Publications

We study the role of information and access in capacity-constrained selection problems with fairness concerns. We develop a statistical discrimination framework, where each applicant has multiple features and is potentially strategic. The model formalizes the tradeoff between the (potentially positive) informational role of a feature and its (negative) exclusionary nature when members of different social groups have unequal access to this feature. Our framework finds a natural application to policy debates on dropping standardized testing in admissions. Our primary takeaway is that the decision to drop a feature (such as test scores) cannot be made without the joint context of the information provided by other features and how the requirement affects the applicant pool composition. Dropping a feature may exacerbate disparities by decreasing the amount of information available for each applicant, especially those from nontraditional backgrounds. However, in the presence of access barriers to a feature, the interaction between the informational environment and the effect of access barriers on the applicant pool size becomes highly complex. Furthermore, we consider an extension with two schools and costly tests, where strategic students decide whether to take the test or not. Our theoretical results reveal that the students’ test-taking behavior can be nonmonotonic. We characterize the two-school policy equilibria and show that each school’s optimal decision to drop the test critically depends on the other school’s test policy. Finally, using calibrated simulations, we demonstrate the presence of practical instances where the decision to eliminate standardized testing improves or worsens all metrics.

Motivated by our collaboration with a major refugee resettlement agency in the United States, we study a dynamic matching problem where each new arrival (a refugee case) must be matched immediately and irrevocably to one of the static resources (a location with a fixed annual quota). In addition to consuming the static resource, each case requires postallocation service from a server, such as a translator. Given the time-consuming nature of service, a server may not be available at a given time, thus we refer to it as a dynamic resource. Upon matching, the case will wait to avail service in a first-come-first-serve manner. Bursty matching to a location may result in undesirable congestion at its corresponding server. Consequently, the central planner (the agency) faces a dynamic matching problem with an objective that combines the matching reward (captured by pair-specific employment outcomes) with the cost for congestion for dynamic resources and overallocation for the static ones. Motivated by the observed fluctuations in the composition of refugee pools across the years, we design algorithms that do not rely on distributional knowledge constructed based on past years’ data. To that end, we develop learning-based algorithms that are asymptotically optimal in certain regimes, easy to interpret, and computationally fast. Our design is based on learning the dual variables of the underlying optimization problem; however, the main challenge lies in the time-varying nature of the dual variables associated with dynamic resources. To overcome this challenge, our theoretical development brings together techniques from Lyapunov analysis, adversarial online learning, and stochastic optimization. On the application side, when tested on real data from our partner agency and incorporating practical considerations, our method outperforms existing ones, making it a viable candidate for replacing the current practice upon experimentation.

Conversational recommender systems powered by generative AI can enhance personalization by facilitating information elicitation through follow-up questions. However, engaging in these conversations imposes a communication cost on users. As platforms with different objectives and monetization models deploy these systems, a central question is: how does the platform’s objective and sellers’ strategic response shape the design of these systems in terms of their elicitation strategy? We develop a parsimonious model of conversational elicitation in which interaction generates noisy preference information and imposes a communication cost borne by the user. A user-welfare-maximizing platform elicits more information when accurate niche matching yields large gains, even when niche users are rare. In contrast, under a conversion objective, for the same setting, the optimal strategy is to immediately recommend the same mainstream option to all users with no or minimal preference elicitation because the incremental conversion benefit from improved matching is bounded, while communication costs are borne by all users. When prices are endogenous and the platform earns a commission, increased elicitation is again optimal because improved screening raises equilibrium prices and platform revenue; however, these price responses can counteract consumer benefits and reduce user welfare. The model also highlights that the optimal elicitation intensity increases with preference heterogeneity, helping explain why conversational systems ask more in highly differentiated categories than in low-heterogeneity ones. We complement the theory with a dataset of long-form product queries that vary in length and informational content. Using our dataset and LLM-based user simulation, we quantify how additional information impacts user decisions and demonstrate that the magnitude of this impact depends on the degree of preference heterogeneity. Additionally, this dataset provides a testbed for measuring the (incremental) value of preference elicitation and may be of independent interest.

