Search

Algorithms (in some form) are already widely used in the criminal justice system. We draw lessons from this experience for what is to come for the rest of society as machine learning diffuses. We find economists and other social scientists have a key role to play in shaping the impact of algorithms, in part through improving the tools used to build them.

Algorithms trained to predict mismeasured proxy variables can reproduce and scale up racial bias. This mechanism of algorithmic bias is distinct from others in the literature and harder to detect. We show this using examples from health care, but the forces we consider apply to a range of other important social sectors.

There are widespread concerns that the growing use of machine learning algorithms in important decisions may reproduce and reinforce existing discrimination against legally protected groups. Most of the attention to date on issues of “algorithmic bias” or “algorithmic fairness” has come from computer scientists and machine learning researchers. We argue that concerns about algorithmic fairness are at least as much about questions of how discrimination manifests itself in data, decision-making under uncertainty, and optimal regulation. To fully answer these questions, an economic framework is necessary--and as a result, economists have much to contribute.

A debt trap occurs when someone takes on a high-interest-rate loan and is barely able to pay back the interest, and thus perpetually finds themselves in debt (often by refinancing). Studying such practices is important for understanding financial decision-making of households in dire circumstances, and also for setting appropriate consumer protection policies. We conduct a simple experiment in three sites in which we paid off high-interest moneylender debt of individuals. Most borrowers returned to debt within six weeks. One to two years after intervention, treatment individuals were borrowing at the same rate as control households. (JEL D14, D18, D91)

Concerns about the dissemination of spurious results have led to calls for pre-analysis plans (PAPs) to avoid ex-post “p-hacking.” But often the conceptual hypotheses being tested do not imply the level of specificity required for a PAP. In this paper we suggest a framework for PAPs that capitalize on the availability of causal machine-learning (ML) techniques, in which researchers combine specific aspects of the analysis with ML for the flexible estimation of unspecific remainders. A “cheap-lunch” result shows that the inclusion of ML produces limited worst-case costs in power, while offering a substantial upside from systematic specification searches.

Concerns that algorithms may discriminate against certain groups have led to numerous efforts to ‘blind’ the algorithm to race. We argue that this intuitive perspective is misleading and may do harm. Our primary result is exceedingly simple, yet often overlooked. A preference for fairness should not change the choice of estimator. Equity preferences can change how the estimated prediction function is used (e.g., different threshold for different groups) but the function itself should not change. We show in an empirical example for college admissions that the inclusion of variables such as race can increase both equity and efficiency.

Machine learning tools are beginning to be deployed en masse in health care. While the statistical underpinnings of these techniques have been questioned with regard to causality and stability, we highlight a different concern here, relating to measurement issues. A characteristic feature of health data, unlike other applications of machine learning, is that neither y nor x is measured perfectly. Far from a minor nuance, this can undermine the power of machine learning algorithms to drive change in the health care system--and indeed, can cause them to reproduce and even magnify existing errors in human judgment.

Machines are increasingly doing “intelligent” things. Face recognition algorithms use a large dataset of photos labeled as having a face or not to estimate a function that predicts the presence y of a face from pixels x. This similarity to econometrics raises questions: How do these new empirical tools fit with what we know? As empirical economists, how can we use them? We present a way of thinking about machine learning that gives it its own place in the econometric toolbox. Machine learning not only provides new tools, it solves a different problem. Specifically, machine learning revolves around the problem of prediction, while many economic applications revolve around parameter estimation. So applying machine learning to economics requires finding relevant tasks. Machine learning algorithms are now technically easy to use: you can download convenient packages in R or Python. This also raises the risk that the algorithms are applied naively or their output is misinterpreted. We hope to make them conceptually easier to use by providing a crisper understanding of how these algorithms work, where they excel, and where they can stumble—and thus where they can be most usefully applied.

Economists have become increasingly interested in studying the nature of production functions in social policy applications, with the goal of improving productivity. Traditionally models have assumed workers are homogenous inputs. However, in practice, substantial variability in productivity means the marginal productivity of labor depends substantially on which new workers are hired--which requires not an estimate of a causal effect, but rather a prediction. We demonstrate that there can be large social welfare gains from using machine learning tools to predict worker productivity, using data from two important applications - police hiring and teacher tenure decisions.

