Big Data vs. Affirmative Action
Past performance is not an indicator of future results. –Every Prospectus Ever.
So there was this former executive of a successful San Francisco start-up telling us about how she just couldn’t hire a person who had bad credit.
And as you make sense of that statement, you’re very likely to jump to the wrong conclusion about what kind of a person that particular corporate executive was: you’re likely to suspect that she’s a classist elitist jerk because that’s the kind of story that you’re used to hearing that would conclude with “she just couldn’t hire a person who had bad credit.” And this is the same sort of trap we fall in when we encounter structural racism, or any other systemic -ist form of discrimination by algorithm: we look for the bad actor so that we can continue to play by the neutral-or-good rules while reassuring ourselves that we’re nothing like the bad actor. This would be fine if the rules were truly neutral, but they aren’t, and that’s even assuming that the people who wrote them did so in good faith and with the best of intentions.
So this post is for all you kids who are so not-racist that you think Affirmative Action is reverse discrimination that undermines the notion of equality that you’re totally in favor of. I’m going to start by splitting out the bad actors and looking briefly at what we want to believe is neutral to show how even good intentions can lead to -ism or -ist results. Then I’m going to double down on that point to explain the sort of situation the corporate executive I lead with was in, and why it’s a risk to you. And then I’m going to conclude with a value-based justification of affirmative action, regardless of how badly the policy may be implemented.
(Just to be clear, -ism and -ist are the abstract forms of “any kind of -ist — racist, sexist, ageist — behavior, or any -ism — racism, sexism, ageism — et cetera” because I’m making an abstract argument about how discrimination here.)
Let’s start with a basic claim:
People too often think that structural racism works the same way that interpersonal racism does. It doesn't. No malice/awareness necessary.
— Charles M. Blow (@CharlesMBlow) July 19, 2015
We all know what interpersonal racism is: it’s these assholes who caused shit like this half a century ago, but then Rosa Parks and MLK fixed almost everything except for… pretty much everything. But this is what schools want to teach because it sounds like progress has been made and the grown-ups are constantly doing a better job than their predecessors — we call this the Myth of Perpetual Progress and it’s off-topic, but I recommend Loewen’s Lies My Teacher Told Me if you’re surprised.
I’m going to briefly argue that structural -ismist behavior is designed into our most-optimized social systems. And I’m drawing upon a bog of knowledge from Cathy O’Neil, Allstair Croll, and danah boyd in case you want to read more of this ilk (you totally should, that’s why I linked it for you). And I’m going to start with the very simple, aspirational, American-dreamy idea of buying a house. Yes, there are bad actors in Real Estate and everywhere else, too, but I’m going to not delve into them directly so I can focus on non-bad actors. And this is important because when things that aren’t bad are actually still bad, then we can’t excuse the badness by saying that we’ve cherry-picked evidence to make things worse than they are.
Because here’s the thing about the non-bad actors like insurers and creditors: they know what kind of a world they live in. Their first bet is a guess on when you’re going to die. But they’re also keenly interested to know how stable your job is, or if you’re likely to be fired — even if it’s a dodgy firing by an -ist boss that doesn’t know you and doesn’t like your demographic. For mortgages, they’re interested to know where you’re buying a home. And in the ideal free-market model, they’ll lower the cost of service for their low-risk customers by charging their high-risk customers more for less service, commiserate to the risk. The service charges (insurance premiums or loan interest rates) would be higher to recoup money earlier, while service maximum (insurance policy size or mortgage amount) would be capped lower to minimize the maximum amount of loss on payout/default. This is simple: the individual — me, you, any schlub — is specifically paying the institution to take on the fiscal risk of giving some rando — me, you, any schlub — a chunk of change, either to buy some property or in the event of misfortune. The more precisely the institution knows the likelihood that you’ll stop paying them and they’ll lose money on you, the more precisely they can charge you for the favor of taking on your risks to ensure that they make (rather than lose) money.
What this means is that an optimized mortgage equation that knows the elevated likelihood of a black man — his name is Jordan, why not? — being spuriously fired by some asshole boss, and will charge Jordan more and loan him less than it would me. The algorithm isn’t trying to be racist; it’s just acknowledging the ugly likelihood that Jordan’s going to have his life disrupted in a way that will impact the business Jordan is doing with the bank and protecting the bank from it while simultaneously working to earn my business. But this still means that Jordan is paying more for less, a negative effect that makes the bank running the racism-aware algorithm appear “racist” rather than “legitimately concerned about racism.” (We’ll get to what “legitimately concerned” looks like when we’re talking about affirmative action.)
