What’s the latest thinking in fiscal and monetary policy? The Hutchins Roundup keeps you informed of the latest research, charts, and speeches. Want to receive the Hutchins Roundup as an email? Sign up here to get it in your inbox every Thursday. 

Expansions in pass-through business activity have lowered the labor share of income  

The labor share of income in the U.S. corporate sector has been steadily declining in recent decades. Using administrative tax data, Matthew Smith of the U.S. Treasury and co-authors estimate that about a third of this decline has been driven by the rise in pass-through business activity over the 1978-2017 period. Pass-through businesses have tax incentives to classify owners’ labor income as profits or to form partnerships (for financial, legal, and consulting services, etc.) that move labor-intensive activities away from the corporate sector. The authors estimate that had the shift towards pass-through businesses since the 1980s not occurred, the labor share of income in the corporate sector would have been 1.6 percentage points larger in 2017. The authors’ adjustments are concentrated among mid-market firms in services, suggesting that the remaining decline in the labor share is driven by the manufacturing sector and by the rising importance of superstar firms, they conclude.  

Highly concentrated Bitcoin market susceptible to systemic risk  

By combining data from blockchain—the technology used to record transactions of Bitcoin and other cryptocurrencies—with information scraped from blogs and websites, Igor Makarov of the London School of Economics and Antoinette Schoar of MIT construct a dataset of real entities holding Bitcoin and use algorithms to analyze the behavior of Bitcoin market participants. They find that 90% of transactions are not related to substantial economic activity, and 75% of the transactions since 2015 involve buying and selling Bitcoin. In addition, the Bitcoin market is highly concentrated, with the top 10% of bitcoin miners controlling 90% of Bitcoin mining capacity and the top 50 miners (0.1%) controlling close to 50%. In addition, between 60% and 80% of mining capacity is in China. Finally, the top 1,000 investors own about one-fifth of all bitcoins, while the top 10,000 investors own about one-third. Overall, almost half of all bitcoins are owned by individual investors rather than banks and other intermediaries. Such a highly concentrated market, the authors conclude, makes the Bitcoin ecosystem susceptible to systemic risk and wider acceptance of Bitcoin will likely only benefit a small group of people.  

Labor force exits during the pandemic were larger for Black women, Latinas, and women living with children  

Examining differences in the labor force participation of women aged 25 to 54 in the U.S. during the COVID-19 pandemic, Katherine Lim and Mike Zabek from the Federal Reserve Board find that labor force participation of Black and Latina women declined by over 4 percentage points during the first few months of the pandemic. In the fall and winter of 2020, Black women’s labor force participation increased but participation among Latina women remained persistently lower. In addition, the authors show that women with young children were more likely to exit the labor force than women without children, reflecting the unique demands of raising children during the pandemic. Among women with primary-school-age children in the household, excess pandemic-era labor force exits were concentrated among lower-earning women. Moreover, the authors find that the presence of children explains roughly one-fourth of the labor force exits in excess of the pre-pandemic trend for Black and Latina women. The authors conclude that “the coincident increase in labor force exits among women living with children suggests that regular, reliable, and available childcare plays an important role in supporting women’s labor force participation”.

Chart of the week: Bond market investors expect elevated inflation rates to persist 

Quote of the week:  

“The fundamental dilemma that we face at the Fed right now is this: Demand, augmented by unprecedented fiscal stimulus, has been outstripping a temporarily disrupted supply, leading to high inflation. But the fundamental productive capacity of our economy as it existed just before COVID—and, thus, the ability to satisfy that demand without inflation—remains largely as it was, and the factors that are disrupting it appear to be transitory. Looked at purely in that light, constraining demand now, to bring it into line with a transiently interrupted supply, would be premature. Given the lags with which monetary policy acts, we could easily find that demand is damping just as supply is increasing, leading us to undershoot our inflation target—and, in the worst case, we could depress the incentives for supply to return, leading to an extended period of sluggish activity and unnecessarily low employment,” says Randal Quarles, Member, Federal Reserve Board of Governors. 

“I am among those who see a good chance that inflation will remain above 2% next year, but I am not quite ready to conclude that this ‘transitory’ period is already ‘too long.’ We haven’t yet met the more stringent tests for liftoff that we have laid out in forward guidance about the federal funds rate … Importantly, the level of uncertainty around the paths for inflation and employment are higher than normal as we navigate the unprecedented reopening of the world economy. Therefore, we will remain outcome based, waiting to see further improvements in employment and the evolution of inflation pressures in coming months. And, if the broadly held expectation that inflation will recede next year turns out to be wrong or if inflation expectations show signs of becoming unanchored to the upside, I am confident that the monetary policy tools at our disposal can bring inflation down toward our 2% goal.” 

The Brookings Institution is financed through the support of a diverse array of foundations, corporations, governments, individuals, as well as an endowment. A list of donors can be found in our annual reports published online here. The findings, interpretations, and conclusions in this report are solely those of its author(s) and are not influenced by any donation. 

      

Randal Quarles, the first Trump appointment to the Federal Reserve’s Board of Governors, finished his four-year term as the Vice Chair for Supervision on October 13, 2021. To replace him, President Biden has nominated no one. The Fed replaced him with no one. For now, the Fed’s vital supervisory and regulatory priorities will be managed by the Fed’s Board of Governors, through their committee structure.

There is much to lament with this state of affairs. Quarles was the first to hold the position: it was created in 2010 in the Dodd-Frank Act to encourage the Fed to focus more completely on the vital work of bank regulation and supervision, areas that many feared had become neglected during the Greenspan years. Even though the position was created under a signature law of the Obama administration, that administration did not prioritize the formal appointment, relying instead on Fed Governor Tarullo to manage the portfolio, just as former Fed Governors had done. Today, for reasons known only to the administration itself, if known at all, the Biden administration has been plagued by delays in filling Fed and other financial regulatory vacancies. Even though the Vice Chair’s term is fixed by statute at four years, we still have no insight into the people the administration is even considering to succeed Quarles, as the administration has not even announced an intent to nominate anyone to any position at the Fed.

