With Mitchell Hoffman, Toronto Rotman, and Amir Sariri, Purdue Krannert
Updated July 2024

Do workers know which jobs are good? Especially for startups, it is very hard to separate the most promising technologies and business models from lemons. In a field experiment with 26 deep-tech seed-funded startups, we invite almost 20,000 business school alumni to apply for jobs, experimentally varying whether they see coarse signals of underlying firm quality. Showing these signals shifts applicants heavily toward higher-quality firms. Despite this, less than a quarter of firms in our sample and a broader set of AngelList job ads show any credible outside signals of quality that might assist potential workers identify top startups.

Updated March 2020

A miner finds gold in a river but doesn't know where the gold vein is. Different locations are not independent: not finding gold in one cave makes it more likely the vein will be found in as-yet-unsearched spots. Given a prior belief over where the gold vein is, how should you search? Unlike Weitzman, there is no Pandora rule: the probability you find the gold is a function of the full distribution of prior probabilities across locations. Sequential search is better than simultaneous search, but the reason for the difference is entirely different from the well-understood case of independent probabilities.

Updated July 2018 [R&R, Journal of Economic History]

The Wright Brothers invented the airplane in 1903 in the US, but the airplane industry is dominated by European firms by 1914. Of 21 canonical inventions we investigate, a similar "industrial reversal of fortune" occurs. Why? An invention is a discrete technological step where a series of necessary components are sufficiently advanced to make a technological leap. By decomposing inventions in their components, it is possible to see how an "inventing" location to actually be a technological laggard. This was true of the airplane: Europeans even in 1903 were more advanced in essentially every area necessary for the later commercial airplane except for lateral control.

With Erik Hovenkamp, USC Gould School of Law
Updated April 2016

When successive monopolies price noncooperatively, the resulting double markup is higher than that a monopolist would choose, hence vertically related firms often merge to coordinate the markup, maximize joint profits, and improve overall welfare. None of this is true when downstream is an oligopoly, whether the upstream markup is chosen cooperatively or noncooperatively. Therefore, there is a need to complete our intuition on what double marginalization does in the non-monopoly setting. We show precisely when noncooperative double marginalization is profitable, note that cooperative double marginalization chooses an upstream markup such that the downstream firm has a "consistent conjecture" about the rival's response curve, and show that simultaneous cooperative double marginalization induces an approximation of the "consistent conjectures equilibrium" from the theory of conjectural variation even though all upstream and downstream firms are individually Nash agents.

With Jorge Guzman, Columbia GSB
Online Appendix
Replication Data

We use cross-state business registrations to track the geographic movement of startups with high growth potential. In their first five years, 6.6% percent of these startups move across state borders. Though startup births are concentrated geographically, hubs like Silicon Valley and Boston on net lose startups to entrepreneurial migration. A revealed preference approach nonparametrically identifies the average utility of cities to migrant founders. University towns and startup hubs have low relative utility. This pattern is due neither to vertical sorting nor industrial specialization. The higher-quality startups move to lower-tax, business-friendly cities, while less growth-oriented startups move to low-tax, high-amenity cities.

With Jorge Lemus, UIUC, and Guillermo Marshall, UBC Sauder

When innovation in a given field becomes more lucrative, its direction can be distorted even as its rate rises. Higher payoffs attract innovators, making the R&D supply side more competitive. This competition endogenously shifts effort toward less promising but quicker-to-invent projects. We develop a dynamic structural model quantifying the magnitude of this distortion, even when the value of projects which are not invented in equilibrium are not observed. As a case study, we estimate the direction of pharmaceutical innovation during the Covid-19 pandemic, showing strategic distortion away from vaccines. Policy remedies include advance purchase commitments based on ex-ante value, targeted research subsidies, or antitrust exemptions for joint research ventures.

With Heidi Williams, Stanford

Innovation is central to long-run economic growth. This handbook chapter is a great way for young scholars to catch up on the history of the field, the existing base of knowledge, and the biggest open questions. We first articulate the key market failures in markets for innovation, and then discuss how both scientific norms and market-oriented policies help overcome those market failures. We close by discussing recent work on the diffusion of inventions as well as on the links between innovation and inequality.

With Laura Rosella, Emma Buajitti, U Toronto, and Vivek Goel, Waterloo

Economic lockdowns are incredibly costly not just to GDP and jobs, but also to long-term health since income loss is causally related to negative health outcomes. We combine a 54-sector production network model of the Canadian economy with a well-validated model of negative health outcomes by demographic history to estimate the equilibrium overall consequences of various lockdown scenarios. A single month of lockdown at the level of April 2020 leads to over 4500 excess deaths. Because of downstream spillovers from productivity shocks, it is better for GDP, employment and long-term health to close retail and restaurants before you close factories and transport, even though the former industries have many low-wage, young workers. A related paper on the public health methodology, "Downstream Health Impacts of Employment Losses During the Covid-19 Pandemic", is also available, in the Canadian Journal of Public Health, as well as a comment on public health research following the pandemic, in Healthcare Papers.

With Mike Ryall, Toronto Rotman, and Burkhard Schipper, UC Davis

Value capture theory uses perfect information cooperative game theory to study firm profitability in a logically rigorous way, without imposing strong assumptions about conduct. However, information differences, or even unawareness of economic possibilities, are intuitively important for explaining differences in firm outcomes. We provide a logically consistent extension of VCT to incorporate those features. We also explain why an approach using characteristic functions as a primitive are fundamentally incapable of capturing some important aspects of how informational asymmetry affects profits.

