Relational Continuity (RC) in primary care confers many reported benefits, yet it has been in sharp decline. Using multiple econometric techniques on a large consultation-level dataset comprising of 5M patients registered with 300 primary care practices across the UK over the course of 10 years, we find that RC has a significant productivity benefit, with operational and strategic implications for primary care practices and third-party payers.

The rising demand for Emergency Department (ED) care is partially driven by the failure to provide timely and high-quality primary care, resulting in patients being forced to use EDs. It is therefore important to identify primary care practices (PCPs) whose patients place a lower burden on ED departments so that best practices can be identified and disseminated, and practices whose patients place a higher burden in order to provide support. This work develops and validates one such methodology.

We examine how employer-driven volatility in workers' schedules impacts their decision to voluntarily leave their job. Using time-stamped work log data of home health nurses, we construct and study an operational measure of schedule volatility. Using an instrumental variables approach, we find that higher levels of schedule volatility substantially increase workers' likelihood of quitting. Using policy simulations, we illustrate how schedule volatility, and employee turnover, could be mitigated.

Using a data set assembled from detailed ED visit-level data and exhaustive billing data in an integrated health system, we show that there is an increasing concave relationship between ED physician workload and post-ED care use. Further, we identify ED physician test ordering behavior as a mechanism of these effects. Together, these findings suggest that when ED physician workload increases, resource utilization increases in the ED and several other channels of care in the healthcare system.

The COVID-19 pandemic has necessitated widespread deferrals of elective surgeries, which non only increase the cost from deterioration of patients' conditions but also decrease the revenue. We develop a robust optimization framework for the dynamic surgical capacity management problem. Our Implementation for more than 15,000 hernia patients in the United States demonstrates sizable improvements over alternative methods by up to 10% with multiple practical insights on optimal expansion strategy.

We propose and implement a rigorous framework and algorithms to quantify the merits of targeted confinement interventions for pandemic response, and demonstrate them in a case study of COVID-19 in Île-de-France. We find that optimized interventions that differentiate based on both population groups and activities are interpretable, and achieve significantly better health and economic outcomes, while also reducing confinement time for each group, compared to less targeted interventions.

In many healthcare settings, patients receive a series of services during a single visit, e.g., infusion care and orthopedic visit. A key commonality is the tandem structure where each stage involves a non-trivial random service time. We develop the first analytic model to provide each patient an individualized visit itinerary in a tandem health service system. A case study populated by data from a large infusion center shows that our approach makes a remarkable 27% cost reduction over practice.

Delays in admission to rehabilitation care can be both capacity-driven and/or due to processing delays. Standard data however only includes a single measure of delay, and the bed allocation decisions in practice do not follow a systematic policy. We propose a hidden Markov model to estimate the processing times and the status-quo bed allocation policy. We validate our structural model and conduct counterfactual experiments to evaluate various operational interventions aimed at reducing delays.

Hospitals are rife with alarms, many of which are false. This leads to alarm fatigue, in which clinicians become desensitized and may inadvertently ignore real threats. We develop a partially observable Markov decision process model for recommending dynamic, patient-specific alarms. We find that compared to current approaches of setting patients’ alarms, our dynamic patient-centred model significantly reduces the risk of patient harm.

Split liver transplantation (SLT) can save two lives using one liver. To facilitate increased SLT usage, we formulate a multi-queue fluid model, incorporating size matching specifics, dynamic health conditions, transplant type, and fairness. We find the optimal organ allocation policy, and evaluate its performance versus other common allocations.

We study the interaction between a seller and a buyer, both of whom face uncertainty related to downstream demand, over a two-period horizon. Making multi-unit purchase decisions before demand from their respective lower tier realizes, both agents seek to minimize their demand mismatch risk as perceived according to their degree of foresight. Focusing on the effect of buyer foresight, we find that with multi-unit purchases sellers benefit from some degree of buyer foresight but not too much.

Behavioral aspects of consumer return policy design and their interaction with other decisions in retailing have not been investigated to date. Leveraging a generalized newsvendor model, we conduct a randomized experiment to assess how subjects jointly make key decision (order quantity, price, and refund amount) and the effect of salvage value on them. We identify time-dependent behavioral regularities that we explain through a process theory, thus providing a new direction for future research.

Actual inventory can be lower than recorded system inventory due to shrinkage or loss. To handle inventory inaccuracy, managers can decide to clean inventory data before placing order quantities. Alternatively, they can deliberately decide to not clean, but to compensate for the inaccuracy in the order decision. The optimal decision depends on the cost of cleaning and the efficiency loss due to the compensation. We present a set of hypotheses on this trade-off and test these in lab experiments.

How the cost of completing a queue varies with (i) the experience of wait and (ii) the characteristics of the queue are not well understood. In this study, we use the incentive-compatible BDM mechanism to experimentally address these questions. We find that when service speed is slow, experienced wait increases (decreases) the completion cost of impatient (patient) subjects. Also, the length and speed of a queue affect completion costs, but not proportional to their effects on total waiting time

Through a series of experiments, we show that when others are waiting in line, customers accelerate their own service time, sacrificing their own consumption utility. This behavior is driven by concern for others. We show that the negative effect of others queueing on one’s own service time is moderated by the participants' self-wait and visibility between customers in service and those waiting in line.

Contrary to traditional queueing theory, recent field studies in health care and call centers indicate that pooling queues may not lead to operational efficiencies relative to dedicated queues. We use a series of experiments to examine the conditions under which this may be the case and to test servers' customer orientation as a behavioral mechanism that may explain why. We also examine whether higher levels of customer orientation and performance persist across changes in queue configuration.

