Publications

Online reviews serve as a guide for consumer choice. With advancements in large language models (LLMs) and generative AI, the fast and inexpensive creation of human-like text may threaten the feedback function of online reviews if neither readers nor platforms can differentiate between human-written and AI-generated content. In two experiments, we found that humans cannot recognize AI-written reviews. Even with monetary incentives for accuracy, both Type I and Type II errors were common: human reviews were often mistaken for AI-generated reviews, and even more frequently, AI-generated reviews were mistaken for human reviews. This held true across various ratings, emotional tones, review lengths, and participants’ genders, education levels, and AI expertise. Younger participants were somewhat better at distinguishing between human and AI reviews. An additional study revealed that current AI detectors were also fooled by AI-generated reviews. We discuss the implications of our findings on trust erosion, manipulation, regulation, consumer behavior, AI detection, market structure, innovation, and review platforms.

As a board-certified neurosurgeon, Dr. Ryan Grant believed that the medical system treating chronic musculoskeletal (MSK) pain was inherently broken. He believed that he could challenge the way traditional medical institutions treated back pain and build a profitable and sustainable company, even in a tough funding environment. His goal was audacious, but as a successful healthcare entrepreneur with multiple start-ups, Grant was used to winning.

MSK pain was one of the most widespread health conditions, with major impacts on the lives of many individuals as well as their employers and society as a whole. According to the World Health Organization, approximately 1.7 billion people had musculoskeletal conditions. Low back pain was the single leading cause of disability, estimated to affect up to half of all Americans at some point in their lives.

However, according to recent research, the traditional treatment protocol for MSK pain was deficient. The surgical option was overused; multiple studies had demonstrated that less invasive and expensive treatments could often be just as effective as surgery. Treatment was fragmented, requiring patients to visit multiple health practitioners, who often had limited interactions with each other.

In 2020, Grant founded Vori Health along with Dr. Mary O’Connor, a new clinical service company that he believed would help patients with MSK achieve their quality-of-life goals while avoiding costly and unnecessary surgeries. Vori Health devised a system through which clinicians would develop personalized evaluation and treatment plans based on patient-defined needs and then deliver non-surgical therapies to reduce pain levels. A doctor-led team that integrated experts in physical therapy, nutrition, and other lifestyle support options would support each patient’s treatment plan. Vori Health services were primarily delivered virtually, although in-person services were also available.

In 2021, Vori raised over $68 million in Series A financing. Preliminary results for patients had been encouraging and the company benefited from the trend away from fee-for-service payment for medical services. In 2023, Vori's management was contemplating a Series B financing round. They recognized that the investment environment had changed since the company their Series A round. Vori saw that investors were using profitability as a metric over revenue and were less eager to fund organizations holding large customer contracts that required a significant level of resources, preferring companies that could grow organically from their own profits. Vori had to meet these investor expectations while facing a stiff competitive market, including traditional providers that continued to favor the surgery-first model, other virtual services, and new deep-pocketed entrants that had entered the healthcare field.

In light of these developments, Vori needed to develop a go-to-market plan that allowed it to reach its potential customers and scale. This would mean identifying potential healthcare partners as well as fashioning pricing models and operational enhancements to facilitate working within an expanded network. 

US state governments frequently provide firms with targeted subsidies. In exchange, recipients promise to create or retain a certain number of jobs in the subsidizing state. Using novel hand-collected data, we address three questions: (i) the extent to which firms meet job creation targets promised in their applications, (ii) the factors that determine which firms meet the targets, and (iii) the benefits to firms from meeting those targets. We find that 63% of subsidies awarded to publicly traded U.S. firms between 2004 and 2015 meet their promised job creation targets. Firms with poorer labor practices are less likely to meet their targets, as are politically connected firms that receive subsidies in election years. Conversely, promised job targets are also more likely to be met for subsidies accompanied by government press releases but less likely to be met for subsidies accompanied by firm press releases; the latter likely reflects the fact that firms put out press releases for larger subsidies with more ambitious job targets. In terms of consequences, firms that meet job targets are more successful at obtaining subsequent subsidies both in and out of subsidizing states. However, while firms’ success in meeting job targets is associated with an uptick in positive media coverage, this does not flow through to ESG ratings, even on scores specific to community impact. Our results should be of interest to both academics and policymakers interested in the design of state-level economic incentives.