Quarles, a Republican, pursued a bank regulatory and supervisory agenda with expertise and a clear vision. He is no favorite of some Democrats, who do not endorse his vision, have little use for his expertise, and have been eager to see him depart the scene. Whether the Democrats would prefer it otherwise or not, Quarles is not going anywhere for now. He remains a Fed Governor, with the same important responsibilities over regulatory, supervisory, and monetary policy as his colleagues on the Board. That term is fixed for fourteen years and will not expire until 2032.

Here is the good news. Despite the mishandling of these vacancies from the Biden administration, the Fed’s decision not to reassign these priorities to another Governor is exactly the right thing to do. Its other alternatives are not attractive. It could have given now-Governor Quarles the responsibilities despite the expired term, but his ability to operate without the benefit of his statutory status would be significantly curtailed. The other option is hardly better: the Fed could have given these responsibilities to a candidate more in line with Democratic priorities—Fed Governor Lael Brainard, an expert on virtually every regulatory and supervisory question before the Fed, would fit this bill nicely. But Governor Brainard herself is a candidate to succeed Fed Chair Jay Powell, whose term as Chairman expires in January, and any move to reassign her portfolio could look like meddling in the Fed Chair sweepstakes that is still ongoing.

And so, the Vice Chair for Supervision—that unique creature of governance created by Congress just a decade ago—remains vacant, creating the possibility that financial regulation and supervision will not take their place at the forefront of the Fed’s policymaking. What’s more, the replacement of the Vice Chair position with a committee will devolve more authority to the Fed’s staff to handle this highly political and politicized portfolio.

So why is this good news? Because public oversight of the Federal Reserve System is primarily a product of public governance. We need, as a public, to have rigorous debates about who we want our central bankers to be. One such debate is underway as the Biden administration continues to consider the president’s appointment of the Fed Chair. Those who support Jay Powell, the incumbent, praise his leadership during the 2020 pandemic crisis and his management of a major shift in monetary policy regime. His detractors argue that his regulatory priorities are insufficiently aligned with those of the president, especially around bank regulation, financial stability, and climate change. While the tone of this debate can veer toward hyperbole—an American political tradition as old as the Republic—this is what politics looks like. We should welcome it.

What we are not having, however, is that same level of debate around the priorities that the Fed should pursue as a regulator and supervisor. For this debate, we need to have time to consider viable candidates for this position. And we need the Fed not to do this work for us by pretending that the work of bank regulation and supervision has no political content in it.

The position obviously does have political content. The act of regulating and supervising the financial system is almost top to bottom a political exercise. We have elections to let that content and those exercises dictate the course that regulation and supervision should take. Just because the Biden administration has inexplicably dodged its responsibility for sponsoring that debate does not mean that the Fed should skip the debate entirely. By failing to appoint a successor to Quarles, the Fed has turned up the heat on the politicians to give us—the people and institutions affected most by the Fed’s regulatory and supervisory work—the chance to perform our role in vetting the nominees for this job.

Let’s hope the president accepts the Fed’s invitation as quickly as possible.

The Brookings Institution is financed through the support of a diverse array of foundations, corporations, governments, individuals, as well as an endowment.  A list of donors can be found in our annual reports published online here. The findings, interpretations, and conclusions in this report are solely those of its author(s) and are not influenced by any donation.

      

This week in Class Notes:

• More generous tax credits improve mental health and reduce smoking and heavy drinking among single mothers, but higher SNAP benefits have the opposite effect.
• Piece-rate workers in Florida increased their productivity following an increase in the state minimum wage.
• The common finding of a drop in the share of couples where the wife earns a little more than the husband is explained in Finland by a high proportion of co-working spouses with identical earnings.
• This week’s top chart shows how the Tulsa Race Massacre had persistent negative impacts on Black Tulsans’ home ownership rates.
• Ross Douthat argues for family policy that is more generous but without neglecting marriage, employment, or households that do not use day care, in this week’s choice opinion.
• Check out our latest piece on the college enrollment and completion gaps among men.
• For your calendar: Virtual events discussing the economics of mobility, gender’s impact on life course events, and systemic racism and housing.

More generous tax credits improve mental health and reduce smoking and heavy drinking among single mothers, but higher SNAP benefits have the opposite effect

A more generous social safety net should alleviate some of the hardships faced by low-income single mothers. But which benefits help most? This can be hard to assess, because people can get assistance from multiple programs. Lucie Schmidt, Lara Shore-Sheppard, and Tara Watson use a unique multi-program calculator, that accounts for program interactions and eligibility requirements, to assess the impact of safety net generosity on mental health and risky behaviors for single mothers. They find benefits overall from a more generous package. But there are sharp differences between different programs, as well as in the timing of effects. In particular, tax credits give the most positive boost: a $1,000 increase reduces severe psychological distress by 22.5% and reduces smoking by 14.2%. These reductions are largely concentrated during the first half of the year when tax refunds are likely to occur. However, a similar bump in SNAP benefits increases the chances of smoking by almost 8%, and heavy drinking by almost 10%.

Piece-rate workers in Florida increased their productivity following an increase in the state minimum wage

On January 1, 2009, the state of Florida increased its minimum wage from $6.79 to $7.21. Hyejin Ku studies changes in worker effort in response to this raising of the wage floor using data from piece-rate workers on a Florida tomato farm. First, she notes that low-productivity workers have a greater incentive to increase effort than high-productivity workers. Consistent with this expectation, she finds a 4.6% increase in worker productivity among low-productive workers relative to high-productive workers. In addition, she finds that the minimum wage increase does not appear to lead to dis-employment among the lower-productivity workers. This increased worker productivity could offset about half of the rise in labor costs associated with the rise in minimum wage, she estimates, suggesting that higher costs are shared evenly between the employer and the workers.

The common finding of a drop in the share of couples where the wife earns a little more than the husband is explained in Finland by a high proportion of co-working spouses with identical earnings 

Past studies have found a discontinuity and decline in the share of married households when the wife earns more than her husband (greater than a threshold of 0.5). One primary explanation for this has been a gender identity norm, where women feel pressure to earn less than their husbands, and therefore reduce their labor supply when they out-earn their husbands. However, Natalia Zinovyeva and Maryna Tverdostup argue that this discontinuity actually results from a high proportion of couples who have identical earnings, resulting in a sharp drop off in the fraction of couples past the 0.5 threshold. Using data from an employer-employee matched dataset from Finland, the authors find that 15% of co-working spouses are either both self-employed or work at the same firm. For the self-employed couples, reporting equal earnings could reduce income tax liabilities, facilitate easier accounting, or avoid within-family negotiations. Finally, the authors find that, through co-working, female earnings increase above their earnings potential, which is something the gender norm hypothesis would not predict.

