Industrial Organization Program Meeting, Spring 2021

Claussen, Peukert, and Sen run a field experiment to quantify the economic returns to data and informational externalities associated with algorithmic recommendation relative to human curation in the context of online news. Their results show that personalized recommendation can outperform human curation in terms of user engagement, though this crucially depends on the amount of personal data. Limited individual data or breaking news leads the editor to outperform the algorithm. Additional data helps algorithmic performance but diminishing economic returns set in rapidly. Investigating informational externalities highlights that personalized recommendation reduces consumption diversity. Moreover, users associated with lower levels of digital literacy and more extreme political views engage more with algorithmic recommendations.

Battery storage offers a potentially valuable complement to renewable energy. Recognizing this, policymakers have recently incentivized and mandated storage as a means to integrate renewable energy and meet climate goals. This paper evaluates the equilibrium value and adoption trajectory of utility-scale batteries using California data, focusing on the impact of falling battery capital costs, complementarities with renewable energy, and market power. Butters, Dorsey, and Gowrisankaran add three key modeling features relative to the literature: (1) a modeling of equilibrium effects from large-scale batteries that includes ramping constraints, (2) a frontier high-frequency forecasting model of load and prices, and (3) linked competitive dynamic equilibrium models of battery adoption and operations. They find that: (1) the value of battery storage increased sharply between 2016-19 as solar generation increased, (2) battery investment exhibits decreasing returns-to-scale--the per-unit value of batteries drops significantly with total installed battery capacity, (3) battery operations increase California's 2018 expected discounted social surplus from the electricity market by $3.8 billion or $2.42 per MWh of solar energy generated, and (4) California would require a 35% subsidy on batteries to meet its 2024 storage mandate of 1.3 GW of power capacity.

Off label prescriptions represent more than 20 % of drug spending and treatment choices. Tuncel and the researchers investigate the impact of regulation of off-label prescriptions on physicians' behavior, patients' health, treatment costs and pharmaceutical firms' pricing with a structural model of demand and supply. Exploiting rich panel data on physicians' activities and office visits in France over a nine-year period, the researchers develop a model of prescription choice and health outcomes with unobserved patient-level heterogeneity. They identify the demand for on-label and off-label drugs and the effect of prescription choice on health outcomes. Results show that there is selection into treatment based on patient-level unobservables. Consistent with the medical literature, the researchers find that there is heterogeneity in treatment impact of drugs. On the supply side, they develop a Nash-in-Nash bargaining model between the government and the pharmaceutical companies that allows the identification of the marginal costs of drugs. Counterfactual simulations show that when the researchers remove off-label drugs from the choice set of physicians, substitution to on-label drugs at constant prices would lead to an increase in the expenditure on prescription drugs. If they allow bargaining adjustment on drug prices under a ban on off-label prescriptions, a ban would further increase the treatment cost without leading to an improvement in health outcomes.

An increasing number of decisions are guided by machine learning algorithms. In many settings, from consumer credit to criminal justice, those decisions are made by applying an estimator to data on an individual's observed behavior. But when consequential decisions are encoded in rules, individuals may strategically alter their behavior to achieve desired outcomes. This paper develops a new class of estimator that is stable under manipulation, even when the decision rule is fully transparent. Bjorkegren, Blumenstock, and Knight explicitly model the costs of manipulating different behaviors, and identify decision rules that are stable in equilibrium. Through a large field experiment in Kenya, the researchers show that decision rules estimated with their strategy-robust method outperform those based on standard supervised learning approaches.

Models of firm conduct are the cornerstone of both theoretical and empirical work in industrial organization. A recent contribution (Berry and Haile, 2014) has suggested the use of exclusion restrictions to test alternative conduct models. Backus, Conlon, and Sinkinson propose a pairwise testing procedure based on this idea and show that the power of the test to discriminate between models is tied to the formulation of those restrictions as moments and how they reflect the nonlinearity of equilibrium markups. The researchers apply this test to the ready-to-eat cereal market using detailed scanner and consumer data to evaluate the “common ownership” hypothesis, which has received significant attention. Although Backus, Conlon, and Sinkinson show that the potential magnitude of common ownership effects would be large, their test finds that standard own-firm profit maximization is more consistent with the data.