When employees are living a calling at work, they tend to experience greater well-being and the organization also benefits. Despite the integral role of the organization, research has not sufficiently explored what organizational factors might help employees live a calling. Drawing on a tripartite theoretical framework of living a calling— characterized by destiny, personal significance, and social significance— and Work as a Calling Theory, we hypothesize that needs-supplies fit, empowerment, and servant leadership are positively related to living a calling. Further, we hypothesize that the benefits of living a calling extend to the organization via a negative association with deviant behaviors, a positive association with LMX relationships, and that consistency of interests (a facet of grit) is a boundary condition of the proposed relationships. Through testing our hypotheses in a multi-wave, multi-source field study of employees and supervisors in a park district, we find that needs-supplies fit and empowerment facilitate living a calling in an organization. Further, consistency of interests moderates the relationship between living a calling and deviant behaviors and LMX. Our findings indicate how employers might help employees live their callings, and, in turn, mitigate negative and attain positive outcomes.

Volunteer crowdsourcing platforms match volunteers with tasks which are often recurring. To ensure completion of such tasks, platforms frequently use a lever known as ``adoption,'' which amounts to a commitment by the volunteer to repeatedly perform the task. Despite reducing match uncertainty, high levels of adoption can decrease the probability of forming new matches, which in turn can suppress growth. We study how platforms should manage this trade-off. Our research is motivated by a collaboration with Food Rescue U.S. (FRUS), a volunteer-based food recovery organization active in over 30 locations. For platforms such as FRUS, effectively utilizing non-monetary levers, such as adoption, is critical. Motivated by the volunteer management literature and our analysis of FRUS data, we develop a model for two-sided markets which repeatedly match volunteers with tasks. We study the platform's optimal policy for setting the adoption level to maximize the total discounted number of matches. When market participants are homogeneous, we fully characterize the optimal myopic policy and show that it takes a simple extreme form: depending on volunteer characteristics and market thickness, either allow for full adoption or disallow adoption. In the long run, we show that such a policy is either optimal or achieves a constant-factor approximation. We further extend our analysis to settings with heterogeneity and find that the structure of the optimal myopic policy remains the same if volunteers are heterogeneous. However, if tasks are heterogeneous, it can be optimal to only allow adoption for the harder-to-match tasks. Our work sheds light on how two-sided platforms need to carefully control the double-edged impacts that commitment levers have on growth and engagement. Setting a misguided adoption level may result in marketplace decay. At the same time, a one-size-fits-all solution may not be effective, as the optimal design crucially depends on the characteristics of the volunteer population.

Markets and firms offer contrasting methods to arrange production. In markets, contracts govern the purchase of parts and services. In firms, the shared values, customs and norms coming from a corporate culture govern employees' joint development of parts and services. We argue for this distinction as a theory of the firm. Firms exist because corporate culture at times is more efficient to carry out production than are detailed contracts. The firm's boundary encircles the areas of production for which a manager optimally chooses corporate culture as the organizing device. Consistent with empirical evidence, the model explains why some mergers and acquisitions fail, and why corporate cultures are hard to change.

The emergence of the AI-driven digital economy has raised concerns among economists and policymakers regarding the market power of firms that sell data. The unique characteristics of data, such as its large fixed cost and ability to be replicated at zero marginal cost, suggest the potential for natural monopolies in data markets. However, little is known about how data markets function and how data is priced. This paper documents characteristics of data markets, explores the potential market power of data monopolists through theoretical analysis and uses modeling and data together to explore how consumers fare under different data pricing models. The authors develop a dynamic model of a monopolist data seller with two crucial features: Data that many other buyers also have loses value, and data sellers cannot commit not to sell the same data to more buyers in the future. In such circumstances, even data monopolists have limited power to extract profits. Customers who anticipate more future sales of the data they buy will discount the value of the data. Customers’ willingness to pay for something that many others will know tomorrow is low. Thus, the concern shifts from excessive profits to potential under-provision of data.

We present evidence that is consistent with large disparities across firms in their on-the-job learning opportunities, using administrative datasets from Brazil and Italy. We categorize firms into discrete “classes”—which our conceptual framework interprets as skill-learning classes—using a clustering methodology that groups together firms with similar distributions of unexplained wage growth. Mincerian returns to experience vary widely across experiences acquired in different firm classes. Four tests leveraging firm stayers and movers, occupation and industry switchers, hiring wages, and displaced workers point towards a portable and general human capital interpretation. Heterogeneous employment experiences explain an important share of wage variance by age 35, thus contributing to shape wage inequality. Firms’ observable attributes only mildly predict on-the-job learning opportunities.