Top chart: The Tulsa Race Massacre had persistent negative impacts on Black Tulsans’ homeownership rates  

This week’s top chart shows the persistent negative impacts of the Tulsa Race Massacre on Black Tulsans’ homeownership. The researchers estimate that the massacre reduced the likelihood of homeownership among male household heads by 4.2 percentage points in the short-term. The direct effects are larger in 1940 and appear to remain constant until 1980, after which the adverse effect grows even more over time into the 2000s.

Chart source: National Bureau of Economic Research

Choice Opinion: The best family policy package is more generous without neglecting marriage, employment, or households that do not use day care   

“The best family policy deal would give progressives more of the money they want to spend, and give conservative ideas more influence over the way that money is spent…[One] possible compromise would attach a work requirement to the credit for parents with kids older than one, while offering the money free of strings to parents of infants. That would encourage single mothers, especially, to return to the workforce as their babies get older, without forcing them back at a moment when they’re particularly vulnerable and when children have the right to a parent in the home. [What you get is] a family policy that spends generously without disfavoring marriage, work or households that don’t use day care. In a better world, that’s what Republicans and Democrats would be negotiating together, but even in this one, it isn’t out of reach,” writes Ross Douthat.

Self-promotion: Younger men are less likely to enroll and complete schooling at every single level of the educational system  

Ember Smith and Richard V. Reeves explore the gender gaps in college enrollment and completion. Women are much more likely to enroll in college and now account for 59% of enrolled students. But there is also a completion gap among those who do enroll. Men who enrolled in a four-year college in 2013 were 10 percentage points less likely than women to graduate within four years. Six years after enrolling, the gender gap in graduation narrows to six percentage points. This leads to large gaps in the levels of education in the adult population. Among those born between 1985 and 1994, for example, men are eight percentage points less likely than women to have a BA or higher degree. We conclude: “Closing the gender gap in education will require interventions every step of the way…Some initiatives specifically focused on male students should be on the table.”

For your calendar: Virtual events discussing the economics of modern transportation, all things urban economics, and happiness and age during COVID-19

Economics of Mobility Meeting

NBER

October 22, 2021 11:30 AM – 5:30 PM EDT

Gender and the Economy Conference

Board of Governors of the Federal Reserve System

November 8, 2021

10th Annual Housing Conference

AEI

November 17, 2021

      

Millions of American taxpayers itemize their deductions, one of which is for state and local taxes, or the SALT deduction. Most of these filers are at the upper end of the income distribution and live in high-income urban areas. On this episode, Senior Fellow Richard Reeves, director of the Future of the Middle Class Initiative at Brookings, says the SALT deduction mostly benefits the wealthiest taxpayers, gives little or no benefit to the middle class, and should be eliminated entirely. He also talks about the unusual politics of the debate in Washington, where Democratic leaders are calling for repeal of the SALT deduction CAP put in place in the 2017 tax law, championed by congressional Republicans.

Also on this episode, David Wessel, senior fellow and director of the Hutchins Center on Fiscal and Monetary Policy at Brookings, explains why inflation is back, why it’s different this time, and what the Federal Reserve can do about it. Listen to this segment also on SoundCloud.

Follow Brookings podcasts here or on iTunes, send feedback email to BCP@Brookings.edu, and follow us and tweet us at @policypodcasts on Twitter.

The Brookings Cafeteria is part of the Brookings Podcast Network.

      

On December 13, Anton Korinek, the David M. Rubenstein fellow at the Center on Regulation and Markets at Brookings, will sit down with Acemoglu to discuss the complexities of the impact of new AI technologies and how AI could be regulated to achieve its promising and wide-reaching potential.

This event will be part of the Brookings Center on Regulation and Markets’ series, “The economics and regulation of artificial intelligence and emerging technologies,” which focuses on analyzing how AI and other emerging technologies impact the economy, markets, and society, and how they can be regulated most effectively.

Viewers can submit questions for speakers by emailing events@brookings.edu or via Twitter using #FutureofAI.

      

What’s the latest thinking in fiscal and monetary policy? The Hutchins Roundup keeps you informed of the latest research, charts, and speeches. Want to receive the Hutchins Roundup as an email? Sign up here to get it in your inbox every Thursday.

Much of the wealth held by the richest households is accrued from business ownership

Using administrative tax data over the 2001-2016 period, Matthew Smith of the U.S. Treasury, Eric Zwick of the University of Chicago, and Owen Zidar of Princeton estimate that the top 1% of households hold as much wealth as the bottom 90%. The share of wealth held by the top 0.1% increased from 12.9% in 2001 to 15% in 2016. Households in the top 0.1% hold over half of their wealth in public equity and pass-through businesses, the authors find. In contrast, pensions and home equity make up the majority of wealth accrued by the bottom 90%. The wealthiest households also have greater risk exposure and earn higher than average rates of return on their investments, the authors estimate. Taking this heterogeneity in returns into account results in lower estimates for the share of wealth held by the top 0.1%. “Given prominent wealth tax proposals focus on the extreme tail of the wealth distribution, our estimates would reduce mechanical wealth tax revenue estimates,” the authors conclude.

Simplicity of post-FOMC press conference language gives investors additional insight

In 2011, the Federal Reserve chair began holding press conferences after Federal Open Market Committee (FOMC) meetings. Michiel De Pooter of the Federal Reserve Board finds that press conferences generate financial market reactions, suggesting that they give investors information beyond that in the post-meeting statement and Summary of Economic Projections. On average, interest rates fluctuate between 1 and 3 basis points, the S&P 500 Index changes by ½ percent, and the dollar moves around 1/4 percent during the period beginning 10 minutes before and ending 20 minutes after the press conference. The authors find that the language of the press conference is more easily understood than the FOMC statement, and also find that Chair Jay Powell’s language in answering questions has been notably less complex than that of his two predecessors. While Powell’s answers are at an eighth-grade level, Ben Bernanke and Janet Yellen’s were at the college level, according to a commonly used readability standard.

