Research

Abstract: The adoption of online services, such as telemedicine, has increased rapidly over the last few years. To better manage online services and effectively integrate them with in-person services, we need to better understand customer behaviors under the two service modalities. Utilizing data from two large internal medicine outpatient clinics, we take an empirical approach to study service incompletion for in-person and telemedicine appointments respectively. We focus on estimating the causal effect of physician availability on service incompletion. When physicians are unavailable, patients may be more likely to leave without being seen. We introduce a multivariate probit model with instrumental variables to handle estimation challenges due to endogeneity, sample selection, and measurement error. Our estimation results show that intra-day delay increases the telemedicine service incompletion rate by 7.4%, but it does not have a significant effect on the in-person service incompletion rate. This suggests that telemedicine patients may leave without being seen when delayed, while in-person patients are not sensitive to intra-day delay. We conduct counterfactual experiments to optimize the intra-day sequencing rule when having both telemedicine and in-person patients. Our analysis indicates that not correctly differentiating the types of incompletions due to intra-day delays from no-show can lead to highly suboptimal patient sequencing decisions.

Abstract: Remote patient monitoring (RPM) enables providers to observe patients’ health signals outside the clinic and to intervene proactively when deterioration is suspected. In practice, however, wearable sensors can be imperfect, and unnecessary interventions can impose logistical and psychological burdens that erode patient compliance. We study a dynamic scheduling problem in which a provider monitors a panel of patients, updates probabilistic health beliefs from imperfect RPM signals, and decides whom to intervene under limited capacity. Proactive intervention reduces the risk of deterioration and improves short-term health outcomes, but over-intervention induces patient fatigue, increasing future noncompliance and undermining long-run health outcomes. We formulate the problem as a Partially Observable Markov Decision Process (POMDP) and the objective is to maximize quality-adjusted life years (QALYs). We characterize structural properties of the optimal value function and establish threshold-type optimal policies. In particular, in the single-patient setting, it is optimal to intervene when the health belief crosses a fatigue-dependent threshold, and the intervention region expands as fatigue increases. Under capacity competition, the optimal policy is governed by a priority threshold where higher fatigue reduces priority. For larger capacities, we propose a tractable Lagrangian relaxation and truncation policy with a provable performance bound. Numerical experiments quantify the value of RPM information and show that ignoring endogenous fatigue can substantially degrade performance, in magnitude comparable to a large worsening in sensor accuracy.

Abstract: Sunk-cost bias occurs when decisions are influenced by the time, energy, and money already invested, rather than considering the future costs necessary to achieve success. This phenomenon of "irrational behavior" is well-documented in decision-making studies and is generally recognized as a factor that can lead to suboptimal decisions. In this work,  we investigate how sunk cost (and the behavioral bias associated with it) can be used as an operational lever to increase service completion rates in a congested service system. We run a controlled online experiment and find that the abandonment rate is significantly reduced for the group of participants who incur a larger sunk cost. To better capture the dynamics of service systems and their impact on customers' behavior, we study a queueing model with sunk cost and strategic customers, where customers experience a disutility of balking that is proportional to the sunk cost they incur. We characterize the equilibrium behavior of the customers, from which we further derive the optimal strategy for the service provider in terms of whether to provide real-time queue length information to customers as well as the optimal level of sunk cost to impose. Our results show that the sunk cost strategy is effective only when waiting information is provided and that using a non-zero sunk cost is optimal when the queueing system is moderately congested. Through a comprehensive numerical study, we demonstrate that implementing a non-zero sunk cost can substantially improve the throughput of the system. In addition, we reveal an interesting asymmetric pattern in the robustness of the service provider's optimal policy when the customers' sensitivity to sunk cost cannot be accurately estimated, which suggests that if the service provider cannot accurately estimate the customer’s sensitivity to sunk cost, using an underestimated value will give more robust performance improvements.

Abstract: An increase in the novelty of news predicts negative stock market returns and negative macroeconomic outcomes over the next 12 months. We quantify the novelty of news -- changes in the distribution of news text -- through an entropy measure, calculated using a recurrent neural network applied to a large news corpus starting in 1996. A one standard deviation increase in this measure forecasts a 3% annual decline in the market. In out-of-sample regressions, entropy provides more informative forecasts than a wide range of standard predictors, including text-based ones. In cross-sectional regressions, exposure to entropy carries a negative risk premium, suggesting that assets that positively covary with entropy hedge the aggregate risk associated with shifting news language. We show that entropy cannot be explained by the multitude of existing long-short portfolio factors. Our analysis uses over 1.6 million news articles over a period of 27 years.

Abstract: The global pandemic caused by coronavirus (COVID-19) sped up the adoption of telemedicine. We aimed to assess whether factors associated with no-show differed between in-person and telemedicine visits. The focus is on understanding how social economic factors affect patient no-show for the two modalities of visits. A mixed-effect logistic regression was used. We performed stratified analysis for each modality of visit and a combined analysis with interaction terms between exposure variables and visit modality. Patient demographics (age, gender, race, income, partner), lead days, and primary insurance were significantly different between the two visit modalities. Our multivariable regression analyses showed that the impact of sociodemographic factors, such as Medicaid insurance (OR 1.23, p < 0.01 for in-person; OR 1.03, p = 0.57 for telemedicine; p < 0.01 for interaction), Medicare insurance (OR 1.11, p = 0.04 for in-person; OR 0.95, p = 0.32 for telemedicine; p = 0.03 for interaction) and Black race (OR 1.36, p < 0.01 for in-person; OR 1.20, p < 0.01 for telemedicine; p = 0.03 for interaction), on increased odds of no-show was less for telemedicine visits than for in-person visits. In addition, inclement weather and younger age had less impact on no-show for telemedicine visits. Our findings indicated that if adopted successfully, telemedicine had the potential to reduce no-show rate for vulnerable patient groups and reduce the disparity between patients from different socioeconomic backgrounds.