(Originally published by Stanford Human-Centered Artificial Intelligence on January 31, 2023)
A Stanford collaboration with the Department of the Treasury yields the first direct evidence of differences in audit rates by race.
Researchers have long wondered if the IRS uses its audit powers equitably. And now we have learned that it does not.
Black taxpayers receive IRS audit notices at least 2.9 times (and perhaps as much as 4.7 times) more often than non-Black taxpayers, according to a new paper by Daniel E. Ho, the William Benjamin Scott and Luna M. Scott Professor of Law at Stanford Law School, faculty director of the Stanford RegLab, associate director of the Stanford Institute for Human-Centered Artificial Intelligence, and senior fellow at the Stanford Institute for Economic Policy Research; Hadi Elzayn, researcher at the Stanford RegLab; Evelyn Smith, PhD candidate at the University of Michigan; Arun Ramesh, a pre-doctoral fellow at the University of Chicago; Jacob Goldin, a professor of tax law at the University of Chicago; and economists in the U.S. Department of the Treasury’s Office of Tax Analysis.
The disparity is unlikely to be intentional on the part of IRS staff, Ho says. Rather, as the team’s research demonstrated, the racial disparity in audit selection is driven by a set of internal IRS algorithms that Goldin likens to the recipe for Coca-Cola. That is: It’s completely secret.
To better understand this audit selection bias, the research team modeled the racial impact that various alternative audit selection policies might have. The result: a demonstration of how the IRS might be able to tweak its secret algorithm to reduce its racially disparate impact.
“The IRS should drill down to understand and modify its existing audit selection methods to mitigate the disparity we’ve documented,” Ho says. “And we’ve shown they can do that without necessarily sacrificing tax revenue.”
From a Suspected Disparity to a Proven One
Although there have been long-standing questions about whether the IRS uses its audit powers equitably, Ho says, the private nature of tax returns and the confidentiality of the IRS’s approach to audit decisions made it difficult to study. That changed when, on his first day in office, President Biden signed Racial Justice Executive Order 13985 requiring all federal agencies to assess how their programs impact racial and ethnic equity. To apply that order to the IRS tax return audit program, economists at the Treasury Department collaborated with the Stanford RegLab team, allowing them to analyze (on an anonymized basis) more than 148 million tax returns and approximately 780,000 audits for tax year 2014 (an overall audit rate of 0.54%).
Read the full paper, Measuring and Mitigating Racial Disparities in Tax Audits
Even with all that data in hand, the research team faced a major hurdle: Tax returns do not ask for the taxpayer’s racial or ethnic identity. So, the team adapted and improved on a state-of-the-art approach that uses first names, last names, and geography (U.S. Census block groups) to predict the probability that a person identifies as Black. And they validated their racial identification predictions using a sample of voter registration records from North Carolina – a state where, until recently, citizens were required to check a box for race and ethnicity when they registered to vote.
After finding that Black taxpayers were 2.9 to 4.7 times more likely to be audited than non-Black taxpayers, the team explored possible reasons for that disparity. They suspected that the problem lay with an IRS algorithm’s use of the Dependent Database, which flags a potential problem and generates an audit letter to the taxpayer. That instinct proved to be correct in that the bulk of the observed racial disparity involved so-called “correspondence” audits done by mail rather than more complex, in-person “field” audits.
The team also found that the IRS disproportionately audits people who claim the Earned Income Tax Credit (EITC) – a program that assists low- to moderate-income workers. But claiming the EITC only explains a small percentage of the observed racial disparity. The largest source of disparity occurs among EITC claimants. Indeed, Black taxpayers accounted for 21% of EITC claims, but were the focus of 43% of EITC audits.
The racial disparity in audit rates persists regardless of whether EITC claimants are male or female, married or unmarried, raising children or childless. But it is most extreme for single male taxpayers claiming dependents (7.73% for Black claimants; 3.46% for non-Black claimants) and for single male taxpayers who did not claim dependents (5.66% for Black; 2% for non-Black).
Perhaps the most striking statistic is this: A single Black man with dependents who claims the EITC is nearly 20 times as likely to be audited as a non-Black jointly filing (married) taxpayer claiming the EITC.
The Disparate Impact of the IRS Audit Selection Algorithm
Although the team does not know the precise algorithms the IRS uses to select audits, they modeled several possible explanations for the racial disparity in audit rates.