With Yasin Ozcan, NBER

Open Access mandates are common among funders and universities, with the NIH mandate binding on over one-third of publications in top medical journals. Yet prior research has not found substantive measurable benefits from these mandates. We show that open access increases citation of medical research in patents by 25 to 50 percent, an effect measured using double and triple differences that take advantage of a 2008 NIH policy change. Further, we match academic papers to patents using the in-specification references in the actual writeup of the patent, and show that these references are actually very different from the types of references you see in the prior art section. We argue that in-specification references are more likely to measure "the paper trail of knowledge."

With Susan Athey, Stanford GSB, and Joshua Gans, Toronto Rotman

AI generally does not operate alone, but rather in conjunction with human agents. Self-driving long-haul trucks can be overridden by human drivers. Judicial bail predictions are passed along to human judges. In a principal-agent setting, what does "good AI" mean? Using an Aghion-Tirole delegation model, we show that both the delegation of final decision rights to AI or agent, the relative bias of the AI versus the agent, and the efficiency of the AI all affect human effort. In particular, in some cases, the principal prefers an AI that is more biased or less efficient in order to induce more effort from the agent.

With Erik Hovenkamp, USC Gould School of Law

Traditional Chicago School antitrust is biased against the plaintiff. The argument is that market power incorrectly permitted is partially self-correcting: high industry profits attract entry. On the other hand, efficient behavior incorrectly banned is gone forever. This logic does not apply to startup acquisitions. They can be anticompetitive, and permitting these anticompetitive acquisitions makes it harder to compete against the industry leader. Post-Chicago game-theoretic arguments make clear that this incorrect application of the error cost argument extends beyond just startup acquisitions.

With Erik Hovenkamp, USC Gould School of Law

Should there be limits to startup acquisitions, particularly when the startup is not now nor potentially in the future a horizontal competitor to the acquirer? Yes. Acquisitions and licensing can be inefficient when they distort who gets the invention, what inventions the startup works on, and how much R&D the startup does. In laissez faire, with competing vertically differentiated incumbents, there are distortions in all three dimensions. We propose a simple rule that solves the first, and limits the second and third, while also requiring very little action by an antitrust authority along the equilibrium path.

Winner, Best Paper by a Junior Author, IIOC 2021

With Yasin Ozcan, NBER & Bhaven Sampat, Columbia Med School
Replication Data

Scientific knowledge is cited both on the "front page" and in the specification text of patents. We present the first dataset of in-text citations, tracking all articles 244 prominent journals since 1976, in all US patents from 1984 to March 2018. Most in-text citations are not on the front page, and most front page citations are not in-text. We present three empirical exercises suggesting that, in line with their legal purpose, in-text citations are more likely to represent real knowledge flows to inventors. The dataset used in this draft is publicly available.

Replication Data

Using six years of data of "star" flyouts involving 226 economics PhD students, I show that stars are more male and more international than PhD students as a whole, that the geographic concentration of stars is even more American post-PhD than during the PhD, that nearly all stars have a technical or economics undergraduate degree, that the private sector remains uncompetitive for top economics students, and that RA work with economics faculty is rapidly becoming a common route to post-PhD job market success. Many of these results about where "stars" come from, what they do, and where they go are strikingly different from existing research on the economics market, which tends to focus on the sample of all PhD students.

With Joshua Gans, University of Toronto

Ridesharing platforms like Uber compete on both price and wait time. When I add idled drivers, I flatten the distribution of wait times for consumers, making them more homogenous. This increases willingness-to-pay, makes customers more homogenous and hence easier to extract consumer surplus from, but also increases the incentive for rivals to steal customers with lower prices. A planner cares only about the first effect, the second makes private firms more willing to provide idleness, and the third makes private firms less willing to do so. Multihoming on the driver or consumer side constrains the extent to which idled drivers can be provided. Combining these ideas, the optimal competitive market structure for a ridesharing network could be monopoly or competition under different multihoming configurations, depending on the relative strength of the two wedges between the planner and firms.

With Jorge Lemus, UIUC
Online Appendix B & Appendix C

Long ago, Arrow and others noted that innovations do not meet the usual assumptions necessary for efficient decentralized production, hence the necessity of policies like patents or prize contests to rectify the market's inefficiency. Theoretical models exploring innovation policy tend to focus either on the amount of effort exerted on innovation, or on the extent of spillovers from one firm to another. We are interested not in how hard firms work, but on which types of projects they work. Using some neat tricks to make a very general model of innovation direction tractable, we show that every policy which rewards inventors and depends only on inventions that are actually seen will distort innovation direction. Patents, prizes, and laissez faire are all examples of such distortionary policies.

With Andras Tilcsik, Toronto Rotman and Brooklynn Zhu, BI Norwegian

Coachability is often considered critical for entrepreneurs. The difficulty comes in separating entrepreneurs who are more skilled in a particular area from those who are simply stubborn. Examining four years of participating firms in a science-based incubator, we find that firms who ignore milestones suggested by experienced mentors appear equally successful ex-post as those who follow those milestones, and that even with deep knowledge of firm and founder backgrounds, we are unable to separate the "stubborn" from the "skilled". We argue that incubators, grant programs, and other settings where founder information is much more limited than in our setting will find it equally challenging to identify "coachability" in a usable way.