We study and develop an algorithmic framework for incorporating "ex-ante" constraints—constraints on outcomes that hold only on average—into stateful sequential search problems with costly inspection. Our framework encompasses the classical Weitzman's Pandora's box [Weitzman,1978] as well as its extensions to joint Markovian scheduling [Dumitriu et al., 2003; Gittins, 1979], which model richer processes such as multistage search with multiple layers of inspection. Ex-ante constraints are particularly motivated by social considerations in algorithmic hiring, where they can adjust outcome distributions to promote equity and access. While most work in the algorithmic fairness literature in computer science and economics has focused on incorporating such constraints into machine learning tasks like classification and regression, far less attention has been devoted to operational problems such as sequential search, with their unique intricacies. Our work aims to bridge this gap. Building on the optimality of index-based policies in the unconstrained versions of these problems, we show that optimal policies under a single ex-ante constraint (e.g., demographic parity) retain an index-based structure but require (i) dual-based adjustments of the indices and (ii) randomization between two such adjustments via a "tie-breaking rule," both easy to compute and economically interpretable. We then extend our results to multiple affine constraints by reducing the problem to a variant of the exact Carathéodory problem and providing a polynomial-time algorithm that constructs an optimal randomized dual-adjusted index-based policy satisfying all constraints simultaneously. For general affine and convex constraints, we develop a primal-dual algorithm that randomizes over a polynomial number of dual-based adjustments, yielding a near-feasible, near-optimal policy. These results rely on the key observation that a suitable relaxation of the Lagrange dual function for these constrained problems admits index-based policies akin to those in the unconstrained setting. Finally, through a numerical study, we investigate the implications of imposing socially aware ex-ante constraints and their socially desirable outcomes.

Apportionment is the act of distributing the seats of a legislature among political parties (or states) in proportion to their vote shares (or populations). A famous impossibility by Balinski and Young (2001) shows that no apportionment method can be proportional up to one seat (quota) while also responding monotonically to changes in the votes (population monotonicity). Grimmett (2004) proposed to overcome this impossibility by randomizing the apportionment, which can achieve quota as well as perfect proportionality and monotonicity — at least in terms of the expected number of seats awarded to each party. Still, the correlations between the seats awarded to different parties may exhibit bizarre non-monotonicities. When parties or voters care about joint events, such as whether a coalition of parties reaches a majority, these non-monotonicities can cause paradoxes, including incentives for strategic voting. In this paper, we propose monotonicity axioms ruling out these paradoxes, and study which of them can be satisfied jointly with Grimmett’s axioms. Essentially, we require that, if a set of parties all receive more votes, the probability of those parties jointly receiving more seats should increase. Our work draws on a rich literature on unequal probability sampling in statistics (studied as dependent randomized rounding in computer science). Our main result shows that a sampling scheme due to Sampford (1967) satisfies Grimmett’s axioms and a notion of higher-order correlation monotonicity.

To meet voluntary climate targets, firms often complement internal decarbonization efforts by purchasing carbon credits in the voluntary carbon market (VCM), which finance projects that reduce emissions elsewhere. However, these emissions reductions are difficult to verify, and growing evidence of overcrediting has cast doubt on the VCM's potential to genuinely offset emissions. We investigate how the VCM's defining features shape its climate effectiveness. Our model captures three central elements: adverse selection, as high-quality projects that truly reduce emissions are costlier yet difficult to distinguish from low-quality ones; imperfect third-party certification, as projects are screened based on a noisy signal of quality; and buyer preferences for non-carbon attributes, as some firms value credits that generate observable social or economic co-benefits beyond reducing emissions. We show that the market fails to sustain trade if certification is sufficiently noisy, as quality uncertainty erodes buyer confidence and triggers a market-for-lemons collapse. However, demand for co-benefits can sustain markets that would otherwise collapse. Yet in such cases, the market remains active but yields limited carbon abatement, as most traded credits are low-quality. We then examine policy and market design interventions reflecting recent developments in practice, such as penalizing buyers for greenwashing and offering credit portfolios. We show that these measures can be counterproductive for carbon mitigation if certification remains inaccurate. Accordingly, we demonstrate that the certifier’s incentives for accuracy can be strengthened by modifying its fee structure so that its revenue is tied to the market value rather than the volume of credits.

