Risk scores, label bias, and everything but the kitchen sink.

In designing risk assessment algorithms, many scholars promote a "kitchen sink" approach, reasoning that more information yields more accurate predictions. We show, however, that this rationale often fails when algorithms are trained to predict a proxy of the true outcome, for instance, predicting arrest as a proxy for criminal behavior. With this "label bias," one should exclude a feature if its correlation with the proxy and its correlation with the true outcome have opposite signs, conditional on the other model features. This criterion is often satisfied when a feature is weakly correlated with the true outcome, and, additionally, that feature and the true outcome are both direct causes of the proxy outcome. For example, criminal behavior and geography may be weakly correlated and, due to patterns of police deployment, direct causes of one's arrest record-suggesting that excluding geography in criminal risk assessment will weaken an algorithm's performance in predicting arrest but will improve its capacity to predict actual crime.

The disparate impacts of college admissions policies on Asian American applicants.

There is debate over whether Asian American students face additional barriers, relative to white students, when applying to selective colleges. Here we present the results from analyzing 685,709 applications submitted over five application cycles to 11 highly selective colleges (the "Ivy-11"). We estimate that Asian American applicants had 28% lower odds of ultimately attending an Ivy-11 school than white applicants with similar academic and extracurricular qualifications. The gap was particularly pronounced for students of South Asian descent (49% lower odds). Given the high yield rates and competitive financial aid policies of the schools we consider, the disparity in attendance rates is likely driven, at least in part, by admissions decisions. In particular, we offer evidence that this pattern stems from two factors. First, many selective colleges give preference to the children of alumni in admissions. We find that white applicants were substantially more likely to have such legacy status than Asian applicants. Second, we identify geographic disparities potentially reflective of admissions policies that disadvantage students from certain regions of the United States. We hope these results inform discussions on equity in higher education.

Showing high-achieving college applicants past admissions outcomes increases undermatching.

More than 40% of US high school students have access to Naviance, a proprietary tool designed to guide college search and application decisions. The tool displays, for individual colleges, the standardized test scores, grade-point averages, and admissions outcomes of past applicants from a student's high school, so long as a sufficient number of students from previous cohorts applied to a given college. This information is intended to help students focus their efforts on applying to the most suitable colleges, but it may also influence application decisions in undesirable ways. Using data on 70,000 college applicants across 220 public high schools, we assess the effects of access to Naviance on application undermatch, or applying only to schools for which a candidate is academically overqualified. By leveraging variation in the year that high schools adopted the tool, we estimate that Naviance increased application undermatching by more than 50% among 17,000 high-achieving students in our dataset. This phenomenon may be due to increased conservatism: Students may be less likely to apply to colleges when they know their academic qualifications fall below the average of admitted students from their high school. These results illustrate how information on college competitiveness, when not appropriately presented and contextualized, can lead to unintended consequences.

Forgotten but not gone: A multi-state analysis of modern-day debt imprisonment.

In almost every state, courts can jail those who fail to pay fines, fees, and other court debts-even those resulting from traffic or other non-criminal violations. While debtors' prisons for private debts have been widely illegal in the United States for more than 150 years, the effect of courts aggressively pursuing unpaid fines and fees is that many Americans are nevertheless jailed for unpaid debts. However, heterogeneous, incomplete, and siloed records have made it difficult to understand the scope of debt imprisonment practices. We culled data from millions of records collected through hundreds of public records requests to county jails to produce a first-of-its-kind dataset documenting imprisonment for court debts in three U.S. states. Using these data, we present novel order-of-magnitude estimates of the prevalence of debt imprisonment, finding that between 2005 and 2018, around 38,000 residents of Texas and around 8,000 residents of Wisconsin were jailed each year for failure to pay (FTP), with the median individual spending one day in jail in both Texas and Wisconsin. Drawing on additional data on FTP warrants from Oklahoma, we also find that unpaid fines and fees leading to debt imprisonment most commonly come from traffic offenses, for which a typical Oklahoma court debtor owes around $250, or $500 if a warrant was issued for their arrest.