We examine whether investors value workforce gender diversity. Consistent with the view that investors believe that workforce gender diversity can be valuable in major firms, we use event studies to demonstrate that U.S. technology firms and U.S. financial firms experience more positive stock price reactions when it is revealed that they have relatively higher (versus lower) workforce gender diversity numbers. For instance, we find that Google’s revelation of relatively low workforce gender diversity numbers triggered a negative stock price reaction, whereas eBay’s revelation of relatively high workforce gender diversity numbers triggered a positive stock price reaction. These stock price reactions are both economically and statistically significant; e.g., we estimate that if a technology firm had revealed gender diversity numbers that were one standard deviation higher, its market valuation would have increased by $1.11 billion. Corroborating this plausibly causal field evidence, we also find positive investor reactions to workforce gender diversity in randomized experiments using Prolific participants with investing experience; these reactions seem to be underpinned by investors’ beliefs about potential upsides of diversity for the firm (e.g., reduced legal risks; increased creativity) but not by investors’ beliefs about potential downsides of diversity for the firm (e.g., increased conflict). Our findings highlight the importance of understanding investors’ intuitions or beliefs about major organizational phenomena such as workforce gender diversity. Our results also point towards a new type of business case for diversity, driven by investors: if major firms had more workforce gender diversity, investors may “reward” them with substantially higher valuations.

We study mislearning from equilibrium prices, and contrast this with mislearning from exogenous fundamentals. We micro-found mislearning from prices with a psy- chologically founded theory of “Partial Equilibrium Thinking” (PET), where traders learn fundamental information from prices, but fail to realize others do so too. PET leads to over-reaction, and upward sloping demand curves, thus contributing to more inelastic markets. The degree of individual-level over-reaction, and the extent of in- elasticity varies with the composition of traders, and with the informativeness of new information. More generally, unlike mislearning from fundamentals, mislearning from prices i) generates a two-way feedback between prices and beliefs that can provide an arbitrarily large amount of amplification, and ii) can rationalize both over-reaction and more inelastic markets. The two classes of biases are not mutually exclusive. Instead, they interact in very natural ways, and mislearning from prices can vastly amplify mislearning from fundamentals.

We view innovation investments as real options and explore the implications of risk (volatility) as well as a newly defined outcome independent measure of ambiguity—Knightian uncertainty— for innovation decisions. The empirical analysis uses stock returns to compute an implementable measure of ambiguity. We also control for risk and other determinants of innovation. We find a consistently significant negative effect of ambiguity on R&D, patents, and citations, as predicted. The effect of risk on R&D is positive and significant, but the corresponding effect on patents and citations is negative and significant. Ambiguity matters more for high-tech firms, consistent with intuition.

We propose a novel theory-based approach to the reinforcement learning problem of maximizing profits when faced with an unknown demand curve. Our method is based on multi-armed bandits, which are a collection of minimal assumption nonparametric models that balance exploration and exploitation for maximizing rewards while learning across arms. We leverage the informational externality inherent in price experimentation across arms (price levels) in two ways: correlation between demands corresponds to closer price levels, and demand curves are weakly monotonically decreasing. The first information externality is captured by the use of Gaussian process bandits. We expand on this literature by incorporating the second information externality (monotonicity) into Gaussian process bandits; we provide both a monotonic version of GP-UCB and GP-TS. Incorporating these informational externalities limits unnecessary exploration of certain prices and performs better (characterized by greater profitability or reduced experimentation) than current benchmark approaches. Additionally, our method can be used in conjunction with methods like partial identification. Finally, we provide further variants of these algorithms which account for heteroscedasticity in the noise of the purchase data. Across a wide spectrum of true demand distributions and price sets, our algorithm demonstrated a significant increase in rewards, most notably for underlying WTP distributions where the optimal is low (among the set of prices considered). Our algorithm performed consistently, achieving over 95% of the optimal rewards in every simulation setting tested.