Problem definition. Expanding infrastructure in transport, energy, and digital sectors is important for achieving the economic and sustainability goals of many developing economies. The values of projects involving such infrastructure expansion can be interdependent due to inherent network effects. However, institutional constraints and limited coordination capacity often force governments to award projects through separate auctions. Brazil’s National Logistics Plan (PNL) exemplifies this setting, where interdependent transport investments are procured separately despite strong network effects across projects. We study how a government buyer should design separate per-project procurements when the overall value depends on the resulting infrastructure network, and bundling is not feasible. Methodology/results. We analyze: (i) parallel procurement, in which auction rules are fixed in advance; (ii) sequential procurement, in which their sequence is fixed, but later rules can adapt to earlier executed projects. In both settings, we show that the optimal mechanism retains a simple form akin to canonical results in auction theory: each project is awarded to the lowest virtual-cost bidder if its virtual cost falls below a project-specific threshold that reflects its expected contribution to the equilibrium-induced network. Beyond the optimal mechanism, we propose a simple and interpretable ``credit-aware'' heuristic that adjusts each project's threshold by a simple one-step estimate of its expected network contribution. To demonstrate the importance of network effects in procurement design, we construct a calibrated case study of a major railroad corridor expansion under the PNL. Modeling the Brazilian infrastructure as a hub--spoke network, and the expansion as two key added segments, we estimate their network effect by solving a congestion game aligned with current practice. We show that joint commissioning of these segments adds 16.4% additional economic value, and the common practice of ignoring network effect in procurement design leaves 30–36% of attainable surplus unrealized. Our credit-aware heuristic recovers about 90% of the optimal surplus. Managerial implications. As developing economies invest in infrastructure expansion, it is important to evaluate these projects holistically, as network effects can be strong and first-order. Even when institutional constraints require projects to be procured separately, network effects can still be incorporated effectively through careful adjustments to existing mechanisms, and even simple changes can yield substantial economic and environmental gains.

The design of algorithms for political redistricting generally takes one of two approaches: optimize an objec- tive such as compactness or, drawing on fair division, construct a protocol whose outcomes guarantee partisan fairness. We aim to have the best of both worlds by optimizing an objective subject to a binary fairness constraint. As the fairness constraint we adopt the geometric target, which requires the number of seats won by each party to be at least the average (rounded down) of its outcomes under its worst and best possible partitions of the state.

A seller wants to sell an item to n buyers. Buyer valuations are drawn i.i.d. from a distribution unknown to the seller; the seller only knows that the support is included in [a,b]. To be robust, the seller chooses a DSIC mechanism that optimizes the worst-case performance relative to the ideal expected revenue the seller could have collected with knowledge of buyers' valuations. Our analysis unifies the regret and the ratio objectives. 
For these objectives, we derive an optimal mechanism and the corresponding performance in quasi-closed form, as a function of the support information [a,b] and the number of buyers n. Our analysis reveals three regimes of support information and a new class of robust mechanisms. i.) When a/b is below a threshold, the optimal mechanism is a second-price auction (SPA) with random reserve, a focal class in earlier literature. ii.) When a/b is above another threshold, SPAs are strictly suboptimal, and an optimal mechanism belongs to a class of mechanisms we introduce, which we call pooling auctions (POOL); whenever the highest value is above a threshold, the mechanism still allocates to the highest bidder, but otherwise the mechanism allocates to a uniformly random buyer, i.e., pools low types. iii.) When a/b is between two thresholds, a randomization between SPA and POOL is optimal. 
We also characterize optimal mechanisms within nested central subclasses of mechanisms: standard mechanisms that only allocate to the highest bidder, SPA with random reserve, and SPA with no reserve. We show strict separations in terms of performance across classes, implying that deviating from standard mechanisms is necessary for robustness.

