We develop new quasi-experimental tools to understand algorithmic discrimination and build non-discriminatory algorithms when the outcome of interest is only selectively observed. We first show that algorithmic discrimination arises when the available algorithmic inputs are systematically different for individuals with the same objective potential outcomes. We then show how algorithmic discrimination can be eliminated by measuring and purging these conditional input disparities. Leveraging the quasi-random assignment of bail judges in New York City, we find that our new algorithms not only eliminate algorithmic discrimination but also generate more accurate predictions by correcting for the selective observability of misconduct outcomes.

Human decision-makers frequently override the recommendations generated by predictive algorithms, but it is unclear whether these discretionary overrides add valuable private information or reintroduce human biases and mistakes. We develop new quasi-experimental tools to measure the impact of human discretion over an algorithm on the accuracy of decisions, even when the outcome of interest is only selectively observed, in the context of bail decisions. We find that 90% of the judges in our setting underperform the algorithm when they make a discretionary override, with most making override decisions that are no better than random. Yet the remaining 10% of judges outperform the algorithm in terms of both accuracy and fairness when they make a discretionary override. We provide suggestive evidence on the behaviour underlying these differences in judge performance, showing that the high-performing judges are more likely to use relevant private information and are less likely to overreact to highly salient events compared to the low-performing judges.

We ask whether increased public scrutiny leads to the more effective use of predictive algorithms. We focus on the context of bail, where judges face heightened public scrutiny during competitive partisan elections. We find that judges up for reelection are much more likely to follow the algorithmic recommendation to detain high-risk defendants just before an election. However, release decisions return to normal shortly after the election, and there is little change in pretrial misconduct rates, indicating that heightened public scrutiny, at least through competitive partisan elections, will not lead to the more effective use of predictive algorithms in bail.

We develop new quasi-experimental tools to measure disparate impact, regardless of its source, in the context of bail decisions.We show that omitted variables bias in pretrial release rate comparisons can be purged by using the quasi-random assignment of judges to estimate average pretrial misconduct risk by race.We find that two-thirds of the release rate disparity between white and Black defendants in New York City is due to the disparate impact of release decisions.We then develop a hierarchical marginal treatment effect model to study the drivers of disparate impact, finding evidence of both racial bias and statistical discrimination.

We experimentally test several approaches to increasing the demand for workers with a criminal record on a nationwide staffing platform by addressing potential downside risk and productivity concerns. The staffing platform asked hiring managers to make a series of hypothetical hiring decisions that affected whether workers with a criminal record could accept their jobs in the future. We find that 39% of businesses in our sample are willing to work with individuals with a criminal record at baseline, which rises to over 50% when businesses are offered crime and safety insurance, a single performance review, or a limited background check covering just the past year. Wage subsidies can achieve similar increases but at a substantially higher cost. Based on our findings, the staffing platform relaxed the criminal background check requirement and offered crime and safety insurance to interested businesses.

This article tests for bias in consumer lending using administrative data from a high-cost lender in the U.K. We motivate our analysis using a new principal-agent model of bias where loan examiners are incentivized to maximize a short-term outcome, not long-term profits, leading to bias against illiquid applicants at the margin of loan decisions. We identify the profitability of marginal applicants using the quasi-random assignment of loan examiners, finding significant bias against immigrant and older applicants when using the firm's preferred measure of long-run profits but not when using the short-run measure used to evaluate examiner performance. In this case, market incentives based on characteristics that vary across groups lead to inefficient group-based bias.

Algorithmic decision-making can lead to discrimination against legally protected groups, but measuring such discrimination is often hampered by a fundamental selection challenge. We develop new quasi-experimental tools to overcome this challenge and measure algorithmic discrimination in pretrial bail decisions. We show that the selection challenge reduces to the challenge of measuring four moments, which can be estimated by extrapolating quasi-experimental variation across as-good-as-randomly assigned decision-makers. Estimates from New York City show that both a sophisticated machine learning algorithm and a simpler regression model discriminate against Black defendants even though defendant race and ethnicity are not included in the training data.

In this article, we review a growing empirical literature on the effectiveness and fairness of the US pretrial system and discuss its policy implications. Despite the importance of this stage of the criminal legal process, researchers have only recently begun to explore how the pretrial system balances individual rights and public interests. We describe the empirical challenges that have prevented progress in this area and how recent work has made use of new data sources and quasi-experimental approaches to credibly estimate both the individual harms (such as loss of employment or government assistance) and public benefits (such as preventing non-appearance at court and new crimes) of cash bail and pretrial detention. These new data and approaches show that the current pretrial system imposes substantial short-and long-term economic harms on detained defendants in terms of lost earnings and government assistance, while providing little in the way of decreased criminal activity for the public interest. Non-appearances at court do significantly decrease for detained defendants, but the magnitudes cannot justify the economic harms to individuals observed in the data. A second set of studies shows that that the costs of cash bail and pretrial detention are disproportionately borne by Black and Hispanic individuals, giving rise to large and unfair racial differences in cash bail and detention that cannot be explained by underlying differences in pretrial misconduct risk. We then turn to policy implications and describe areas of future work that would enable a deeper understanding of what drives these undesirable outcomes.

We measure the economic costs of the US pretrial system using several complementary approaches and data sources. The pretrial system operates as one of the earliest points of entry in the criminal justice system. It typically represents an individual's first opportunity to be incarcerated, potentially leading to subsequent long-term damage in the form of family separation, work interruption, loss of housing, and so on. We find that individuals lose almost $30,000 in forgone earnings and social benefits when detained in jail while awaiting the resolution of their criminal cases. These adverse consequences are also present in aggregate measures of economic well-being, with increases in county pretrial detention rates associated with increases in poverty rates and decreases in employment rates. Counties with high levels of pretrial detention also exhibit significantly lower levels of intergenerational mobility among children, consistent with pretrial detention having an adverse impact on young children who may be the dependents of individuals affected by the pretrial system.