Assuming that Jordan does business with the bank anyway, he’s going to promptly run into two downstream effects. First, he can’t buy as much or as nice of property. For example, he might find that properties near a polluted industrial district, or near railroad tracks, or in an higher-crime neighborhood, or under the low flight path of the airport are the only sorts of properties he can afford — and they bring stresses, carcinogens, and a shorter life span. (There was a clear correlation in southern California as I recall, but sadly I cannot remember the book I was reading that was harping on this evidence.) And that’s bad, but secondly there’s an ongoing economic impact: the American middle class keeps a lot of its wealth in its home real estate, and the lower mortgage cap basically caps Jordan’s ability to develop his home as a bastion of wealth.
If you’re in my target audience of teenagers, this may come as a surprise — but CNBC reports that
“Homeownership plays a pivotal role in the U.S. economy and has historically been one of the primary sources of wealth accumulation for middle-class families,” said Lawrence Yun, chief economist for the Realtors. … “Unfortunately, due to an underperforming labor market, insufficient housing supply and overly stringent underwriting standards since the recession, homeownership has plunged to a rate not seen in over two decades,” Yun added. “As a result, the country has become more unequal as the number of homeowners has fallen while the number of renters has significantly risen.”
So let me break down some sample math for you on the expectation that you’re a teenager and haven’t thought about this at all yet: when I was renting a few years ago, I was spending about $1000 on a mediocre apartment per month — that’s just spent and gone. But now I’ve got a house, and I’m putting $900 per month into the principal (that’s the part of the house I own) and spending $400 on interest and spending $300 on taxes. So clearly my budget is more restricted per month, but both the interest and the taxes are tax deductible, and the principal counts towards my net worth — so at the end of the year, after tax deductions, I’m only spending — in the “spent and gone” sense of the word — $500 per month on a nice house. And while rent has gone up noticeably since I’ve bought this house, my mortgage payments have not even though the market value of the house has increased. Yes, it can be hard to have that principal payment leaning on my budget every month, but when I’m done with this house I should be able to sell it and walk away with a huge pile of cash that I simply wouldn’t have been able to accumulate if I’d been renting.
That’s why Jordan wants to buy a house, wants to buy a house that will nicely increase in value while he owns it, and wants to buy a house that he’ll be able to re-sell when he’s ready to move on: it is an important part of his ability to advance socioeconomically, for both him and his family (if you’ve imagined him having a family), and he needs a nice big loan in order to do it because otherwise he’ll be blowing too much money on rent for too long and not make that socioeconomic advance at, for example, the rate I am.
But look back at Yun’s statement: “overly stringent underwriting standards.” That’s when banks are risk averse to people who may not be able to pay all the money back, for any reason — inclusive of being a victim of -ist stupidities in their place of (former) employment. Is that kind of algorithmic discrimination legal? Probably not. Does it happen anyway? Yes. Very clever statisticians are cooking up new ways to read historical data that won’t seem like racism — because they’re not really racists, right? — or sexism or any other -ism while more-precisely charging customers for the risks they pose to the bank. And that historical data is a sight to behold. Let’s say for a moment that we’re looking at a 30-year fixed interest rate mortgage (this is pretty normal; it will take twice as long as you’ve been alive to pay off a house), so we look back 30 years into the past to see how risks played out — and realize that 30 years ago was 1985 when the crack-cocaine epidemic ravaged inner cities, igniting a racially-charged moral panic of “at-risk” youth. 30 years before that was a decade before civil rights, so employment for minorities was super-precarious. As humans, we want to believe — as we were taught in school — “Wow, so much has changed!” But the algorithm is just going to crunch its numbers as if nothing ever changes.
Let me put that another way: large-scale personal discrimination, even-and-especially in the past informs our algorithms how dangerous it is to be in a non-powerful demographic in a country that is apparently overrun by -ist assholes. And our algorithms, in an as amoral (if not sociopathically dispassionate) way as they possibly can, return their variable response to the known risks in a way that is rarely questioned because we trust in the safe neutrality of the rules because they’re here to protect us from those -ist assholes, right?
