DISC Research Seminar Series

TBD

TBD

TBD

Little is known about how people-centric factors affect the shape of service time distributions, despite distributional statistics (variance or quantiles) being key drivers of system performance in many service industries. We investigate the impact of three people-centric factors—worker experience, fatigue, and rest—on the average, variance, and quantiles of service times in paramedic operations. Our analysis uses data on the performance of 368,634 paramedic teams in the London Ambulance Service over 10 years.

Herding cues have traditionally been understood as factors that promote herd behavior in online marketplaces. However, our study challenges this conventional belief by proposing that herding cues can sometimes have the opposite effect: inducing contrarian behavior, where individuals intentionally go against the prevailing trend. We begin by defining contrarian behavior in online marketplaces and suggesting that customers' decision to engage in contrarian behavior is influenced primarily by their perception of the wisdom and irrationality of the herd, as well as their desire for uniqueness. Perceived wisdom and perceived irrationality, in turn, are positively affected by online herding cues (herd size and herd dominance). Furthermore, we suggest that the relationship between herding cues and contrarian behavior is non-linear, as these cues exert different effects on the perceived wisdom and irrationality of the herd. To test our hypotheses, we employ a multi-method approach. Study 1 utilizes archival data from an online service platform to reveal the existence and extent of contrarian behavior and how herding cues may affect customers’ contrarian tendencies. Study 2, on the other hand, employs a controlled experiment to investigate the psychological mechanisms underlying contrarian behavior. Results from the studies attest to our research model. Findings from this research offer important implications for both research and practice.

The mixed logit model, widely used in operations, marketing, and econometrics, represents choice probabilities as mixtures of multinomial logits. This study investigates the effectiveness of representing the mixed logit as a single-hidden-layer neural network, which approximates the mixture distribution with an equally weighted distribution over a finite number of consumer types. Despite its simple architecture, the model's statistical and computational properties have not been thoroughly examined. From a statistical perspective, we demonstrate that the approximation error of the neural network does not suffer from the curse of dimensionality, and that overparameterization does not lead to overfitting when proper regularization is applied. From a computational perspective, we prove that the noisy stochastic gradient descent algorithm can find the global optimizer of the entropy-regularized non-convex parameter learning problem with a nearly optimal convergence rate. Experiments on synthetic and real datasets validate our theoretical findings, highlighting the potential of overparameterized neural network representations, coupled with efficient training algorithms, to effectively learn choice models with strong performance guarantees.

We provide experimental evidence of the impact of peer-driven knowledge sharing on skill production in K-12 classrooms. We designed a technology platform that identifies individual students’ knowledge gaps and organizes team contests inside classrooms to induce knowledge sharing among team members. In a large-scale field experiment on over 3000 students in 86 Grade 3-6 math classes, we randomly assigned classes into team classes (students work on the platform in teams), individual classes (students work on the platform individually), and control classes (students don’t work on the platform). Our estimates indicate that students in team classes obtain 0.25σ and 0.17σ higher cognitive scores than those in control and individual classes. We observe cognitive skill improvements across team students of all abilities – high, medium, and low –indicating that high-ability students benefit and are not hurt by working in teams with their lower-ability counterparts. We also provide empirical evidence that feedback on student knowledge gaps to team members and knowledge sharing among team members are the underlying reasons for higher academic achievements in team classes. We also measure students’ social preferences for generosity, reciprocity, and trust toward fellow students. Our estimates reveal that competing in teams reduced team students’ social preferences for students in other teams but remained unchanged for teammates. Our study provides evidence that carefully designed peer interactions with the help of technology can significantly improve skill production in K-12 classrooms.

Although food insecurity remains a critical issue in developing countries, an average of 30–40% of food is wasted during the production stage of the supply chain, with a significant portion occurring in restaurants and hotels. This study pro- poses a simple, practical, and low-cost solution to reduce food waste in restaurants. In a six-week randomized controlled field experiment conducted across 71 restaurants in Ghana, we tested the effectiveness of two interventions–one driven by public-interest motivations and the other by private-interest motivations. Kitchen workers in the treatment restaurants received these interventions through posters in kitchen areas and periodic SMS messages. Both interventions significantly reduced daily food waste during meal preparation, with the private-interest intervention achieving a 19% reduction and the public-interest intervention resulting in an 9% reduction. This finding is striking, given that Ghana is a collectivistic, consensus-oriented society with strong communal values and a deep commitment to caring for others. Further analyses reveal that this result can also be attributed to cultural factors (e.g., power distance), socioeconomic elements (e.g., workers’ tenure and income), and the relationship between workers and the restaurant manager.