Researchers were already expecting a gradual enrollment slowdown before the onset of COVID-19. Public school enrollment edged up only 2% between 2012 and 2019, holding near 50 million students, while the U.S. total fertility rate had slipped to 1.71 births per woman—well below the replacement level—foreshadowing a smaller school-age cohort. The pandemic turned that slow decline into a sudden shock. Studies document steep post-2020 losses in Massachusetts, Virginia, Michigan, and California. Research at the national level shows similar trends in urban and high-poverty districts, and a surge in both homeschooling and private-schooling that still leaves millions of children “missing” from any formal roll. In addition to concerns around student progress, shrinking headcounts also create immediate fiscal stress because most state and federal aid flows on a per-pupil basis. District leaders have already considered adjusting school capacity, redistricting, or even closing campuses to balance budgets— steps that are often politically sensitive but considered in response to fiscal pressures. Recent evidence confirms that steeper enrollment losses measurably raise the odds of permanent closure. Enrollment shifts have not fallen evenly across student groups. Recent evidence shows that kindergarten enrollment fell most sharply for black and low-income children, whereas the smaller declines observed in later grades were concentrated among white and higher-income students already enrolled in public schools. Such patterns may heighten long-standing worries about potential re-segregation and resource inequality. Against this backdrop, policymakers and district officials are experimenting with strategies to stem further enrollment losses or mitigate their effects. Some, like New York City, have pledged to preserve school budgets even as rolls shrink. Others hope new curricula or enhanced parent outreach will attract students back. This report provides detailed estimates of recent shifts in public school enrollment. By linking the latest National Center for Education Statistics data with federal population estimates, the report shows how enrollment shifts differ across districts with distinct racial and economic profiles. In addition, it projects how continued enrollment declines could drive future school closures and alter the number of seats traditional districts will need through 2050.

Classical Bayesian mechanism design relies on the common prior assumption, but the common prior is often not available in practice. We study the design of prior-independent mechanisms that relax this assumption: The seller is selling an indivisible item to n buyers such that the buyers’ valuations are drawn from a joint distribution that is unknown to both the buyers and the seller, buyers do not need to form beliefs about competitors, and the seller assumes the distribution is adversarially chosen from a specified class. We measure performance through the worst-case regret, or the difference between the expected revenue achievable with perfect knowledge of buyers’ valuations and the actual mechanism revenue. We study a broad set of classes of valuation distributions that capture a wide spectrum of possible dependencies: independent and identically distributed (i.i.d.) distributions, mixtures of i.i.d. distributions, affiliated and exchangeable distributions, exchangeable distributions, and all joint distributions. We derive in quasi closed form the minimax values and the associated optimal mechanism. In particular, we show that the first three classes admit the same minimax regret value, which is decreasing with the number of competitors, whereas the last two have the same minimax regret equal to that of the case n = 1. Furthermore, we show that the minimax optimal mechanisms have a simple form across all settings: a second-price auction with random reserve prices, which shows its robustness in prior-independent mechanism design. En route to our results, we also develop a principled methodology to determine the form of the optimal mechanism and worst-case distribution via first-order conditions that should be of independent interest in other minimax problems.