First, they tried an “oracle” approach using a dataset called the National Research Project (NRP) – a set of nearly 72,000 tax returns that is randomly chosen for audit. Because each tax return in this dataset was subjected to a line-by-line audit, the amount of underreported tax liability is known. So the researchers considered what would happen if they pretended that an omniscient IRS selects taxpayers for audit based on the known amount of underreported tax in the NRP dataset. The result: The racial disparity in audit selection flips – an omniscient IRS aimed at capturing the most underreported income tax would audit more non-Black taxpayers than Black taxpayers.
Since auditors are not omniscient, the team also used the NRP dataset to train a model to predict – for the full 2014 dataset – the likelihood that a taxpayer has underreported income, and the magnitude of a taxpayer’s underreporting. They found that an approach focused just on the likelihood that there’s underreporting of at least $100 would result in auditing more Black taxpayers (as was observed). By contrast, a focus on the magnitude of underreporting (the amount of money unpaid by a taxpayer) would yield a result much closer to the oracle: More non-Black taxpayers would be audited than Black.
Why might this be the case? “The choice to focus on whether there is underreporting, as opposed to magnitude of underreporting, is connected to broader structural sources of economic inequality and racial justice,” Smith says. Because far more Black taxpayers have lower income, they have less opportunity to underreport substantial amounts of income. By contrast, Smith says, “focusing audits on the amount of underreported income will disproportionately end up focusing on higher income individuals who are less likely to be Black taxpayers.”
Finally, the team wondered if the racial disparity in audits springs from IRS and even congressional concerns about refundable tax credits including not only the EITC but several others. When someone claims one of these tax credits, which are part of our country’s social safety net, they receive a refund amount even if they didn’t pay any taxes. And some in government believe it’s more important to make sure we avoid paying money to someone who claims it inappropriately than it is to make sure we collect all of the tax dollars that are due from someone engaged in some other type of tax evasion.
To test the hypothesis that this approach would have a disparate impact on Black taxpayers, the team examined what would happen if the IRS focused audits specifically on the underreporting that is due to over-claiming of refundable tax credits (the EITC as well as two others) rather than total underreporting. Their findings: This policy would result in Black taxpayers being audited at the rates similar to what the team observed in the 2014 data.
The Racial Impact of IRS Resource Constraints
Seventy percent of IRS audits happen through the mail and 50% involve EITC claimants. That’s because, compared with labor-intensive field audits, correspondence audits of EITC claimants are easy to trigger, cost very little, and require minimal effort by IRS personnel. Unfortunately, the burden of correspondence audits on EITC claimants is more likely to fall on lower income individuals whose tax returns are less complex and less likely to lead to litigation, as this research team noted in a recent paper about vertical equity in tax audits.
In their new work, the team found that additional aspects of the IRS audit selection process driven by an impetus to reduce costs have a racially disparate impact. For example, even among correspondence audits of EITC claimants, the IRS devotes fewer resources to auditing EITC returns with business income. It’s likely that’s partly because auditing EITC returns with business income would be more expensive (approximately $385 per audit compared to $29 per audit for EITC claimants with no business income), Elzayn says. And, the team found, this cost-saving measure has a disparate impact on Black taxpayers, who make up only 10% of EITC claimants reporting business income but 20% of EITC claimants who don’t report business income.
Yet even if IRS resource limits explain some of the racial disparities the team observed, they don’t explain all of them. “Even holding fixed how many audits are devoted to EITC claimants who report business income, we still observe racial disparities,” Elzayn says.
Next Steps for the IRS
The study’s authors have not made any formal recommendations for how to make the IRS audit selection algorithm more just. Instead, they have documented the likely effects of alternative policies, which provides the IRS with several potential pathways for alleviating the racial impact of its audit selection system. These include predicting and focusing on the magnitude of taxpayers’ underreported income rather than just the likelihood of it; viewing dollars as equal whether they are to be paid in refundable credits or received in taxes; and using IRS resources to audit more complex returns rather than focusing only on the simpler ones that are cheaper to audit.
Before Biden signed the Racial Justice Executive Order that engendered this research project, the IRS had neither the impetus nor the ability to do that. Now that they know the equity implications of how they select audits, Ho hopes they will tweak their confidential audit selection algorithm.
“Racial disparities in income are well known, and what the IRS chooses to focus on has big implications for whether audits complement, or undercut, a progressive tax system,” Ho says.
Stanford HAI’s mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.