Designing equitable algorithms.

Predictive algorithms are now commonly used to distribute society's resources and sanctions. But these algorithms can entrench and exacerbate inequities. To guard against this possibility, many have suggested that algorithms be subject to formal fairness constraints. Here we argue, however, that popular constraints-while intuitively appealing-often worsen outcomes for individuals in marginalized groups, and can even leave all groups worse off. We outline a more holistic path forward for improving the equity of algorithmically guided decisions.

Measuring racial and ethnic disparities in traffic enforcement with large-scale telematics data.

Past studies have found that racial and ethnic minorities are more likely than White drivers to be pulled over by the police for alleged traffic infractions, including a combination of speeding and equipment violations. It has been difficult, though, to measure the extent to which these disparities stem from discriminatory enforcement rather than from differences in offense rates. Here, in the context of speeding enforcement, we address this challenge by leveraging a novel source of telematics data, which include second-by-second driving speed for hundreds of thousands of individuals in 10 major cities across the United States. We find that time spent speeding is approximately uncorrelated with neighborhood demographics, yet, in several cities, officers focused speeding enforcement in small, demographically nonrepresentative areas. In some cities, speeding enforcement was concentrated in predominantly non-White neighborhoods, while, in others, enforcement was concentrated in predominately White neighborhoods. Averaging across the 10 cities we examined, and adjusting for observed speeding behavior, we find that speeding enforcement was moderately more concentrated in non-White neighborhoods. Our results show that current enforcement practices can lead to inequities across race and ethnicity.

A large-scale analysis of racial disparities in police stops across the United States.

We assessed racial disparities in policing in the United States by compiling and analysing a dataset detailing nearly 100 million traffic stops conducted across the country. We found that black drivers were less likely to be stopped after sunset, when a 'veil of darkness' masks one's race, suggesting bias in stop decisions. Furthermore, by examining the rate at which stopped drivers were searched and the likelihood that searches turned up contraband, we found evidence that the bar for searching black and Hispanic drivers was lower than that for searching white drivers. Finally, we found that legalization of recreational marijuana reduced the number of searches of white, black and Hispanic drivers-but the bar for searching black and Hispanic drivers was still lower than that for white drivers post-legalization. Our results indicate that police stops and search decisions suffer from persistent racial bias and point to the value of policy interventions to mitigate these disparities.

Racial disparities in automated speech recognition.

Automated speech recognition (ASR) systems, which use sophisticated machine-learning algorithms to convert spoken language to text, have become increasingly widespread, powering popular virtual assistants, facilitating automated closed captioning, and enabling digital dictation platforms for health care. Over the last several years, the quality of these systems has dramatically improved, due both to advances in deep learning and to the collection of large-scale datasets used to train the systems. There is concern, however, that these tools do not work equally well for all subgroups of the population. Here, we examine the ability of five state-of-the-art ASR systems-developed by Amazon, Apple, Google, IBM, and Microsoft-to transcribe structured interviews conducted with 42 white speakers and 73 black speakers. In total, this corpus spans five US cities and consists of 19.8 h of audio matched on the age and gender of the speaker. We found that all five ASR systems exhibited substantial racial disparities, with an average word error rate (WER) of 0.35 for black speakers compared with 0.19 for white speakers. We trace these disparities to the underlying acoustic models used by the ASR systems as the race gap was equally large on a subset of identical phrases spoken by black and white individuals in our corpus. We conclude by proposing strategies-such as using more diverse training datasets that include African American Vernacular English-to reduce these performance differences and ensure speech recognition technology is inclusive.

The limits of human predictions of recidivism.

Dressel and Farid recently found that laypeople were as accurate as statistical algorithms in predicting whether a defendant would reoffend, casting doubt on the value of risk assessment tools in the criminal justice system. We report the results of a replication and extension of Dressel and Farid's experiment. Under conditions similar to the original study, we found nearly identical results, with humans and algorithms performing comparably. However, algorithms beat humans in the three other datasets we examined. The performance gap between humans and algorithms was particularly pronounced when, in a departure from the original study, participants were not provided with immediate feedback on the accuracy of their responses. Algorithms also outperformed humans when the information provided for predictions included an enriched (versus restricted) set of risk factors. These results suggest that algorithms can outperform human predictions of recidivism in ecologically valid settings.