Two-sided platforms rely on their recommendation algorithms to help visitors successfully find a match. However, on platforms such as VolunteerMatch, which has facilitated millions of connections between volunteers and nonprofits, a sizable fraction of website traffic arrives directly to a nonprofit’s volunteering page via an external link, thus bypassing the platform’s recommendation algorithm. We study how such platforms should account for this external traffic in the design of their recommendation algorithms, given the goal of maximizing successful matches. We model the platform’s problem as a special case of online matching, where (using VolunteerMatch terminology) volunteers arrive sequentially and probabilistically match with one opportunity, each of which has a finite need for volunteers. In our framework, external traffic is interested only in their targeted opportunity; by contrast, internal traffic may be interested in many opportunities, and the platform’s online algorithm selects which opportunity to recommend. In evaluating the performance of different algorithms, we refine the notion of competitive ratio by parameterizing it based on the amount of external traffic. After demonstrating the shortcomings of a commonly used algorithm that is optimal in the absence of external traffic, we propose a new algorithm, adaptive capacity (AC), which accounts for matches differently based on whether they originate from internal or external traffic. We provide a lower bound on AC’s competitive ratio that is increasing in the amount of external traffic and that is close to (and, in some regimes, exactly matches) the parameterized upper bound we establish on the competitive ratio of any online algorithm. We complement our theoretical results with a numerical study motivated by VolunteerMatch data where we demonstrate the strong performance of AC relative to current practice and further our understanding of the difference between AC and other commonly used algorithms.

We study non-monetary markets where objects that arrive over time are allocated to unit-demand agents with private types, such as in the allocation of public housing or deceased-donor organs. An agent’s value for an object is supermodular in her type and the object quality, and her payoff is her value minus her waiting cost. The social planner’s objective is a weighted sum of allocative efficiency (i.e. the sum of values) and welfare (i.e. the sum of payoffs). We identify optimal mechanisms in the class of direct-revelation mechanisms. When the social planner can design the information disclosed to the agents about the objects, the optimal mechanism has a simple implementation: a first-come first-served waitlist with deferrals. In this implementation, the object qualities are partitioned into intervals; only the interval containing the object quality is disclosed to agents. When the planner places a higher weight on welfare, optimal disclosure policies become coarser.

In collaboration with VolunteerMatch (VM)---the world's largest online platform for connecting volunteers with nonprofits---we designed and implemented a new display ranking algorithm. VM's original ranking algorithm was intended to maximize efficiency (i.e., the total number of connections), but as a consequence it repeatedly displayed the same few opportunities at the top of its ranking, effectively limiting access to volunteers for the other opportunities. To incorporate VM's desire for equity (defined as the weekly number of opportunities with at least one connection) along with efficiency, we propose a modeling framework for online display ranking in settings where it is important to manage the trade-off between the total number of connections and the equitable allocation of these connections. We take an adversarial approach in evaluating the performance of online algorithms and show that a class of algorithms that applies a penalty to opportunities after each connection provides a strong (and, in certain regimes, optimal) performance guarantee. Inspired by our theoretical results yet mindful of practical considerations on VM's platform, we propose SmartSort, a simple score-based ranking algorithm which enjoys comparable guarantees in many regimes. We implemented SmartSort in two experiments, covering Dallas-Fort Worth and all of Southern California. Using a difference-in-differences analysis, we find that the implementation of SmartSort led to a 8-9% increase in the weekly average number of opportunities with at least one connection (consistent across both experiments) without any meaningful decrease in the total number of connections, implying a Pareto improvement for VM. Based on the success of our experiments, SmartSort has now been deployed nationwide. If SmartSort has a similar distributional effect on a national scale, every year, an additional 30,000 connections will go to opportunities that would have otherwise lacked access to volunteers.