Targeting assistance to the poor is a central problem in development. We study the problem of designing a proxy means test when the implementing agent is corruptible. Conditioning on more poverty indicators may worsen targeting in this environment because of a novel tradeoff between statistical accuracy and enforceability. We then test necessary conditions for this tradeoff using data on Below Poverty Line card allocation in India. Less eligible households pay larger bribes and are less likely to obtain cards, but widespread rule violations yield a de facto allocation much less progressive than the de jure one. Enforceability appears to matter. (JEL D12, I32, I38, O12, O15)

In standard models with asymmetric information, the parties involved are assumed to have private information about their own characteristics. In the health insurance market, for example, customers are typically assumed to know more about their health status than insurers do, and if the customers use this information in deciding whether to buy insurance, we have a classic case of adverse selection. Modern data-gathering technologies, however, can reverse this situation. For example, because cell-phone providers keep and analyze detailed records, they can know more about a consumer’s expected usage than the customer herself does. Similarly, a credit card company may know more about a customer’s probability of incurring a late fee than the customer herself. We explore the consequences of this reversal in the information asymmetry. 1 A bare-bones model examines how providing consumers with information about their own usage affects prices and welfare. We first establish the rather obvious result that, taking prices as given, consumers benefit from having better information about their own usage: they can choose the pricing plans that are more suitable for them. Recent

Economists have long been interested in the idea that there is a direct circular relation between poverty and low productivity, and not just one that is mediated by market failures, usually in asset markets. The nutrition-based efficiency wage model (Partha Dasgupta and Debraj Ray, 1987) is the canonical example of models where this happens: However it has been variously suggested (see for example T. N. Srinivasan, 1994) that the link from nutrition to productivity and especially the link from productivity to nutrition is too weak to be any more than a small part of the story. Partha Dasgupta himself acknowledges this when he writes "nutrition-productivity construct provides a metaphor,..., for ... an economic environment harboring poverty traps" (Partha Dasgupta, 1997, page 5). We propose an alternative approach to this question based on the idea that attention is a scarce resource that is important for productivity. Specifically, people may not be able to fully attend to their jobs if they are also worrying about problems at home and being distracted in this way reduces productivity. But not paying attention at home is also costly: early symptoms of a child's sickness may go unnoticed; water may run out at the end of the day; kerosene for lighting lamps at home might run out and make it hard to do homework; etc. Finally, the extent to which home life distracts depends on the nature of home life. Specifically, certain goods (e.g. a good baby sitter, a 24-hour piped water supply, a connection to a power supply grid) can reduce the extent of home life distraction. These three assumptions generate an interesting relation between income and productivity that is at the core of our model. The non-poor in this model, by virtue of owning distraction-saving goods and services at home, are able to focus more on their work. Hence they will be more productive at work and will be able to afford more distraction-saving goods. This simple two-way relationship between income and productivity produces a discontinuity in the relation between human capital and earnings which is certain cases can lead to a poverty trap, even in the absence of any market failures.

In his famous article, “The Nature of the Firm,” Ronald Coase (1937) raised two fundamental questions that have spawned a large body of research: Do firm boundaries affect the allocation of resources? And, what determines where firm boundaries are drawn? While the first of these questions has received some theoretical attention (notably, Oliver Williamson [1975, 1985], Benjamin Klein et al. [1978], and Sanford Grossman and Oliver Hart [1986]), it has largely been ignored empirically. Instead, the empirical work in this area, discussed in the other articles in this session, has addressed the second question by analyzing the determinants of vertical integration. Thus, while we know something about the forces that determine firm boundaries, we know relatively little about how these boundaries affect actual firm behavior. This is a major limitation in our understanding of the nature of the firm. To begin to assess how firm boundaries affect behavior, we analyze whether there are differences between integrated and nonintegrated chemical manufacturers in their investments in production capacity. We focus on producers of vinyl chloride monomer (VCM), the sole use of which is in the production of the widely used waterproof plastic, polyvinyl chloride (PVC). VCM is a homogenous commodity and is traded in relatively liquid markets. Moreover, there is no obvious production link between VCM and PVC other than that one is an input into the other. For example, PVC is not a byproduct of VCM production. Nevertheless, twothirds of VCM producers in our sample are integrated downstream into PVC. The existing literature would ask why we observe this degree of integration. We ask instead whether integrated and nonintegrated VCM producers invest differently in production capacity.

Who sets CEO pay? Our standard answer to this question has been shaped by principal agent theory: shareholders set CEO pay. They use pay to limit the moral hazard problem caused by the low ownership stakes of CEOs. Through bonuses, options, or long term contracts, shareholders can motivate the CEO to maximize firm wealth. In other words, shareholders use pay to provide incentives, a view we refer to as the contracting view. An alternative view, championed by practitioners such as Crystal (1991), argues that CEOs set their own pay. They manipulate the compensation committee and hence the pay process itself to pay themselves what they can. The only constraints they face may be the availability of funds or more general fears, such as not wanting to be singled out in the Wall Street Journal as being overpaid. We refer to this second view as the skimming view. In this paper, we investigate the relevance of these two views.