Research

What should regulators of complex algorithms regulate? We propose a model of oversight over 'black-box' algorithms used in high-stakes applications such as lending, medical testing, or hiring. In our model, a regulator is limited in how much she can learn about a black-box model deployed by an agent with misaligned preferences. The regulator faces two choices: first, whether to allow for the use of complex algorithms; and second, which key properties of algorithms to regulate. We show that limiting agents to algorithms that are simple enough to be fully transparent is inefficient as long as the misalignment is limited and complex algorithms have sufficiently better performance than simple ones. Allowing for complex algorithms can improve welfare, but the gains depend on how the regulator regulates them. Regulation that focuses on the overall average behavior of algorithms, for example based on standard explainer tools, will generally be inefficient. Targeted regulation that focuses on the source of incentive misalignment, e.g., excess false positives or racial disparities, can provide second-best solutions. We provide empirical support for our theoretical findings using an application in consumer lending, where we document that complex models regulated based on context-specific explanation tools outperform simple, fully transparent models. This gain from complex models represents a Pareto improvement across our empirical applications that is preferred both by the lender and from the perspective of the financial regulator.

"Right-to-counsel" programs provide free legal assistance to tenants in eviction court. Legal assistance can delay or prevent eviction. However, large-scale legal assistance programs can also generate costs for tenants due to equilibrium rental market responses. In this paper, we study how right to counsel impacts rental markets when implemented at scale, and quantify the policy's impact on tenant welfare. Leveraging the geographic rollout of New York City's program, we find listed rent prices rose by $22-$38/month within two years of policy implementation, with larger increases in areas with higher baseline eviction rates. We do not find evidence that landlords adjusted on other margins, such as tenant screening or improvements to habitability. Guided by these results, we develop a framework to evaluate the policy's welfare implications for tenants, incorporating the trade-off between protection from eviction and higher rent prices. We quantify the parameters of our framework using linked data on eviction court cases, rental housing listings, and tenant earnings trajectories. Despite the direct benefits and insurance value of stronger eviction protections, the estimated price increases are large enough to generate a small net reduction in ex-ante tenant welfare.

Recent research has documented the prevalence and consequences of evictions, but our understanding of underlying drivers of the eviction rate and the scope for policy to affect it remains limited. In this paper, we study landlords' decisions to evict tenants and how these decisions may be influenced by policy. We combine novel lease-level ledger data from low-income rental markets with a model of the landlord's eviction decision to characterize the persistence of shocks to tenant default risk, landlords' information about these shocks, and landlords' cost of eviction. Our data show that nonpayment is common, is frequently tolerated by landlords, and is often followed by recovery, suggesting that landlords face a trade-off between initiating a costly eviction or waiting to learn whether a tenant can continue paying. Our dynamic discrete choice model of the eviction decision captures this trade-off. Estimates indicate that filing an eviction costs landlords the equivalent of 2-3 months of rent, and that the majority of evictions involve tenants who are unlikely to pay going forward. This implies that uniformly applied policies can generate additional forbearance for tenants, but they do not prevent most evictions. We find that 15% of those evicted would have resumed paying rent, suggesting a role for more targeted interventions. Among the policy instruments we consider, direct financial incentives for landlords---such as taxes and subsidies---are more likely to durably prevent evictions than procedural delays.

We show that widely used consumer credit scores exhibit disparities in classification errors: more minority applicants predicted not to repay a loan would have repaid, and more minority applicants predicted to repay do not. These disparities arise because of disparities in noise, not bias, in predicting risk. Three candidate solutions fail to reduce credit score noise disparities: machine learning prediction technology; building group-specific credit scoring models; and adding non-traditional credit bureau data. Rather, up to 70% of these disparities result from greater data sparsity for disadvantaged groups, suggesting that credit scoring can make historical disparities in credit access more persistent.

Over 30m US adults do not use formal consumer credit. How many of these are inefficiently excluded because they lack a credit history or have a poor credit score? We develop a framework to characterize the efficiency-maximizing system of credit histories and credit scoring, subject to the constraints imposed by the severity of adverse selection, and by the ability of credit histories to predict future risk. We find US consumer credit features a moderate amount of adverse selection and persistent consumer types. This adverse selection generates substantial welfare loss: a majority of today's non-borrowers would be first-best efficient to lend to. Credit reporting helps alleviate the costs of adverse selection, with the current US system recovering roughly two-thirds of the welfare that would be lost in a no-credit-reporting counterfactual, relative to a full-information first-best. We find that requiring histories to be shorter -- or to forget past default sooner -- would induce some market unraveling but also would help non-borrowing consumers escape the "no history trap."

The regulation of algorithmic decisions, ranging from credit scoring to employment screening, presents unique challenges and novel opportunities for achieving regulatory goals. We analyze a framework for algorithmic regulation that emphasizes the importance of temporal stages in the regulatory pipeline: ex-ante (pre-training), ex-interim (post-training but pre-deployment), and ex-post (post-deployment). Regulators can choose both the pipeline stage targeted by the legal rule ("rule timing") and the stage at which compliance is assessed ("scrutiny timing"). We situate emerging and proposed AI regulations within this framework and analyze the tradeoffs between different regulatory regimes. We highlight how ex-interim rules offer a unique opportunity in algorithmic settings compared to the rigidity of ex-ante rules or the bluntness of ex-post rules and explore the considerations that guide whether regulators might scrutinize ex-interim rules before or after deployment.

Recent federal policy shifts underscore the urgency of reassesing disparate impact’s role in modern markets. Although the requirement to consider less discriminatory alternatives (LDAs) has long been part of disparate impact analyses under fair lending, employment, and housing law, it has received little practical or doctrinal attention. That is starting to change. We offer, focusing on the fair lending setting, a framework for LDA analysis that confronts a core tension: the tradeoff between disparities in outcomes (who gets approved for loans and on what terms) and disparities in validity (how well risk is measured across groups). Efforts to reduce disparities along one dimension can often worsen the other. Through a simulation exercise, we show how LDA analysis can help navigate these competing fairness goals, and inform what is required for disparate impact to endure as both a viable legal doctrine and a compelling vision of fairness.

Public finance and real estate are uniquely intertwined in China, where local governments also serve as monopolist sellers of land. We shed new light on how land sale decisions and land prices depend on local governments’ financing objectives. First, we document the large (ten-fold) price premium paid for residential-zoned relative to industrial-zoned land and show this price premium can be explained by the greater future tax revenues generated by industrial land; the choice to sell land as industrial rather than residential generates an IRR of 7.70%, which is comparable to local governments’ cost of capital in bond markets. Second, local governments are sensitive to financing constraints: industrial land supply decreases with governments’ bond yields. Third, local governments’ land sales are sensitive to the intergovernmental tax sharing, such that industrial land sales increase with the share of taxes captured by local governments. Thus, shocks to local public finances can be expected to affect the Chinese real estate market and vice versa.