Operations research, originating during World War II, is the scientific study of operations aimed at improving decision-making and organizational performance. Initially focused on military logistics, its scope has expanded to address diverse operational problems in business, government, and non-profit sectors. These include scheduling, capacity planning, routing, and resource allocation. Through mathematical modeling and analysis, operations research seeks not just to describe but to optimize operations by aligning them with organizational goals such as maximizing profit, minimizing costs, or enhancing effectiveness. The field has evolved from simple problem identification to complex mathematical modeling, emphasizing the importance of framing the right problem within a well-understood system context. Applied operations research assumes that the identified problem can be modeled mathematically, that the models and assumptions are valid, and that organizational objectives and constraints are clearly defined and quantifiable. The ultimate aim is actionable recommendations that improve real-world decision-making. Grounded in the belief in mathematical rigor, operations research integrates objectives and constraints to deliver feasible solutions. By leveraging analytical tools, it supports better decision-making, ensuring that operations are not only efficient but also aligned with strategic priorities, making it a practical and impactful discipline across sectors.

The transparency and control of earnings are major concerns for gig economy workers across platforms such as ride-hailing and food delivery. While workers advocate for greater transparency, platforms selectively disclose information, shaping workers' decision-making and earnings. Recently, the Federal Trade Commission (FTC) has highlighted lack of transparency as a key issue, and platforms have responded by introducing upfront pay quotes that provide pertrip compensation details for workers. Using a game-theoretic model, we analyze the strategic interactions between the platform and workers, incorporating tools from information design to examine how different transparency policies-specifically, a fixed commission rate versus upfront pay quotes-shape equilibrium outcomes. We find that greater transparency can paradoxically increase platform control, as it allows platforms to fine-tune pay structures in ways that ultimately reduce worker autonomy. Moreover, while full information benefits the platform when it has flexibility in commission setting, it can backfire under commitment constraints, leading to lower profits than a no-information policy. Our findings highlight that transparency is not inherently beneficial for workers. Instead, its effects depend on how it interacts with pay policies. In particular, simple mechanisms, such as a fixed commission rate, can provide workers with more stability and bargaining power than per-trip transparency. These insights offer important guidance for policymakers and platform designers navigating the trade-offs of transparency in the gig economy. Keywords: Platforms, transparency, queueing, gig eco

In collaboration with Feeding America, we aim to redesign Real-Time—a tool on its food sourcing and rescue platform, MealConnect—that facilitates the connection of ad-hoc, time-sensitive food donations to local agencies (e.g., meal programs) through an offer process. In making offer decisions, Real-Time and similar food rescue platforms face a challenge in balancing efficiency and equity due to heterogeneity in response rates across agencies: offering to many agencies upfront improves efficiency (the likelihood of acceptance) but may disadvantage those with lower response rates. On top of this, the ad hoc nature of donations introduces future uncertainty, adding to the challenge of achieving the dual goals of efficiency and equity. Motivated by these challenges, we study a sequential offer scheduling problem in which donations arrive sequentially and are connected through a multi-stage offer process. The goal is to maximize an objective that balances efficiency and equity, promoting allocations proportional to agencies needs. We first develop a dynamic programming (DP)-based algorithm that optimally solves the one-donation problem and yields an intuitive nested offer schedule. Moving beyond one donation and motivated by the ad hoc nature of donations, we take a robust approach to designing sequential offer policies that do not rely on any knowledge about future donations. We design a penalty-based offer policy that solves a modified one-donation DP by properly penalizing current allocation to hedge against future uncertainty. We establish strong (and optimal in asymptotic regimes) performance guarantees for our policy. We further demonstrate the importance of hedging: a greedy policy that solves each one-donation DP without accounting for future arrivals fails to achieve a comparable guarantee. Numerical results on real data from Feeding America’s MealConnect platform demonstrate that our proposed approach significantly improves both efficiency and equity relative to current practice and several benchmarks.

