Search

In observational studies of discrimination, the most common statistical approaches consider either the rate at which decisions are made (benchmark tests) or the success rate of those decisions (outcome tests). Both tests, however, have well-known statistical limitations, sometimes suggesting discrimination even when there is none. Despite the fallibility of the benchmark and outcome tests individually, here we prove a surprisingly strong statistical guarantee: under a common non-parametric assumption, at least one of the two tests must be correct; consequently, when both tests agree, they are guaranteed to yield correct conclusions. We present empirical evidence that the underlying assumption holds approximately in several important domains, including lending, education, and criminal justice -- and that our hybrid test is robust to the moderate violations of the assumption that we observe in practice. Applying this approach to 2.8 million police stops across California, we find evidence of widespread racial discrimination.

Regulatory efforts to protect against algorithmic bias have taken on increased urgency with rapid advances in large language models (LLMs), which are machine learning models that can achieve performance rivaling human experts on a wide array of tasks. A key theme of these initiatives is algorithmic "auditing," but current regulations -- as well as the scientific literature -- provide little guidance on how to conduct these assessments. Here we propose and investigate one approach for auditing algorithms: correspondence experiments, a widely applied tool for detecting bias in human judgements. In the employment context, correspondence experiments aim to measure the extent to which race and gender impact decisions by experimentally manipulating elements of submitted application materials that suggest an applicant's demographic traits, such as their listed name. We apply this method to audit candidate assessments produced by several state-of-the-art LLMs, using a novel corpus of applications to K-12 teaching positions in a large public school district. We find evidence of moderate race and gender disparities, a pattern largely robust to varying the types of application material input to the models, as well as the framing of the task to the LLMs. We conclude by discussing some important limitations of correspondence experiments for auditing algorithms.

Managers, employers, policymakers, and others often seek to understand whether decisions are biased against certain groups. One popular analytic strategy is to estimate disparities after adjusting for observed covariates, typically with a regression model. This approach, however, suffers from two key statistical challenges. First, omitted-variable bias can skew results if the model does not adjust for all relevant factors; second, and conversely, included-variable bias -- a lesser-known phenomenon -- can skew results if the set of covariates includes irrelevant factors. Here we introduce a new, three-step statistical method, which we call risk-adjusted regression, to address both concerns in settings where decision makers have clearly measurable objectives. In the first step, we use all available covariates to estimate the value, or inversely, the risk, of taking a certain action, such as approving a loan application or hiring a job candidate. Second, we measure disparities in decisions after adjusting for these risk estimates alone, mitigating the problem of included-variable bias. Finally, in the third step, we assess the sensitivity of results to potential mismeasurement of risk, addressing concerns about omitted-variable bias. To do so, we develop a novel, non-parametric sensitivity analysis that yields tight bounds on the true disparity in terms of the average gap between true and estimated risk -- a single interpretable parameter that facilitates credible estimates. We demonstrate this approach on a detailed dataset of 2.2 million police stops of pedestrians in New York City, and show that traditional statistical tests of discrimination can substantially underestimate the magnitude of disparities.

The field of fair machine learning aims to ensure that decisions guided by algorithms are equitable. Over the last decade, several formal, mathematical definitions of fairness have gained prominence. Here we first assemble and categorize these definitions into two broad families: (1) those that constrain the effects of decisions on disparities; and (2) those that constrain the effects of legally protected characteristics, like race and gender, on decisions. We then show, analytically and empirically, that both families of definitions typically result in strongly Pareto dominated decision policies. For example, in the case of college admissions, adhering to popular formal conceptions of fairness would simultaneously result in lower student-body diversity and a less academically prepared class, relative to what one could achieve by explicitly tailoring admissions policies to achieve desired outcomes. In this sense, requiring that these fairness definitions hold can, perversely, harm the very groups they were designed to protect. In contrast to axiomatic notions of fairness, we argue that the equitable design of algorithms requires grappling with their context-specific consequences, akin to the equitable design of policy. We conclude by listing several open challenges in fair machine learning and offering strategies to ensure algorithms are better aligned with policy goals.