You should be legitimately concerned at this point.
But let’s give you an example that’s in your more-immediate future: So there was this former executive of a successful San Francisco start-up telling us about how she just couldn’t hire a person who had bad credit, even though she really wanted to. We’ll say she was trying to hire Jordan (because I don’t know who she was actually trying to hire). See, the problem is that no matter how much money she offers Jordan now, even for moving expenses and as a joining bonus, if Jordan has a shitty credit score — a huge menagerie of college loans, maybe a couple of missed credit card payments because OMG Textbooks Are Expensive, maybe a used car loan on a lemon — Jordan will be rejected when he tries to rent an apartment, any apartment, and will thus be unable to move to San Francisco to take the well-paying job he needs to pay off that huge menagerie of college loans. The VP may want to pay Jordan with wheelbarrows full of cash (and she did!), but that desire to improve Jordan’s future means nothing when he’s disallowed from entering the housing market because of the financial 8-ball he’s starting his adult life behind. And that’s not racism at play, that’s just an authoritative algorithm spreading misery from the past into the future. And you should be legitimately concerned at this point.
You should be legitimately concerned because, as danah boyd succinctly put it “Technology doesn’t produce accountability [and] Removing discretion often backfires.” And when most of our economic lives are governed by black boxes full of trade-secret algorithms replacing the old arcane actuarial tables, then even in a best-possible-faith situation the negative impacts of past -ist behaviors will be used to encode -ism into our future.
I started off this post with the routinely disregarded financial wisdom that “past performance is not an indicator of future results,” a disclaimer that goes on every stock and mutual fund prospectus. But we ought to now repurpose it to serve as a reminder that it’s a mistake to pre-judge somebody’s future simply based on what we can reasonably infer from their past. It’s a mistake to not rent an apartment to a bright young engineer starting a promising career at one of the hottest companies in the Bay Area just because college is a financial tribulation these days. It’s a moral abdication to make people pay more for the risks they incur based on how people, civilization, and chance treat them rather than trying to help them mitigate those very risks. And thus it’s wrong to not take actions that affirm people’s ability to perform better and achieve more when they’re not being disadvantaged by their demographic correlation to risk. That’s what we’re supposed to mean by “Affirmative Action.” It’s not supposed to be a question of quotas (as dumbly implemented as the statistics they’re intended to counteract), nor should it be about lowering standards; it’s about seeing how well somebody does despite being almost certainly disadvantaged and then giving them the opportunity to show what they can do when they’re not disadvantaged anymore.
And so our legitimate concern ought to sometimes take moral action to affirm the idea that citizens of a just society would not be historically disadvantaged by other people’s hostile/predatory behaviors; here justice imposes a counter-factual responsibility to look to a person’s possible future regardless of their statistical past. Contrariwise, it is immoral to abdicate this responsibility to a machine built to enact corporate policy. Past performance is not an indicator of future results.
So I’ve covered what I wanted to cover and avoided what I wanted to avoid. But there is one last thing that I should briefly address, so kids, pay attention: affirmative action, even in its dumbest and sloppiest implementation is not about reverse racism. It’s about extending opportunity to somebody who worked hard to earn it with relative, if not absolute, achievement. It’s not really about you at all. And if you are concerned, then allow me to suggest that your concern is not that other people are being given opportunities, but rather that the adults of the world are doing a piss-poor job of creating opportunities in general such that they seem scarce in a way that make The Hunger Games a totally reasonable allegory rather than the obvious horror show unfit for consumption in civil society that it totally is. That’s why you almost certainly mis-attributed a cold elitism to the executive that just couldn’t hire a person who had bad credit before I explained what was going on. But you doubt this, consider: the simple economic fact of the matter is that “the richest 80 people in the world alone now possess more combined riches than do the poorest half of the world’s population” and that ratio is getting sharper over time.
Or, to put it another way, when “you wouldn’t be so annoyed that Jordan got to interview for that job and you didn’t, except you just know you’d have done better in the knife-fighting part of the interview than Jordan” — the problem isn’t who got the interview (affirmative action), it’s that there was a knife-fighting component to it (scarce opportunity). I know this doesn’t help your situation, but I do hope it helps your perspective.
Update – Further Reading: “When Big Data Becomes Bad Data” by Lauren Kirchner at ProPublica.