Declining interest rates disproportionately benefit “superstar” firms

Using firm-level data from 1962-2019, Thomas Kroen, Ernest Liu, and Atif Mian from Princeton and Amir Sufi from the University of Chicago find that falling interest rates disproportionately benefit “superstar” firms—those with valuations in the top 5 percent for their industry. In particular, a decline in interest rates disproportionately lowers the cost of borrowing for superstar firms, perhaps because conditions that lead to ultra-low rates tend to make other firms relatively riskier. In turn, superstar firms raise additional debt financing, repurchase shares, boost capital investment, and conduct acquisitions more aggressively than other firms. All these effects increase as the level of the interest rate approaches zero. These results suggest that the low interest rates observed in recent years may partly explain the rise of superstar firms and increase in market concentration in the U.S.

Chart of the week: Consumer price of natural gas rising quickly as winter approaches

Source: The Wall Street Journal

Quote of the week:

“I continue to believe currently elevated inflation is episodic, driven by pandemic conditions such as disruptions in supply chains and labor markets. A major caveat, though, is that the severe and pervasive supply chain issues will probably last longer than most of us initially expected. Up to now, indicators do not suggest that long-run inflation expectations are dangerously untethered. But the episodic pressures could grind on long enough to unanchor expectations,” says Rafael Bostic, President of the Atlanta Federal Reserve Bank.

“In my view, then, the fate… of price stability could be on the line in coming months. I think inflation is likely to remain above 2 percent going forward. How far forward I cannot say. But upside risks are salient. I believe the conditions I’ve described argue for a removal of the Committee’s emergency monetary policy stance, starting with the reduction of monthly asset purchases, as we discussed in last month’s meeting.”

The Brookings Institution is financed through the support of a diverse array of foundations, corporations, governments, individuals, as well as an endowment. A list of donors can be found in our annual reports published online here. The findings, interpretations, and conclusions in this report are solely those of its author(s) and are not influenced by any donation.

      

Over the past decade there have been rapid increases in the performance of artificial intelligence (AI). For example, since 2015, AI has performed better than humans at image recognition, as measured in an annual image recognition contest called ImageNet. These achievements have led to a similarly rapid increase in corporate investment in AI and funding for venture-backed AI startups in a range of applications. According to the 2021 “AI Index Report,” of the categories in which it tracks investment, the use of AI in drugs and drug discovery received $13.8 billion in private investment in 2020, 4.5 times higher than 2019. There have been similarly dramatic advances in the capabilities of and interest in other emerging technologies, including applications of blockchain (such as cryptocurrency and non-fungible tokens) and recombinant DNA (perhaps best exemplified by the rapid development of COVID-19 vaccines).

Such rapid technological advances are likely to herald an unprecedented period of innovation and technological change which will fundamentally alter current industries and markets. What is different from previous periods of technological progress is the speed at which these developments are happening and the extent to which they will shape the operations of markets around the world. These technological breakthroughs garner excitement as well as trepidation. On the one hand, new technologies may lead to rapid economic growth and improvements in human well-being. On the other hand, they may affect workers, customers, and society in unanticipated or inequitable ways. Indeed, a common theme in popular press articles is that AI and robots will take jobs away from humans.

Policymakers around the world have started to grapple with these competing claims, seeking to implement policies that harness the benefits of AI and other emerging technologies while mitigating any downsides. The 2021 “AI Index Report” reveals that mentions of AI in the U.S. Congressional Record for the 116^th Congress (2019 – 2021) are triple that of the 115^th Congress (2017 – 2019). However, these technological breakthroughs pose a particular challenge for policymakers and regulators: they will have to be much more flexible and adaptable in their approaches to governing these new markets. In some respect, the technology is so far ahead of the regulations that new approaches to policymaking and regulatory decision making are needed.

Drawing on methods and insights from economics, political economy, public policy, business strategy, political science, and law, the Center on Regulation and Markets’ new series on “The Economics and Regulation of Artificial Intelligence and Emerging Technologies” will explore these issues. The series will feature contributions from experts of diverse viewpoints to elucidate underexplored ways in which AI and other emerging technologies affect the economy and society, and whether—and how—to regulate them most effectively. It is our hope that the new series provides a valuable resource for policymakers, regulators, business executives, academics, and the public.

The Brookings Institution is financed through the support of a diverse array of foundations, corporations, governments, individuals, as well as an endowment. A list of donors can be found in our annual reports published online here. The findings, interpretations, and conclusions in this report are solely those of its author(s) and are not influenced by any donation.

      

So far we have traced the evolution of legal standards for sexual assault and sexual harassment, and their current defects and challenges. There is, however, a significant area of our national discussion that is not fully covered by these discussions, because it involves a complex and uneasy mixture of federal law (Title IX, discussed in Chapter 5) and informal tribunals: sexual assault and harassment on college campuses. Because my previous discussions have covered the most salient issues in each area of law, I need not devote a full chapter to this case, nor do I wish such a disproportionate focus to suggest that women who attend college deserve more attention than women who do not. Unequal access to higher education is already a major problem of justice in our society, compounding other disadvantages based on race and class. There is no reason to perpetuate the injustice by paying more attention to the problems of those women who have managed to arrive at a college or university. One of the great strengths of the traditions I have described is the fact that working-class and minority women (for example Cheryl Araujo, Mechelle Vinson, Mary Carr) have been among their salient plaintiffs.

Yet, because the institutional structures are different, the topic of campus assault requires separate treatment, albeit briefly. Nobody knows exactly how large a problem this is, but one recent survey by the Association of American Universities found that around 20 percent of female undergraduates are victims of sexual assault or sexual misconduct at some point during their college life.¹ Other studies have found frequent sexual abuse of males as well, amounting to 6 to 8 percent. Although there are disputes over methodology and definition, there’s no doubt about the severity of the issue. It would appear, however, that attending college does not make a woman more likely to suffer sexual assault.²

Sexual harassment and sexual assault have long included abuses of power between faculty and students, but on the whole, these cases have been understood as workplace abuses of power, and are dealt with under clear public rules, in much the manner of other workplaces. Thus, Chapter 5 has already basically dealt with these cases. In this Interlude I focus on student-student assault and harassment.

