recognized as a priority for local governments aiming to reduce poverty among smallholder farmers. Traditional market access strategies, which connect farmers to wholesale intermediaries and reach consumers indirectly, have been criticized for exposing farmers to exploitation. Although it is commonly believed that enabling farmers to sell directly to consumers can mitigate this issue, farmers often rely on service intermediaries to facilitate direct sales and continue to face exploitation. Consequently, it remains unclear which type of market access strategy—indirect or direct—is more beneficial to smallholder farmers. We address this question by comparing representative strategies of the two types: contract farming, where farmers sell to a buying firm at a predetermined wholesale price, and rural livestreaming, where farmers sell directly to consumers via live broadcasts hosted by a media company that charges a percentage commission fee. Methodology/results: We construct game-theoretic models of the two strategies and compare farmers’ income in equilibrium. We show that, relative to contract farming, rural livestreaming can mitigate exploitation and improve farmers’ income for niche crops or crops with limited diseconomies of scale in planting, but may have the opposite effect for crops with mass appeal or steeply increasing marginal costs. Yield uncertainty can strengthen the relative advantage of rural livestreaming, whereas subsidies to farmers can weaken or even eliminate it. We validate our results via a case study of the market access strategies for smallholder farmers in Western China. Managerial implications: Policymakers should be mindful of the operational and market characteristics of local crops when choosing between direct and indirect market access strategies to improve farmers’ income. Moreover, coordination among poverty reduction instruments is important to avoid unintended consequences.