AI is increasingly used to guide high-stakes educational decisions, yet its effectiveness depends on whether people follow its advice. We present the first large-scale field experiment with adolescents in this context, conducted across 14 public high schools. We compare the adoption of identical college-application recommendations from human counselors versus an AI-based algorithm. Contrary to the common assumption that objective, data-based recommendations favor AI, we find that algorithm aversion intensifies when recommendations are based on objective criteria (e.g., grades, admission chances) and dissipates when criteria are more subjective. We find that student perceptions of the recommender's intent strongly drive this aversion, consistently across scenarios and statistical approaches; perceptions of alignment with personal goals, ability, and comprehension also play significant roles. The results further reveal substantial heterogeneity in recommendation adoption rates. We observe stronger aversion among female students, students from rural schools, lower-GPA students, and those with stronger prestige-seeking attitudes. Using an optimization approach, we demonstrate how a policymaker can navigate the heterogeneity in recommendation adoption rates to optimally prioritize the assignment of human versus AI-based algorithmic recommenders, under varying social priorities and limited capacity of human counselors. We find that a targeting policy relying on few readily available student and school features can approximate the first-best, personalized targeting policy effectively. These insights underscore the importance of understanding student preferences and trust in an AI system's intent, and of adopting hybrid approaches that blend human guidance with AI tools to design effective recommendation systems.

Many industries, including the NFL with the Rooney Rule and law firms with the Mansfield Rule, have adopted interview-stage diversity interventions requiring a minimum representation of disadvantaged groups in the interview set. However, the effectiveness of such policies remains inconclusive. In light of this, we develop a framework of a two-stage hiring process, where rational firms, with limited interview and hiring capacities, aim to maximize the match value of their hires. The labor market consists of two equally sized social groups, m and w, with identical ex-post match value distributions. Match values are revealed only post-interview, while interview decisions rely on partially informative pre-interview scores. Pre-interview scores are more informative for group m, while interviews reveal more for group w; as a result, if firms could interview all candidates, both groups would be equally hired. However, due to limited interview capacity and information asymmetry, we show that requiring equal representation in the interview stage does not translate into equal representation in the hiring outcome, even though interviews are more informative for group w. In certain regimes, with or without intervention, a firm may interview more group w candidates but still hire fewer. At an individual level, we show that strong candidates from both groups benefit from the intervention as the candidate-level competition weakens. For borderline candidates, only group w candidates gain at the expense of group m. To understand the impact of non-universal interview-stage interventions on the market, we study a model with two vertically differentiated firms, where only the top firm adopts the intervention. We characterize the unique equilibrium and demonstrate potentially negative effects: we show that in certain regimes the lower firm hires fewer group w candidates due to increased firm-level competition for them, and further find examples where overall fewer group w candidates are hired across the market. At an individual level, while superstar candidates in both groups benefit, surprisingly the impact on borderline candidates may reverse: the lower firm may replace borderline group w candidates with borderline group m candidates in its interview set, effectively reducing the chance of those borderline group w candidates being hired. Overall, our findings highlight challenges in diversifying the labor market at early hiring stages due to information asymmetry, filtering, and competition. Beyond our context, our natural framework of a market with two-stage hiring may be of independent interest.

We develop an economic model of autonomous vehicle (AV) ride-hailing markets, in which uncertain aggregate demand is served with a combination of a fixed fleet of AVs and a flexible pool of human drivers (HVs). Dispatch efficiencies increase with scale because of density effects. We analyze market outcomes in this setting under four market configurations, defined by two dispatch platform structures (common platform versus independent platforms) and two levels of supply competition (monopoly AV versus competitive AV). A key result of our analysis is that the lower cost of AVs does not necessarily translate into lower prices; the price impact of AVs is ambiguous and depends critically on both the dispatch platform structure and the level of AV supply competition. In the extreme case, we show that if AVs and HVs operate on independent dispatch platforms, there is a monopoly AV supplier, and labor supply elasticity is sufficiently high, then prices are even higher than in a pure-HV market. Indeed, to guarantee consistently lower prices (relative to a pure HV market) in all scenarios and under all supply and density elasticities, a common dispatch platform between AVs and HVs is required. Furthermore, competitive AVs lead to lower prices than monopoly AVs in every such scenario. Our results illustrate the critical role that market configuration plays in realizing potential welfare gains from AVs.