Confounding Variables from Historical Bias

Note: co-authored with Christine Assaf, originally published in the now-defunct PASS Blog.

Historical data analysis that is naïve to past discrimination is doomed to parrot bias. How do we combat bias in our data analytics? 

First, we already know that complex, multi-talented teams are best-suited to face complex, multi-faceted problems. A more diverse group of researchers and engineers is more likely to recognize and think harder about bias problems that may impact them personally. “It’s difficult to build technology that serves the world without first building a team that reflects the diversity of the world,” wrote Microsoft President Brad Smith in Tools and Weapons (2019). 

Second, data collection is often provided to us through real-world interactions, subject to real-world bias. No real-world data exists in a vacuum. We should understand that bias in data collection and analysis may be inevitable but is not acceptable. It is not your responsibility to end bias (though that’s a worthy cause), but rather to be proactively informed and transparent.  

Look for systemic outcomes, not intentions. Only in outcomes are potential disparate impacts measured. Let’s review a couple of examples. 

Many Americans are familiar with the ACT test, an exam many students take as part of the college application process.  In 2016, ACT admitted an achievement gap in composite scores based on family income. According to ACT’s own data, there is a 3-4 point difference in scores between poorer and wealthier households, and the gap continues to widen. 

Credit to ACT for disclosing their research. Transparency is part of accounting for bias in historical data and data collection and is critically important to furthering a larger conversation about inequality of opportunity. 

Recently, more than 1,000 American institutions of higher learning have adopted test-optional admissions policies, meaning they no longer ask for the ACT (or SAT, a similar exam) on applications. An overwhelming amount of studies have been conducted suggesting that the ACT does NOT predict college graduation outcomes as strongly as other factors, including high school GPA and household income. 

Knowing the variables involved in your analysis is important. When conducting analysis, researchers must review, identify, and anticipate variables. You will never find the variables unless you are looking for them.   

This is why a proposed new rule by the United States Dept of Housing and Urban Development in 2019 stirred a massive reaction from technologists and ethicists alike. The proposed rules, yet to be implemented, would make it nearly impossible for a company to be sued when racial minorities are disproportionately denied housing, as mortgage lenders or landlords could simply blame their algorithm to avoid penalty. 

Centuries of racially-divided housing policies in the United States evolved into legalized housing discrimination known as redlining, ensconced in federal-backed mortgage lending starting in the 1930s. The long history of legal racial housing discrimination in the United States was arguably not directly addressed until the 1977 Community Reinvestment Act. Yet today, 75% of neighborhoods “redlined” on government maps 80 years ago continue to struggle economically, and minority communities continue to be denied housing loans at rates far higher than their white counterparts. If discriminatory outcomes in housing are to change for the better, the algorithmic orchestration of mortgage lending should not be excused from scrutiny.  

In both cases, we examined industries where algorithms draw from data revealing larger societal outcomes. These outcomes are the result of trends of economic inequality and a pervasive opportunity gap. Are such data systems to be trusted as the result of an algorithm, and thereby inherently ethical? No, not without scrutiny.  

These provide examples of when data professionals must be aware of the historical and societal contexts from which our data is drawn, and how the outcomes of our data findings could be leveraged. Would our outcomes contribute to justice?  For example, the industries of financial marketing, healthcare, criminal justice, education, or public contracting have histories checkered with injustice. We should learn that history. 

Transparency is desirable. It is needed to aid an informed societal conversation. We should not that assume an algorithm can overcome historical biases or other latent discrimination in its source data. Informed scrutiny should gaze upon historical data with a brave and honest eye.

Measure Ethical Data Analysis by Outcomes, Not Intent

Note: co-authored with Christine Assaf, originally published in the now-defunct PASS Blog.

In January 2020, Robert Williams was arrested on his front lawn in a suburb of Detroit, Mich., after police scanned state driver’s license photos and matched his face to grainy surveillance camera footage. He was held for 30 hours in police custody on suspicion of theft of 5 luxury watches, months earlier. He was innocent. In the resulting lawsuit and public scrutiny, the Detroit Chief of Police admitted in a public meeting in June 2020 that their facial recognition software would misidentify suspects 95-97% of the time.

A 2019 NIST study found larger error rates for darker-skinned samples across 99 different facial recognition software providers, including those used by the Michigan State Police software package.

In a 2017 MIT study, leading facial recognition software technologies provided by Microsoft, Face++, and IBM were put to the test on front-facing portrait style headshots. Their accuracy was lowest when evaluating darker-skinned faces. For example, in the case of IBM’s Watson, there was a 34.4% difference in error rate between lighter males and darker females, while 93.6% of Azure Face API’s gender identification failures were on darker-skinned subjects.

The reasons for the embarrassing inaccuracy are being pursued, and could be speculated, from the photo contrast levels to an unrepresentative data set used to train the models. However, it is instead the inaccurate outcomes that are the cause for concern. The potential for a civil rights nightmare is clear and present.

Robert Williams, the Michigan man arrested based on a faulty match, was black. In many Western countries including the United States, communities of color are more scrutinized by the criminal justice system. There is identifiable discrimination based on criminal justice outcomes, including ethnic and gender biases in sentencing, especially when judges have discretion.

Data collection and analysis without regard to the potential for disparate outcomes may reinforce and institutionalize a society’s discriminatory history. Appropriately, Microsoft, IBM, and Amazon announced in June that they would no longer license their facial recognition platforms for use by law enforcement.

Ethical outcomes don’t care about your intentions. Systems can only be judged by their impact, not by intentions vulnerable to revisionism and disconnected from outcomes. Algorithms, as we in the data community well understand, are not infallible, incorruptible oracles. We know better.

When we presented on this topic in May 2020 at a virtual conference, we were asked somewhat incredulously if we favored regulation of software development.

Frankly, our answer is yes.

Engineers who make bridges need a regulated licensing structure around a Professional Engineer stamp, because we don’t want bridges collapsing. Doctors need board certifications to make sure people get evidence-based medical care. From the Therac-25 software bug to the Boeing 737 MAX, perhaps we as data professionals specifically, and software developers generally, do need some regulation. Or, at least, ethical commitments to practice.

Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence and a professor at the University of Washington, has proposed just that: a modified version of a medical doctor’s Hippocratic Oath. He proposed a Hippocratic Oath for artificial intelligence practitioners, with this statement at its heart: “I will consider the impact of my work on fairness both in perpetuating historical biases, which is caused by the blind extrapolation from past data to future predictions, and in creating new conditions that increase economic or other inequality.”

There is considerable need for this type of commitment to our quickly evolving use of modern datasets. Colin Allen, a researcher in Cognitive Science and History at the University of Pittsburgh, summarized the need nearly a decade ago: “Just as we can envisage machines with increasing degrees of autonomy from human oversight, we can envisage machines whose controls involve increasing degrees of sensitivity to things that matter ethically. Not perfect machines, to be sure, but better.”