We develop new quasi-experimental tools to measure racial discrimination in the context of bail decisions. Observational comparisons of white and black pretrial release rates suffer from omitted variables bias when there are unobserved racial differences in pretrial misconduct potential. We show that the bias in these observational comparisons is a function of average white and black misconduct risk, which can be estimated from the quasi-random assignment of bail judges. Estimates from New York City show that less than one-third of the release rate disparity between white and black defendants is explained by unobserved differences in misconduct potential, with more than two-thirds explained by racial discrimination. We then develop a hierarchical marginal treatment effects model that imposes additional structure on the quasi-experimental variation to investigate the drivers of this discrimination. Model estimates show that discrimination in bail decisions is driven by both racial bias and statistical discrimination, with the latter coming from a higher level of average risk and less precise risk signals for black defendants. David Arnold Industrial Relations Section Louis A. Simpson International Bldg. Princeton University Princeton, NJ 08544-2098 dharnold@princeton.edu Will S. Dobbie Harvard Kennedy School 79 John F. Kennedy St. Cambridge, MA 02138 and NBER will_dobbie@hks.harvard.edu Peter Hull University of Chicago 5757 South University Avenue Chicago, IL 60637 and NBER hull@uchicago.edu

ABSTRACTWe study the financial and labor market impacts of bad credit reports. Using difference‐in‐differences variation from the staggered removal of bankruptcy flags, we show that bankruptcy flag removal leads to economically large increases in credit limits and borrowing. Using administrative tax records linked to personal bankruptcy records, we estimate economically small effects of flag removal on employment and earnings outcomes. We rationalize these contrasting results by showing that, conditional on basic observables, “hidden” bankruptcy flags are strongly correlated with adverse credit market outcomes but have no predictive power for measures of job performance.

In Arnold, Dobbie, and Yang (2018, ADY), we find that marginally released white defendants have higher rates of pre-trial misconduct than marginally released black defendants. We interpret these findings as evidence of racial bias against black defendants through the lens of the marginal outcome test originally developed by Becker (1957). Canay, Mogstad, and Mountjoy (2020, CMM) question the interpretation of our empirical findings and the logical validity of the marginal outcome test. However, CMM’s conclusions are based on an incomplete definition of racial bias that is different from the one used in ADY. Under ADY’s definition of bias, the marginal outcome test is logically valid and a useful tool for studying discrimination in real-world settings.

We study the drivers of financial distress using a large-scale field experiment that offered randomly selected borrowers a combination of (i) immediate payment reductions to target short-run liquidity write-downs to target long-run debt constraints. We identify the separate effects of the payment reductions and interest write-downs using both the experiment and cross-sectional variation in treatment intensity. We find that the interest write-downs significantly improved both financial and labor market outcomes, despite not taking effect for three to five years. In sharp contrast, there were no positive effects of the more immediate payment reductions. These results run counter to the widespread view that financial distress is largely the result of short-run constraints. (JEL G56, K35)

We estimate the impact of charter schools on early-life labor market outcomes in Texas. We find that, at the mean, charter schools have no impact on test scores and a negative impact on earnings. No Excuses charter schools increase test scores and 4-year college enrollment but have a statistically insignificant impact on earnings, although the coefficient is almost identical to what one would expect given the correlation between test scores and wages. Other types of charter schools decrease test scores, 4-year college enrollment, and earnings, and surprisingly the decrease in wages is more negative than one would anticipate.

In this Article, we provide a new statistical and legal framework to understand the legality and fairness of predictive algorithms under the Equal Protection Clause. We begin by reviewing the main legal concerns regarding the use of protected characteristics such as race and the correlates of protected characteristics such as criminal history. The use of race and nonrace correlates in predictive algorithms generates direct and proxy effects of race, respectively, that can lead to racial disparities that many view as unwarranted and discriminatory. These effects have led to the mainstream legal consensus that the use of race and nonrace correlates in predictive algorithms is both problematic and potentially unconstitutional under the Equal Protection Clause. This mainstream position is also reflected in practice, with all commonly used predictive algorithms excluding race and many excluding nonrace correlates such as employment and education. Next, we challenge the mainstream legal position that the use of a protected characteristic always violates the Equal Protection Clause. We develop a statistical framework that formalizes exactly how the direct and proxy effects of race can lead to algorithmic predictions that disadvantage minorities relative to nonminorities. While an overly formalistic solution requires exclusion of race and all potential nonrace correlates, we show that this type of algorithm is unlikely to work in practice because nearly all algorithmic inputs are correlated with race. We then show that there are two simple statistical solutions that can eliminate the direct and proxy effects of race, and which are implementable even when all inputs are correlated with race. We argue that our proposed algorithms uphold the principles of the equal protection doctrine because they ensure that individuals are not treated differently on the basis of membership in a protected class, in stark contrast to commonly used algorithms that unfairly disadvantage minorities despite the exclusion of race. We conclude by empirically testing our proposed algorithms in the context of the New York City pretrial system. We show that nearly all commonly used algorithms violate certain principles underlying the Equal Protection Clause by including variables that are correlated with race, generating substantial proxy effects that unfairly disadvantage Black individuals relative to white individuals. Both of our proposed algorithms substantially reduce the number of Black defendants detained compared to commonly used algorithms by eliminating these proxy effects. These findings suggest a fundamental rethinking of the equal protection doctrine as it applies to predictive algorithms and the folly of relying on commonly used algorithms.