The literature on this topic is vast and controversies are heated, in part because the Obama administration guidelines have now been replaced by different guidelines developed by the Department of Education under the aegis of Secretary of Education Betsy DeVos. However, the controversies cross political lines. Thus the group of Harvard Law School professors who protested against the Obama guidelines as unfair to accused men, anticipating the DeVos critique (I’ll describe their intervention as Stage Two below) included some conservatives, but also faculty from the left and even extreme left of the faculty.

I’ll cover the salient issues briefly, without discussing all the ins and outs of all the controversies. Thus the intention of this brief discussion is to indicate, in a general way, how my overall view in this book’s detailed chapters would approach campus cases, rather than to construct a comprehensive argument.³

Alcohol

A large proportion of sexual assaults and alleged sexual assaults occur when one party, or usually both parties, have been drinking heavily. Heavy drinking makes memory gappy and adjudication very difficult. In general campuses need to do much better with alcohol education and treatment. But one recommendation that most college administrators would support is: lower the drinking age. This approach seems counterintuitive, but it is really sensible. Right now, if adults are present where there is under-age drinking (and most students are under twenty-one), they can be charged with contributing to the delinquency of a minor. So they refrain from providing badly needed supervision, including help for students who have passed out. If the drinking age were reduced to eighteen, adults could attend parties and be prepared to give assistance.

Another alcohol-related issue that needs addressing, in both education and adjudication: sex with a person who has passed out or is close to that point is an assault. This is a species of my point about affirmative consent, but it needs to be repeated again and again. The standard, however, is far from clear in application. Many cases before campus tribunals concern the thorny and as yet unresolved question of how impaired a person must be in order not to be capable of decision-making. Since the evidence comes, typically, from two impaired individuals, it is hard for them to remember how impaired they were. Third-party evidence is usually helpful, but is not always available.

Campus Tribunals

There is considerable confusion in the public mind over why campuses do not simply turn accusations over to the police. So it’s important to point out that campuses have membership conditions, usually spelled out in the admissions contract, that go beyond the letter of the law and that need to be enforced by the campus itself. Plagiarism, not attending class, cheating on exams–all of these things are likely to be punished, sometimes with suspension or expulsion, even though they are not crimes. Similarly campuses may adopt sexual requirements that go beyond the law. Some of these are extreme: honor codes at some religious schools penalize all non-marital sexual conduct. I think such restrictions are counterproductive, creating cultures of silence (if a woman discloses that she has been raped, she can be penalized for engaging in sex). But there are also some reasonable requirements, such as affirmative consent, that are not necessarily the law of the land.

Moreover, the criminal justice system takes a long time, and victims need swift justice in order to deal with the trauma and go forward as students.

Finally, if a perpetrator is convicted in the criminal justice system, that record is ruinous for future life and employment. Campus convictions come in degrees, and many involve mandatory counseling and other lesser penalties. For this reason, having the criminal justice system as the only option, would deter reporting and bringing charges, since victims often hesitate before ruining the perpetrator’s life, and yet they seek some measure of recognition. They want the wrong done to them to be acknowledged—both that it happened and that it was wrong—and they want accountability for the perpetrator; but typically they are not seeking maximal revenge. Nor do they want lengthy involvement with the formal criminal justice system.

These are reasons why campus tribunals are not replaceable by the criminal justice system. However, it must also be said that these tribunals often do their job poorly. Faculty and administrators who serve on them are rarely well trained, and they do not always understand the quasi-legal issues with clarity. Procedures are often poorly defined, and the accused, who typically lack legal representation, are at a disadvantage.

Procedural Issues for Tribunals

How, then, can these tribunals be made to work better?

In this section I’ll refer to several key stages in the debate. Stage One was the “Dear Colleague” letter issued by the Obama administration, laying out standards to which all universities must conform to receive federal money.⁴ Stage Two involved a series of objections to these standards, some issued by Betsy DeVos once she became secretary of education,⁵ but similar objections were raised earlier by legal professionals—most famously by a group of twenty-eight Harvard Law School professors, drawn from both the left and the right, in a letter published initially in the Boston Globe but widely reprinted.⁶ Next, in Stage Three, came the new Department of Education draft rule, which, like all administrative rules was subject to “notice and comment,”⁷ and received over 124,000 comments.⁸ Finally, in Stage Four (May 2020), the Department of Education issued its Final Rule, which is now legally binding on all colleges and universities that receive federal money.⁹ I’ll proceed issue by issue.

First, all involved need to get clear about the best burden of proof. This issue has been one of the largest political disputes. Three standards are currently in use in our legal system. The most stringent, used throughout the United States in the criminal justice system, is proof beyond a reasonable doubt. Many countries do not use this standard for criminal trials, but our tradition has judged that convicting an innocent person is more heinous and more to be avoided than letting a guilty person go free. Together with this exacting standard, our criminal justice system gives the accused a constitutional right to the “effective” and cost-free assistance of legal counsel, although great disparities still exist between public defenders provided free of charge and the sort of lawyer that a more affluent defendant typically would engage—not always because of quality, but because public defenders are overworked and usually don’t have enough time to devote to each client. But at least there is cost-free representation. Furthermore, our Constitution’s “confrontation clause” gives accused parties the right to confront witnesses testifying to their guilt. Over time other rights have been inferred from constitutional guarantees, the most famous being the Miranda warnings that must be read to defendants on arrest, warning them of their right to counsel and their right to remain silent. So our system is protective of defendants in multiple ways.

In civil trials, the standard, instead, is “preponderance of the evidence,” which means anything over 50 percent. Obviously this is a much weaker standard. Nor are free lawyers always provided in civil cases (some states do, most don’t). Still, the civil litigation system has firm procedural structures that safeguard the parties—especially a lengthy period of “discovery,” which gives both sides a chance to examine the other side’s evidence. Without such structural safeguards, and without legal counsel assisting the parties, many people feel that the “preponderance” standard is likely to lead to error.

A third intermediate standard is “clear and convincing evidence,” which is used in ways specified by the relevant state laws, often in areas such as paternity and child custody. This standard is typically thought to mean that it is about 75 percent likely that the person did what is alleged.