The algorithm-augmented business model promises service companies the benefits of both algorithms and humans. But companies will only realize this promise if their workers rely on algorithms, and there is conflicting evidence about workers’ willingness to do this. We design a laboratory experiment to resolve this conflict, and find that workers are generally unwilling to rely on algorithms but that they become more willing to do so in response to high customer loads and learning interventions.

We conduct a series of online experiments to isolate the pure effects of three types of Relative Performance Feedback, RPF, on middle-ranked workers' performance. We find that providing any type of feedback reduces performance compared to no feedback. Aligned with theory, delivering feedback increases the focal employee's shame and social comparison involvement (SCI), which measures the focal individual’s level of engagement in social comparison while performing the task.

Many human-machine interactions focus on the optimization of the system output yet tend to overlook human behavior. Using an intervention-based field experiment in a semi-automated warehouse, we study the impact of a behavioral intervention that provides humans with more information about the picking process and enables them to choose out of a set of pre-defined goals. We find that human performance is enhanced by 6%. Our insights enrich the discussion on human-machine interactions.

In this project, we study the fairness implications of using machine learning to assist a human decision-maker. Relative to a baseline where machine decisions are implemented directly, we show in a formal model that the inclusion of a biased human decision-maker can revert common relationships between accuracy and fairness. Specifically, we document that excluding information about protected groups from the prediction may fail to reduce, and may even increase, ultimate disparities.

In many practical settings, a human reviews recommendations from a decision support algorithm and either approves or adjusts the recommendation. Automation may reduce a ML system's longer-term ability to predict effective adjustments and leads to predictive performance degradation over time. We (empirically) demonstrate this effect and show how to include the loss of learning into the automation decision.

We provide a stylized model in which a principal chooses a classifier, D, with known properties for a Bayesian decision-maker who observes the outcome of D before determining their own label in a binary classification problem. The decision-maker has a utility which deviates from the principal ’s whenever they take an action which contradicts the classifier. We characterize the optimal posterior decision and show how the optimal classifier for assistance depends on the decision-maker's prior.

Using 50K observations of purchase tasks, we examine the mechanism behind purchasers’ decisions on the timing of the orders. Our analysis focuses on the impact of exposure to suppliers and rare events throughout the history of experiences of the purchasers. We find that exposure to specific suppliers increases the safety time for the current order; rare events in the prior orders are increasing the safety time. We also observe that purchasers rely on self-experienced orders rather than peers.

In recent applications of Markov decision processes, it is common to estimate model parameters from data. When data are collected from a population, one faces a modeling question of whether to estimate different models for subpopulations. This work provides theoretical results and empirical methods for making the decision of whether to model subpopulations separately or not. We also present how to use our results to select the best stratification and empirical results using various instances.

An action associated with a small cost is commonly taken by humans for preventing a potential big loss in a wide range of domains. For example, technicians decide whether to do preventive maintenance now or to wait for the next updated machine status information. We model such decisions as a dynamic cost-loss game. We analytically show that using a probability threshold is an optimal policy. We next conduct a controlled laboratory observing how human decision makers behave in this environment.

Resource-based contracts fail to motivate suppliers to provide reliable services as they are paid for after-sales services. Performance-based contracts shifts much of the downtime risk to the supplier by making him responsible for machine uptime, but then customers might reduce care efforts. We propose the full-care contract to achieve both high reliability and care, and show that the FCC not only the best contract analytically but also can motivate risk averse suppliers to bear the chain risk.

Using a behavioral game of supply rationing, we investigate the consequences of sharing information about competing retailers’ historical orders on the Bullwhip Effect. We find that decision makers act more strategically and closer to Nash equilibrium predictions if information about competitors’ historical orders is shared; however, sharing the entire order history does not accelerate the convergence to theoretical Nash equilibrium as much as sharing orders from the last period only.

Contact centers face operational challenges - proxies for customer experience are subject to uncertainty. A main source is silent abandonment customers (Sab). Sab leave while in queue with no indication. As a result, capacity is wasted. We develop methodologies to identify Sab and to estimate patience. We show how accounting for Sab in a queueing model improves the estimation accuracy of key measures of performance. Finally, we suggest strategies to operationally cope with Sab.

In the service industry, it is common that companies send agents to perform tasks at customer locations. Thereby, many companies rely on their agent's ability to construct tours manually. These manual tours are often non-optimal, causing high travel cost. In our work, we developed a new agent-to-customer assignment approach, designed to promote the manual construction of distance minimal tours. In our experimental study we investigate its effect on the human routing performance.

We study how consumers decide to pay with loyalty points or money by developing a model and estimating it on airline loyalty program data. We find that mental accounting, subjective perceived value, and reference exchange rate of points play important roles, and consumers’ primary points earning source and total earning level are jointly associated with their attitudes toward points and money. We show how a firm can implement pricing policies to efficiently influence consumers’ payment choices.

Agile project management techniques have become commonplace in many organizations. We experimentally examine how these techniques affect performance in two innovation settings: (1) a design setting and (2) a search setting. Our results caution against uniform adoption of the Agile approach, and suggest that the choice of the approach should depend on the nature of the project and on the risk appetite of the project manager.

We study multiple parallel innovation contests where contest organizers elicit solutions to a set of problems from solvers. Prior theoretical work shows that if problem solution requires less novelty then organizers should restrict participation with one contest. We experimentally show that this may not hold true in practice due to the heterogenity of solver efforts and encouraging solvers to participate in multiple contests can yield a larger organizer profit even for less novel problems.