Predicting consumer behavior with Web search.

Recent work has demonstrated that Web search volume can "predict the present," meaning that it can be used to accurately track outcomes such as unemployment levels, auto and home sales, and disease prevalence in near real time. Here we show that what consumers are searching for online can also predict their collective future behavior days or even weeks in advance. Specifically we use search query volume to forecast the opening weekend box-office revenue for feature films, first-month sales of video games, and the rank of songs on the Billboard Hot 100 chart, finding in all cases that search counts are highly predictive of future outcomes. We also find that search counts generally boost the performance of baseline models fit on other publicly available data, where the boost varies from modest to dramatic, depending on the application in question. Finally, we reexamine previous work on tracking flu trends and show that, perhaps surprisingly, the utility of search data relative to a simple autoregressive model is modest. We conclude that in the absence of other data sources, or where small improvements in predictive performance are material, search queries provide a useful guide to the near future.

Real and perceived attitude agreement in social networks.

It is often asserted that friends and acquaintances have more similar beliefs and attitudes than do strangers; yet empirical studies disagree over exactly how much diversity of opinion exists within local social networks and, relatedly, how much awareness individuals have of their neighbors' views. This article reports results from a network survey, conducted on the Facebook social networking platform, in which participants were asked about their own political attitudes, as well as their beliefs about their friends' attitudes. Although considerable attitude similarity exists among friends, the results show that friends disagree more than they think they do. In particular, friends are typically unaware of their disagreements, even when they say they discuss the topic, suggesting that discussion is not the primary means by which friends infer each other's views on particular issues. Rather, it appears that respondents infer opinions in part by relying on stereotypes of their friends and in part by projecting their own views. The resulting gap between real and perceived agreement may have implications for the dynamics of political polarization and theories of social influence in general.

Assessing respondent-driven sampling.

Respondent-driven sampling (RDS) is a network-based technique for estimating traits in hard-to-reach populations, for example, the prevalence of HIV among drug injectors. In recent years RDS has been used in more than 120 studies in more than 20 countries and by leading public health organizations, including the Centers for Disease Control and Prevention in the United States. Despite the widespread use and growing popularity of RDS, there has been little empirical validation of the methodology. Here we investigate the performance of RDS by simulating sampling from 85 known, network populations. Across a variety of traits we find that RDS is substantially less accurate than generally acknowledged and that reported RDS confidence intervals are misleadingly narrow. Moreover, because we model a best-case scenario in which the theoretical RDS sampling assumptions hold exactly, it is unlikely that RDS performs any better in practice than in our simulations. Notably, the poor performance of RDS is driven not by the bias but by the high variance of estimates, a possibility that had been largely overlooked in the RDS literature. Given the consistency of our results across networks and our generous sampling conditions, we conclude that RDS as currently practiced may not be suitable for key aspects of public health surveillance where it is now extensively applied.

Respondent-driven sampling as Markov chain Monte Carlo.

Respondent-driven sampling (RDS) is a recently introduced, and now widely used, technique for estimating disease prevalence in hidden populations. RDS data are collected through a snowball mechanism, in which current sample members recruit future sample members. In this paper we present RDS as Markov chain Monte Carlo importance sampling, and we examine the effects of community structure and the recruitment procedure on the variance of RDS estimates. Past work has assumed that the variance of RDS estimates is primarily affected by segregation between healthy and infected individuals. We examine an illustrative model to show that this is not necessarily the case, and that bottlenecks anywhere in the networks can substantially affect estimates. We also show that variance is inflated by a common design feature in which the sample members are encouraged to recruit multiple future sample members. The paper concludes with suggestions for implementing and evaluating RDS studies.

Simultaneous corticosteroid avoidance and calcineurin inhibitor minimization in renal transplantation.