Before the “Dear Colleague” letter issued by the Obama administration,¹⁰ most universities used “clear and convincing evidence” as the standard in sexual assault tribunals. The Obama administration insisted, instead, on the civil “preponderance of the evidence” standard. The Harvard Law School faculty letter, and DeVos in her own remarks, held that this standard was not protective enough of the accused. So far, it seems that nobody favors the “reasonable doubt” standard, which would be very difficult to apply in the informal and evidentially challenged situation of a tribunal. So the choice is between the other two standards, and in the end the Department of Education’s Final Rule gives every college that choice.

It’s important to be clear that a college tribunal will not take away a defendant’s liberty. That dire consequence is our legal system’s primary reason for choosing reasonable doubt. Courts, however, have repeatedly held that educational opportunities are economic or property interests, not matters of freedom. So it seems that there is nothing at all odd about using either the civil justice standard of preponderance, or the tougher standard of clear and convincing evidence. This is where the debate occurs.

In real life, both sides have merit. Preponderance defenders believe, rightly, that in the typical alcohol-fueled interaction any stronger standard will be very difficult to meet. However, it is also true that education, albeit a property interest, is one of special defining importance in our society. So it’s important to be protective of the accused. And the civil standard is probably a bad idea in a setting that lacks the procedural safeguards that are usually present in civil trials. Clear and convincing makes more sense, I believe; but if a school should opt for preponderance—as I said, the Final Rule ultimately, and rather surprisingly, gives institutions a choice between these two—a careful tribunal would probably think in terms of a kind of preponderance plus, not necessarily convicting someone where the evidence suggests a mere 50.5 percent likelihood of guilt. The 50.5 approach would really not be protective enough of the accused. Many preponderance-based tribunals actually interpret the standards somewhat more strongly. Whatever the standard, members of tribunals need better training about the whole issue of evidence and the burden of proof.

A second issue of great importance is the definition of sexual harassment. The campus process typically runs together the two things our legal system has carefully kept apart—namely sexual assault or abuse, and (workplace) sexual harassment. There is no harm in this combination so long as sub-definitions are clearly drawn. Sexual assault is typically defined as a single act, not a pattern of actions: you only need to rape a woman once to be guilty of rape! Sexual harassment, by contrast, has two forms. If there is a quid pro quo, a single act suffices. But in “hostile environment” harassment, the plaintiff needs to show a pattern of actions that are sufficiently “serious” and “pervasive,” as well as “unwelcome.” One demeaning comment or gross overture will not suffice. This distinction seems correct.

In terms of this legal background, the Dear Colleague letter was far from adequate. It defined sexual harassment as “unwelcome conduct of a sexual nature,” including “unwelcome sexual advances, requests for sexual favors, and other verbal, non-verbal, or physical conduct of a sexual nature.” This meant in practice that one gross or demeaning comment, with no prior evidence of its unwelcomeness, would be actionable. The Department of Education’s Final Rule, by contrast (Stage Four), hews closely to legal standards accepted elsewhere in our legal system. There are three categories of sexual harassment: (1) “any instance of quid pro quo harassment by a school’s employee,” (2) “any unwelcome conduct that a reasonable person would find so severe, pervasive, and objectively offensive that it denies a person equal educational access”; and (3) “any instance of sexual assault as defined in the Clery Act [a federal statute dealing with campus security], dating violence, domestic violence, or stalking, as defined in the Violence Against Women Act.” In other words, a single unannounced act can still be sexual assault or a quid pro quo, but verbal harassment must form a pattern that meets the Supreme Court standard of pervasiveness and severity, as determined from the point of view of a reasonable observer. The Final Rule protects someone who makes a deeply offensive remark without advance notice of its unwelcomeness and who does not persist.

On most grounds the Department of Education’s Final Rule is an advance over the Obama administration’s rule, and also over the Department of Education’s first rule (Stage Three), under DeVos, which did not include dating violence, domestic violence, or stalking. The Final Rule is perhaps too narrow in its requirement that the accuser show that the harassment is not just severe, pervasive, and objectively offensive, but that it also has a deleterious effect on the person’s equal educational access. Campuses are academic organizations, but they are also social organizations. Social harassment does not always affect someone’s ability to study, and why should that need to be shown? Why isn’t the poisoning of the person’s campus social life sufficient? There are other issues that have been raised, but on balance the “notice and comment” process seems to have worked pretty well.

I shall not go into the details of the various discussions of the questioning and confrontation process in the old and new rules. What I want to focus on, instead, is what I consider to be one of the largest problems with campus tribunals, which has not been addressed by any of these rules: the lack of access to free legal counsel for the accused. Most institutions not only do not provide a lawyer for the accused party; they actively discourage the hiring of lawyers. Typically the accused is permitted to have one supporter or advisor, but when the accused asks if this person can be a lawyer, they are usually discouraged. This is wrong. “Advisors” are typically faculty or administrators who have no legal training and who cannot do an energetic job of defending their client’s rights. And it is also wrong to require people to hire their own lawyers. Free legal assistance would go a long way to dispelling the worries of the twenty-eight Harvard Law School faculty members (Stage Two) about the system’s unfairness. Columbia University does provide free legal counsel for the accused, and so, now, does Harvard Law School (though not the rest of Harvard). My own university has recently begun to implement a policy offering free legal counsel to both defendants and plaintiffs. I have not been able to find out how many other institutions do this. And some federal grant money is available to support accused students at state universities. But the linchpin of our justice system is legal representation. Perhaps this requirement could be waived for minor offenses for which the likely penalty is alcohol counseling, for example; but in cases where the accused faces expulsion it should be mandatory, no matter what it costs. Colleges and universities have many doctors, nurses, and psychologists on their payrolls. And they do have a staff of lawyers, only not for this purpose. They should enlarge their legal departments to include lawyers at the service of students, for just this sort of problem.

I’ve said that tribunals are often poorly trained. The best solution to this problem, since membership of tribunals rotates, is mandatory sexual assault and sexual harassment training for all faculty and administrators. Such training is now required in most universities, as it is in most businesses. At the University of Chicago, each administrator and faculty member must complete the course online every year. It is not perfect, but it does supply a uniform level of awareness.