Reading disabilities affect 10%–20% of students in the US. Untreated students fall behind their typically developing peers, leading to poor long-term outcomes. While instructional interventions can help, they are most effective when implemented early. Inexpensive screening tests can be used to monitor and flag at-risk students who may need expensive follow-up diagnostic evaluations that determine eligibility for intervention. However, conventional wisdom holds that the accuracy of these tests increase with grade level. Schools that do not have the capacity to do follow-up evaluations on every student flagged by screening are therefore believed to face an operational trade-off in allocating resources for evaluations, balancing the need for early intervention against budget constraints and legal obligations to honor direct parent or teacher requests. We examine how school administrators can choose evaluation policies to maximize benefits from intervention for students and ensure equitable allocation across diverse backgrounds. We model identification by optimizing over a time-dependent Bernoulli process which incorporates the screening test accuracy and the benefits from intervention at different grade levels. In collaboration with researchers from the Florida Center for Reading Research, we use longitudinal data from school districts across the state to empirically estimate these parameters and numerically solve for the optimal policies. Our study provides actionable insights for school administrators making resource allocation decisions and policy makers considering changes to laws governing the identification process. In this context, counter to conventional wisdom the screening test accuracy does not increase with grade level. To maximize the benefit to students under the current identification process, schools should simply evaluate as many students as their budget allows as early as possible. At existing budget levels, this policy also results in maximally equitable allocations. Changes to the identification process that ease legal obligations can increase benefits by up to 66% and decrease disparities by up to 100% without additional funding.

Persuasive communication functions not only through content but also delivery, e.g., facial expression, tone of voice, and diction. This paper examines the persuasiveness of delivery in start-up pitches. Using machine learning (ML) algorithms to process full pitch videos, we quantify persuasion in visual, vocal, and verbal dimensions. Positive (i.e., passionate, warm) pitches increase funding probability. Yet conditional on funding, high-positivity startups underperform. Women are more heavily judged on delivery when evaluating single-gender teams, but they are neglected when co-pitching with men in mixed-gender teams. Using an experiment, we show persuasion delivery works mainly through leading investors to form inaccurate beliefs.

We investigate the role of private equity (PE) in the resolution of failed banks after the 2008
financial crisis. Using proprietary failed bank acquisition data from the FDIC combined with data
on PE investors, we find that PE investors made substantial investments in underperforming and
riskier failed banks. Further, these acquisitions tended to be in geographies where the other local
banks were also distressed. Our results suggest that PE investors helped channel capital to
underperforming failed banks when the “natural” potential bank acquirers were themselves
constrained, filling the gap created by a weak, undercapitalized banking sector. Next, we use a
quasi-random empirical design based on proprietary bidding data to examine ex post performance
and real effects. We find that PE-acquired banks performed better ex post, with positive real effects
for the local economy. Our results suggest that private equity investors had a positive role in
stabilizing the financial system in the crisis through their involvement in failed bank resolution.

Non-informational cues, such as facial expressions, can significantly influence judgments and interpersonal impressions. While past research has explored how smiling affects business outcomes in offline or in-store contexts, relatively less is known about how smiling influences consumer choice in e-commerce settings even when there is no face-to-face interaction. In this paper, we use a longitudinal Airbnb dataset and a facial attribute classifier to quantify the effect of a smile in the host's profile photo on property demand and identify factors that influence when a host's smile is likely to have the biggest effect. A smile in the host's profile photo increases property demand by 3.5% on average. This effect is moderated by a variety of host and property characteristics that provide evidence for the role of uncertainty underlying why smiling increases demand. Specifically, when there is greater uncertainty regarding either the quality of the accommodations or the interaction with the host, a host smile will have a greater effect on demand. Online experiments confirm this pattern, offering further support for uncertainty perceptions driving the effect of smiling on increased Airbnb demand, and show that the effect of smiling on demand generalizes beyond Airbnb.

We study an advertiser’s targeting strategy and its effects on consumer data privacy choices, both of which determine the advertiser’s targeting accuracy. Targeted ads, serving as implicit recommendations when consumer preferences are uncertain, not only influence the consumer’s beliefs and purchasing decisions but also amplify the advertiser’s temptation towards strategic mistargeting—sending ads to poorly matched consumers. Our analysis reveals that advertisers may, paradoxically, choose less precise targeting as accuracy improves. Even if prediction is perfect, the advertiser still targets the wrong consumers, leading to strategic mistargeting. Nev- ertheless, consumer surplus can remain positive due to improved identification of well-matched consumers, thereby reducing the incentive for consumers to withhold information. However, the scenario shifts with endogenous pricing; better prediction leads to more precise targeting, although mistargeting persists. To exploit the recommendation effect of advertising, the ad- vertiser raises prices instead of diluting recommendation power, lowering consumer welfare and prompting consumers to opt out of data collection. Furthermore, we investigate the impact of consumer data opt-out decisions under varying privacy policy defaults (opt-in vs. opt-out). These decisions significantly affect equilibrium outcomes, influencing both the advertiser’s tar- geting strategies and consumer welfare. Our findings highlight the complex relationship between targeting accuracy, privacy choices, and advertisers’ incentives