Steroids and calcineurin inhibitors (CNI) have been mainstays of immunosuppression but both have numerous side effects that are associated with substantial morbidity and mortality. This study was carried out to determine if steroids can be eliminated with early discontinuation of cyclosporine A (CsA) and later discontinuation of mycophenolate mofetil (MMF). Ninety-six patients with kidney transplants were entered into four subgroups of two pilot studies. All patients received Thymoglobulin induction, rapamycin (RAPA), and the immunonutrients arginine and an oil containing omega-3 fatty acids. Mycophenolate mofetil was started in standard doses and discontinued by 2 years. CsA was given in reduced doses for either 4, 6, or 12 months. Follow-up was 12-36 months. Thirteen first rejection episodes occurred during the first year (14%). Combining all patients, 86% were rejection-free at 1 year, 80% at 2 years and 79% at 3 years. No kidney has been lost to acute rejection. Ninety percent of the 84 patients at risk at the end of the study were steroid-free and 87% were off CNI. Fifty-seven percent of 54 patients with a functioning kidney at 3 years were receiving monotherapy with RAPA. We conclude that this therapeutic strategy is worthy of a prospective multi-center clinical trial.

The influence of immunomodulatory diets on transplant success and complications.

BACKGROUND: Animal studies have shown that dietary supplementation with arginine and lipids containing the omega-3 and omega-9 fatty acids prolong allograft survival in animals receiving a short course of low-dose cyclosporine. They also reduce cardiovascular complications and infections in humans. METHODS: Adult renal transplant patients receiving standard immunosuppression were stratified according to gender, diabetic state, donor source (LD or CD), and first versus repeat transplant, and randomized to receive or not receive supplemental arginine and canola oil (containing both omega-3 and omega-9 fatty acids) twice daily. Patients were followed for a minimum of 3 years. RESULTS: Seventy-six patients were randomized to the supplement group (S) and 71 patients to the control group (C). Intent-to-treat analysis revealed that S patients had fewer post-30 day first rejection episodes (5.4%) when compared with the C group (23.7%) (P=0.01) and fewer post-30 day episodes of calcineurin inhibitor (CNI) drug toxicity (9.2% vs. 35.3%, P=0.003). S patients developed new onset diabetes mellitus (NODM) less frequently by 3 years (2.3% vs. 14.5%, P=0.04), had fewer cardiac events (5.0% vs. 17.1%, P=0.05), and fewer episodes of sepsis (6.5% vs. 18.7%, P=0.05). CONCLUSIONS: Dietary supplementation with L-arginine and canola oil is a safe, inexpensive, and unique treatment, which is associated with decreased rejection rates and CNI toxicity after the first month in renal transplant patients. Due to reductions in NODM and cardiac events, long-term benefits for patient survival may be particularly important.

Gastric bypass in morbidly obese patients with chronic renal failure and kidney transplant.

BACKGROUND: Morbid obesity occurs frequently in patients with renal failure and is associated with an increased mortality, particularly from cardiovascular disease, as well as a marked increase in comorbid conditions affecting quality of life. Morbid obesity is also associated with an increased risk of complications and death in transplant patients and is often a cause for denial for access to transplantation. METHODS: Thirty morbidly obese patients with chronic renal failure or transplantation underwent gastric bypass (GBP). Nineteen patients had chronic renal failure at the time of GBP, eight had transplantation followed by GBP, and three had GBP and then transplantation. RESULTS: The reduction in excess body mass index (above 25) after GBP at 1, 2, and 3 years was similar to patients without transplantation or chronic renal failure, approximately 70% at 1 year. Comorbid conditions were diminished in each subset of patients, decreasing their risk for potential cardiovascular complications. One patient died 7.9 years after a GBP and 6.1 years after transplantation from cardiovascular disease related to longstanding diabetes that was present before her renal failure. CONCLUSIONS: GBP is a safe and effective means for achieving significant long-term weight loss and relief of comorbid conditions in patients with renal failure on dialysis, in preparation for transplantation, or after transplantation.