The Title IX Process

A welcome element of experienced professionalism is now supplied by the presence of Title IX offices on campuses. Typically they do face-to-face training as well as online training, though not as often. But they also play a crucial role through a strong norm of mandatory reporting, which is helping to close the information gap. If a student discloses sexual harassment or assault to any faculty member or administrator, that person is required immediately to inform the Title IX coordinator, giving the complainant’s name. The coordinator will then contact the complainant, typically promising her complete confidentiality and anonymity if she requests it. The complainant usually also has decisional autonomy: nothing will be done, and the alleged perpetrator will not be contacted, unless the complainant gives a go-ahead. Meanwhile the coordinator can advise the complainant about how the process works.

Mandatory reporting is controversial. Many have feared that it will discourage disclosures: the minute you open up to someone you trust, the information also goes to someone else you don’t know. But on the whole mandatory reporting seems wise. The Title IX staff, in my experience, behave with restraint and professionalism, protecting confidentiality. Once faculty and administrators have experience with the coordinators, my experience is that they do come to trust them. And faculty (and others) are relieved of a huge burden of dealing with the whole of a traumatized person’s subsequent life and choices. Faculty usually are not equipped to shoulder this burden, however well-intentioned they are.

The letter by the twenty-eight Harvard Law School professors objected to too much centralized power being vested in the Title IX office, in the scheme at first proposed by Harvard Law School in its attempt to institutionalize the Obama administration standards. The main problem they identified was that the Title IX office did both investigation and adjudication. Their letter was surely correct to say that this setup is very unfair and unwise. Harvard Law School quickly heeded their criticism, separating the two functions. The primary function of the Title IX office should be—and by now for the most part is—investigative and advisory. The tribunals themselves typically consist of faculty, and sometimes administrators, and are constituted according to procedures subject to faculty autonomy and faculty governance. They have many defects, but they are not an alien bureaucracy invading the campus, as the Harvard letter had feared.

We now need to address the gaps in the process that remain, particularly in the area of legal representation.

We have all learned a great deal from these somewhat painful debates. And progress has been made. Although in some ways DeVos has been a polarizing figure, the Final Rule adopted by the Department of Education under her aegis, thanks to the notice-and-comment process, is debatable but still arguably fair. It seems distinctly superior both to the draft rule and to the standards articulated by the Obama administration. We now need to address the gaps in the process that remain, particularly in the area of legal representation.

      

In sectors as diverse as health care, criminal justice, and finance, algorithms are increasingly used to help make complex decisions that are otherwise troubled by human biases. Imagine criminal justice decisions made without race as a factor or hiring decisions made without gender preference. The upside of AI is clear: human decisionmakers are far from perfect, and algorithms hold great promise for improving the quality of decisions. But disturbing examples of algorithmic bias have come to light. Our own work has shown, for example, that a widely-used algorithm recommended less health care to Black patients despite greater health needs. In this case, a deeply biased algorithm reached massive scale without anyone catching it—not the makers of the algorithm, not the purchasers, not those affected, and not regulators.¹

Since both human and algorithmic decisionmakers introduce the possibility of bias, removing algorithms entirely isn’t always the best approach. In fact, in some cases, biased algorithms may be easier to fix than biased humans. However, it falls on policymakers to ensure that the algorithms helping make complex decisions are doing so in a just and equitable way. To do so, we believe three key steps are required. First, regulators must define bias practically, with respect to its real-world consequences. Second, once the goalposts are clear, regulators must use them to provide much-needed guidance for industry and to define targets for objective, hard-hitting investigations into biased algorithms. Third, as in other fields, regulators should insist on specific internal accountability structures, documentation protocols, and other preventative measures that can stop bias before it happens.

Rather than prescriptive rules, which would quickly become obsolete in a rapidly evolving field, we believe these practices will both set a foundation for mitigating bias and provide clear protocols for investigating bad actors.

Defining the goalposts

We believe the key reason AI is still mostly unregulated is that regulators don’t currently have a vocabulary to articulate the benefits and harms of algorithms and hold them accountable. When we regulate a toaster oven, we know it should not catch on fire. When we regulate a pharmaceutical, we know the benefits should outweigh any side effects. But what do we measure when we regulate algorithms?

Algorithms provide decisionmakers with information—a forecast, a probability, or some other key unknown—in order to improve the quality of decisions. Building on this simple observation, we propose that regulators should ask two questions to hold AI accountable: first, what is the ideal information that an algorithm should be providing? And second, is it doing so accurately, both overall and for protected groups?

Consider the example we examined in Obermeyer et al (2019) of a biased health algorithm that caused untold harm. We used data from one health system to study the way a large family of algorithms works. Population health management algorithms like this one are used around the world by health systems and insurers alike to decide who gets access to so-called “extra help” programs. These ‘care management’ programs provide additional up-front care to patients with chronic conditions with the aim of reducing flare-ups and complications in the future. Patients avoid health problems, and the health care system saves the money it would have spent on ER visits and hospitalizations—a win-win.

What is the ideal information the algorithm should be providing in this instance? It was supposed to identify patients who were going to get sick tomorrow so hospitals could enroll them in the extra help program today. We call that goal of the algorithm its ideal target. But what was the algorithm actually doing? In fact, it was doing something subtly but importantly different: it was predicting not who was going to get sick but who was going to generate high costs for the health care system. This is the algorithm’s actual target. The wedge between these two is a key driver of bias.

The algorithm made its predictions based on cost because an assumption was made that health care costs are a fitting proxy for health care need. While that seems reasonable, not everyone who needs health care gets health care—which means some patients end up having lower costs than others, even though they need the same care. This wedge—between what the algorithm was supposed to be doing and what it was actually doing—led to Black patients being deprioritized for the program, resulting in unmeasured harm against many patients. The ideal target, the decision we care about, was need for care, but the algorithm’s actual target was cost of care. We wanted the algorithm to answer one question, but it was answering something else.

Algorithms are like genies. A man asks a genie for his wish, “I want to be rich!” The genie replies, “Ok, Rich, what’s your next wish?” Algorithms are concrete and literal to a fault. When we investigate the decisions an algorithm is informing, we can understand the gap between what we want it to do and what we actually told it to do. This gap between the ideal target and the actual target results in what we call label choice bias. In follow-up work since the example above, we’ve found bias in a wide range of other algorithms, nearly all driven by biased proxies, leading to biased outcomes. This bias, disturbing as it is, is just a symptom of a deeper problem: algorithms are often not doing what they are supposed to be doing. To catch problems like these, we need to understand algorithms in context. What do we want them to do? What are they actually doing? Is there a discrepancy, and if so, is it different for different groups?

