INFORMS 2021 Program Book

INFORMS Anaheim 2021

TE37

2 - Optimizing Machine Learning to Assist a Human Decision-Maker Bryce McLaughlin, Stanford University, Stanford, CA, United States Many decision systems have begun to incorporate predictions built off of large quantities of past data in an attempt to improve the quality of the decisions. In many such applications, machine-learning algorithms are trained to take an optimal decision based on this data. However, in high-stakes applications such as judicial decisions, medical diagnosis, or fraud detection many decisions will continue to be taken by humans, with machine predictions merely providing one of many inputs. In this work, we therefore ask how machine-learning predictions should change when they are designed to support, rather than replace, a human decision-maker who holds the decision authority. Our work aims to move the focus on AI-driven decisions from substituting human decisions to complementing them. A general take-away is that the optimal solutions to these problems are strikingly different, with optimal assistance focusing on instances that the human gets wrong, while optimal substitution focuses on cases humans typically also get right. As a specific take-away, we provide theoretical results that can guide the implementation of machine-learning algorithms in high-stakes decisions where human decision-makers have the decision authority. TE37 CC Room 210C In Person: Empirical Healthcare Operations Management General Session Chair: Masoud Kamalahmadi, University of Miami, Bloomington, IN, 47405-1701, United States 1 - A Field Experiment on Wait Time Information Provision Danqi Luo, Stanford University, Stanford, CA, 94305-7216, United States, Mohsen Bayati, Erica Plambeck In an ongoing field experiment, we trial three different wait time information provision schemes to low-acuity patients (LAP), patients with ESI level 3, 4, and 5. Through an incentivized text-based survey, patients can electronically self- report their real-time satisfaction on wait time and pain level throughout their stay in the ED. Matching patients’ responses with their electronic medical records (EMR) and the NRC health data (a survey collected by SMMC), we can measure the impact of different wait time information on patients’ waiting satisfaction, outcomes concerning the behavior of left-without-being-seen by a physician, length of the stay in the ED, and pain level. In the first stage of results, we identified that LAPs are less likely to leave the ED without being seen by a physician compared to the baseline when no information is provided. 2 - Nudging Patient Choice by Messaging: A Field Experiment Jiayi Liu, Emory University, Atlanta, GA, 30322, United States, Diwas S. KC To examine the drivers of patient no-shows at outpatient clinics, we conduct a series of field experiments where the messaging regarding their upcoming appointment is randomly assigned. We find that the type of messaging has a significant effect on the queuing behavior of individuals, most notably their no- show behavior. TE39 CC Room 211A In Person: Doing Good with Good OR: I Award Session Chair: Paul Gölz, Pittsburgh, PA, 15206, United States 1 - Learning, Optimization, and Planning under Uncertainty for Wildlife Conservation Lily Xu, Harvard University, Cambridge, MA, United States In collaboration with conservation NGOs, our project helps plan effective ranger patrols to protect endangered animals from poaching. Algorithmically, the problem is to optimize limited resources to maximize the number of snares confiscated. Given limited and incomplete data, we leverage linear programming, multi-armed bandits, and game theory to handle uncertainty about poacher behavior. Our approaches are supported with theorems, experiments, and real- world field tests. Our system is being integrated into existing conservation software to become available to 800 protected areas worldwide. 2 - Data-Driven COVID-19 Vaccine Development for Janssen Michael Lingzhi Li, Massachusetts Institute of Technology, Boston, MA, 02111, United States The COVID-19 pandemic has spurred extensive vaccine research worldwide. Phase III vaccine trials’ success highly depends on future COVID-19 incidence rates at trial sites. To accurately predict these rates, we created DELPHI, a novel data and policy driven epidemiological model. DELPHI is the centerpiece of site selection for the Phase III trial of Ad26.COV2-S, the leading Janssen vaccine

candidate. DELPHI-driven site selection accelerated the trial by 6-8 weeks while reducing the necessary size from 60k to 45k individuals, allowing millions of people earlier access to a life-saving vaccine. 3 - Fair Algorithms for Selecting Citizens’ Assemblies Paul Gölz, Carnegie Mellon University, Pittsburgh, PA, 15206, United States Globally, there has been a recent surge in citizens’ assemblies, which are panels of randomly-selected citizens weighing in on a policy question. Since these panels must proportionally represent many demographic groups, the selection algorithms currently used for choosing panels select different agents with highly unequal probabilities. We develop selection algorithms that satisfy quotas while choosing pool members with probabilities as close to equal as possible. We have implemented one such algorithm, which has been adopted by a number of organizations around the world. TE40 CC Room 211B In Person: Forecasting, Ordering and Allocation under Strategic Behavior General Session Chair: Minseok Park, PhD, Salisbury University, Salisbury, MD, 21801, United States 1 - Spatial Information-sharing on On-demand Service Platforms: A Behavioral Examination Swanand Kulkarni, Georgia Institute of Technology, Atlanta, GA, 30318, United States, Basak Kalkanci We examine how the spatial characteristics of demand-supply mismatch information sharing influence a platform’s matching efficiency. We compare public information sharing mechanisms (where all drivers have the same information) with local information sharing (where only nearby drivers learn about a surge opportunity). We examine drivers’ relocation decisions under competition theoretically and test predictions experimentally. Experiments reveal that local information sharing can be highly effective despite being dominated theoretically. 2 - Accountability and the Starvation Cycle in the Nonprofit Sector Iman Parsa, Arizona State University, Tempe, AZ, 85287, United States, Mahyar Eftekhar, Charles J. Corbett Donors to nonprofit organizations are sensitive to overhead expenses and support charities with lower administration and fundraising costs. This can lead to under- investments in essential organizational and operational infrastructure, or misleading financial reports, and an increasing expectation of donors for low overhead costs. In this paper, we empirically investigate whether the IRS policy change in 2008 has helped in breaking this cycle by providing additional information about nonprofits’ accountability. This empirical study is based on a large dataset containing information of nonprofits in different sectors in the time period 2009-2017. TE41 CC Room 212A In Person: Decision Analysis Game Theory Applications, Stochastic Dynamic Programming, and Public Policy General Session Chair: Manel Baucells, University of Virginia, Charlottesville, VA, 22903-1760, United States 1 - Optimal Incentives to Mitigate Epidemics: A Stackelberg Mean Field Game Approach Gokce Dayanikli, Princeton University, Princeton, NJ, United States, Alexander Aurell, Rene A. Carmona, Mathieu Lauriere Motivated by the models of epidemic control, we consider a Stackelberg mean field game between a principal and a mean field of agents whose states evolve in a finite state space. The agents play a noncooperative game in which they control their transition rates between states to minimize an individual cost. The principal influences the Nash equilibrium through incentives to optimize its own objective. We show an application to an epidemic model of SIR type in which the agents control their contact rates, and the principal is a regulator acting with non pharmaceutical interventions. We propose a numerical approach based on Monte Carlo simulations and machine learning tools for stochastic optimization.

148