Digital platforms rely on user interaction and engagement often through user created content, e.g., in online Q&A forums like Reddit, but the proliferation of uncivil content on these platforms that happens in parallel adversely impacts safety and community. We examine the interplay between incivility and engagement. We theorize that based on emotional arousal and cognitive dissonance, uncivil content can increase user engagement due to emotional arousal, however, its effectiveness varies with the level of incivility due to differences in the level of cognitive dissonance it evokes. Our work suggests that understanding the relationship between incivility and engagement requires a more nuanced definition of incivility than the binary categorization (content is either uncivil or not) used in the IS literature. More specifically, we examine how three levels of incivility in an initial post affect the subsequent quantity, novelty, and incivility of the subsequent content. Leveraging a unique dataset from the largest online community in South Korea, we categorize incivility (extreme, mild, or no incivility) using large language models and estimate their impact on engagement using causal forests. We find that extremely uncivil content, which we posit triggers stronger emotional arousal but also potentially stronger cognitive dissonance, leads to a notable decrease in the novelty of subsequent content. Conversely, mildly uncivil content, by eliciting emotional arousal without substantial cognitive dissonance, promotes a more diverse and novel range of user comments. Still, both levels of uncivil content (extreme and mild) are linked to an increased subsequent extreme incivility and a decreased no-incivility discourse. This outcome underscores the interplay between emotional arousal and cognitive dissonance in shaping user reactions to uncivil content and highlights the importance of exploring incivility beyond a simple binary measure.

Information technologies are vital for addressing the ills affecting healthcare in the US, including high costs, variable outcomes, and significant inequities. One category of transformative technologies is price transparency and decision support tools for prescription drug choice, which currently -operates under a fog of price opacity. This leads to avoidably high spending, even for competitive generic drugs, non-adherence to medication, and poor health outcomes. Can tools that provide timely price information to decision-makers mitigate these problems? We present findings from a widely implemented drug-cost decision-support system at prescribing (DDP) that, within the prescription workflow, provides out-of-pocket cost information and identifies cost-effective, therapeutically equivalent alternatives and fulfillment options reflecting the patient’s benefits status and pharmacy coverage. Using transactional and claims data covering 229,193 prescribers, 1.166 million patients, and 6.504 million visits during 2023, we find evidence of positive impact: lower-cost alternative drugs were identified 24.824% of the time, but with an average switch rate of just 3.698% (after excluding simple same-drug brand-to-generic switches that typically occur at the pharmacy counter). These switches reduce total costs (payer plus patient) by 15.475% relative to an opportunity of 30.861% (for patients’ share, the savings opportunity was 61.982%, with 20.256% realized). Switch behavior varied significantly depending on the characteristics of the medication and disease, the prescriber’s attributes, and recommendation scenarios. We discuss research directions for deeper investigation of these relationships and better designs that can generate greater effectiveness, savings, and improved health outcomes.

Public Generative AI (PGAI) is a class of information technologies that has been developed to learn and generate text, code, arts, and other forms of human artifacts. Due to their public nature, access to and use of these technologies are democratized, removing the usual scale and size barriers to adopting valuable resources. However, based on the classical resource-based view, public Generative AI resources cannot be a source of competitive advantage as competitors and the focal firm could access them with comparable ease. Against this prediction from classical resource-based view, we argue that PGAI can be a source of competitive advantage. This study puts the predictions of classical RBV to the test. Utilizing data from 2,496 public firms, we find that firms with strong use cases for public Generative AI applications boast over 3 percent of value-added. Nonetheless, these competitive gains in the form of value-added are not uniform and depend on the firm’s ability to move beyond simple integration of the public Generative AI resources in day-to-day activities. Gains depend heavily on the firm’s ability to adapt, re-train, and tightly fit these digital resources into the contours of its operation. Because AI resources, like live organs that are transplanted into a host’s body, can re-shape and re-adapt to the firm’s history-dependent idiosyncrasies, we call the ability to adapt, re-train, and tightly fit public Generative AI resources to the firm the resource transplanting ability and contrast it with the better-known resource integration abilities. We show that firms with better input-feed feasibility and higher presence of generative use cases are more successful to turn Generative AI resources into competitive weapons. Our study also shows that the transplanting process makes the adapted and retrained Generative AI resources firm-specific and immovable, therefore rendering those resources a source of negative value externalities, unlike the classical IT resources that have been, by and large, a source of undisputed positive externalities.

-=- Cancelled! Please note that this Research Seminar event is cancelled due to unforeseeable reasons. Contact Professor Xiao Ma if you have any questions -=- The fast penetration of smart mobile and sensor technologies, combined with the wide adoption of location services, has produced a vast volume of atomic, behavior-rich mobile consumer location data. Such “location big data” have led to the pervasive digitization of individual behavior at a very fine-grained level. This information provides us with a new lens through which we can better monitor, understand, evaluate, and optimize the individual decisions and policy making to improve economic welfare. This talk presents some recent studies in which we apply novel machine learning, spatial-temporal data mining, econometrics and randomized experimental methods to discern the socio-economic value and tradeoffs in learning from location data.