To summarize, algorithms provide a key piece of information to a decisionmaker. So regulation must ensure that they are providing that information accurately, both overall and for specific groups. This approach, which focuses on the output of the algorithm—the accuracy of its predictions on an ideal target—has several key advantages. It does not require ‘opening the black box’: algorithms can be audited simply based on the scores they produce. This avoids compromising trade secrets. It also means regulators do not need to regulate the inputs of algorithms or understand the many reasons why they might be biased—non-representative training data, use of an explicit race correction, etc. Instead, regulators can focus on one simple question: is the algorithm predicting its ideal target accurately and equitably? This test will detect many forms of bias, like failure of an algorithm to generalize from one population to another. But unlike many other measures of bias proposed, it will also detect label choice bias, which we’ve found is a major driver of algorithmic bias in many settings.

Translating the goalposts into action

Defining the goalposts can be transformative. In our experience, most people who make algorithms do not want to scale up racial biases. They want to design products that work and to stay on the right side of regulators, but because this is a new and fast-moving field, they often don’t know how. By providing guidance to industry on what bias looks like and how to avoid it, regulators can have a transformative effect on the future of algorithms in many fields.

At the same time, regulators also need to be prepared to enforce the standards and make sure the industry is staying within those goalposts that have been set. There are a couple key tools regulators can use to investigate suspected cases of bias.

First, start with some simple statistical tests to see if the algorithm is performing well on the variable it is actually predicting. These tests are as simple as making a graph or running a regression in any standard statistical package: simply compare the variable being predicted to the algorithms score, by group. This should detect problems caused by non-diverse training data, failures to generalize, or poor performance in underserved groups (this is termed “calibration” in the literature). But keep in mind: good performance in one protected group does not mean the algorithm is free of bias.

Second, hold the algorithm accountable for predicting the ideal target for underserved populations. Once you define your ideal target variable, you can check whether the algorithm does a good job of predicting that variable across populations. Again, this is as simple as a graph or a regression comparing the ideal target to the algorithm score, by group. Notice that we don’t need to understand or explain the inner workings of the “black box” in order to find bias. We just need the outputs of that box (the algorithm’s predictions) and some data representing the outcome we care about (the ideal target) for our population.

Often, we find that algorithms are poorly calibrated when you consider the outcome that is truly important. In the health care example above, where an algorithm informs whether people are offered an extra help program, we found that if you take two patients, one Black and one white, who are equally sick (as defined by active number of chronic conditions), the algorithm is more likely to recommend extra care for the white patient than the Black one (see graph below). This is a case where something has certainly gone wrong.

Solving these problems is simply a matter of matching the actual target (what the algorithm is predicting) more closely to the ideal target (what we want the algorithm to predict). Again considering the health care example, instead of predicting cost, we should predict the variable we truly care about—health care needs. In our research, adjusting the target led to almost double the number of Black patients being selected for the program. Fixing algorithmic bias means solving the problems we are really trying to solve.

Accountability structures to prevent algorithmic bias

Policymakers also need to provide guidance on how to set up preventative structures and practices within one’s organization. In our Algorithmic Bias Playbook we recommend a few steps, and we think that most of these are broadly applicable to most organizations.

In any organization, it is crucial that a single person be responsible for algorithmic bias. This person, whom we’ve termed a steward, should have broad oversight over strategic decisions so that they are equipped to provide executive support to the team but also so they can be truly accountable for mitigating bias. While a single person should be ultimately responsible, they shouldn’t work alone. They must be advised by a diverse group of stakeholders for this effort to be truly successful. Individuals from different organizational capacities with different backgrounds, both internal and external voices, should all be included in an advisory group to the steward so that decisions about algorithmic bias are made collectively.

Aside from a clear accountability structure, organizations also need clear documentation. Algorithms live throughout an organization and provide input on decisions that affect thousands or even millions of people. We recommend that the steward and her team create a master inventory of algorithms that can provide a record of items such as the following: the source, the owner, the intended purpose/ideal target, the actual target, the population served, and any other information that the organization deems useful. Keeping records like these will help uncover where bias can arise and help prioritize efforts to ensure equitable outcomes.

Finally, be clear that ignorance is no protection—as our own ongoing work with state law enforcement agencies has taught us. It is no longer reasonable for an organization to be unaware of algorithmic bias as a potential problem. Negligence is in not doing the work to find, mitigate, and prevent it.

Conclusions

Algorithmic bias is everywhere. Our work has shown that label choice bias is a particularly dangerous form of bias: it is widespread because we often don’t have the specific variable we’re interested in. We’ve discussed an example from health care, where it’s difficult to measure the true outcome we care about—health—and instead organizations often fall back on convenient proxies like cost. Any industry where the variables we care about are not truly available will have this issue. In finance, we may want to measure “creditworthiness,” but instead we measure income, employment, demographics, etc. In crime, we may want to know “propensity to commit crime,” but instead we look at arrest record, education, employment, etc. These are all proxies for the true variable of interest, which means there is room in every one of these cases for label choice bias to exist. They must be investigated and prevented.

The good news is that we can reduce bias in algorithms by asking organizations to do a better job defining that “ideal target” —the thing we really care about. Organizations across the health care industry have come to us for support as they’ve worked to identify and mitigate bias in their own contexts. Some problems we’ve helped address include population health management, operational efficiency, strategic patient engagement, and others. In addition to workshopping specific algorithms, we help organizations apply our framework in practice and establish processes for proactively preventing bias.

Regulators are central to solving the problem of algorithmic bias. While some examples like the Federal Trade Commission guidelines and Food and drug administration regulatory framework are good starts, much more needs to be done. Regulators must identify targets for prompt investigations grounded in real world use cases and provide clear guidance and accountability structures for organizations to follow. Working together, we can move the needle towards more equitable outcomes across sectors.

Ziad Obermeyer has received speaking or consulting fees from AcademyHealth, Anthem, Independence Blue Cross, Premier Inc, and The Academy. The authors did not receive financial support from any firm or person for this article or, other than the aforementioned, from any firm or person with a financial or political interest in this article. They are currently not an officer, director, or board member of any organization with an interest in this article.