While there is a vast body of research on the benefits of FDI in developing countries, whether and how the form of FDI matters have received limited attention. In this paper, Bai, Barwick, Cao, and Li study the impact of FDI via quid pro quo (technology for market access) on facilitating knowledge spillover and quality upgrading. Their context is the Chinese automobile industry, where foreign automakers are required to set up joint ventures (the "quid") with domestic automakers in return for market access (the "quo"). The identification strategy exploits a unique dataset of detailed vehicle quality measures along multiple dimensions and relies on within-product quality variation across dimensions. The researchers show that affiliated domestic automakers adopt more similar quality strength as their joint ventures, compared to non-affiliated pairs. The results suggest that quid pro quo generates additional knowledge spillover to affiliated domestic automakers, on top of any industry-wide spillover. Bai, Barwick, Cao, and Li rule out endogenous joint venture network formation, overlapping customer bases, or direct technology transfer via market transactions as alternative explanations. Analyses leveraging additional micro datasets on worker flows and shared upstream suppliers among automakers demonstrate that labor mobility and supplier network are important channels in mediating knowledge spillover. On the other hand, while quid pro quo facilitates learning, such a requirement is not a prerequisite for knowledge spillover. Counterfactual exercises show that quid pro quo is not the primary driver of the overall quality improvement experienced by domestic automakers.

Digital platforms are increasingly the subject of regulatory scrutiny. In comparison to multiple competitors, a single platform may increase consumer welfare if network effects are large or may decrease welfare due to higher prices or reduction in platform variety. Farronato, Fong, and Fradkin study the net effect of this trade-off in the context of the merger between the two largest platforms for pet-sitting services. They exploit variation in pre-merger market shares and a difference-in-differences approach to causally estimate network effects at the platform and market level. The researchers find that consumers are, on average, not substantially better off with a single combined platform than with two separate and competing platforms. On one hand, users of the acquiring platform benefited from the merger because of network effects. On the other hand, users of the acquired platform experienced worse outcomes. Their results highlight the importance of platform differentiation even when platforms enjoy network effects.

Sahni and Zhang analyze a large-scale randomized field experiment in which a search engine varied the prominence of search ads for 3.3 million US users: one group of users saw the status quo, while the other saw a lower level of advertising (with prominence of search ads decreased). Revealed preference data reject that users are, overall, averse to search advertising targeted to them across a diverse set of searches. At the margin, users prefer the search engine with the higher level of advertising. On the supply side, newer websites are more likely to advertise. Going from the lower to the higher level of advertising increases traffic to newer websites, with the newest decile of websites gaining traffic by 10%. Users also respond more positively to advertising when local businesses in their state create new websites. Taken together, patterns in researchers' data are consistent with an equilibrium in which advertising compensates for important information gaps in organic listings: it conveys relevant new information, which is hard for the search engine to gather, and therefore missed by the organic listings algorithm. Viewing search ads, at the margin Sahni and Zhang study, makes consumers better off on average.

Choice screen auctions have recently been deployed in 31 European countries, allowing consumers to choose their preferred search engine on Google’s Android platform instead of automatically defaulting them to Google’s own search engine. Ostrovsky shows that a seemingly minor detail in the design of these auctions--whether they are conducted on a “per appearance” or a “per install” basis--plays a major role in the mix and characteristics of auction winners, and consequently in their expected overall market share. The researcher also shows that “per install” auctions distort the incentives of alternative search engines toward extracting as much revenue as possible from each user who installs them, at the expense of lowering the expected number of such users. The distortion gets worse as the auction gets more competitive and the number of bidders increases. Empirical evidence from Android choice screen auctions conducted in 2020 is consistent with his theoretical results.