Not enough discrimination?

Posted on by

Aleks pointed me to this article by Stan Liebowitz on the recent financial crisis:

At the crisis’ core are loans that were made with virtually nonexistent underwriting standards – no verification of income or assets; little consideration of the applicant’s ability to make payments; no down payment. Most people instinctively understand that such loans are likely to be unsound. But how did the heavily-regulated banking industry end up able to engage in such foolishness? . . . it was the regulators who relaxed these standards . . . a “landmark” 1992 study from the Boston Fed concluded that mortgage-lending discrimination was systemic. That study was tremendously flawed – a colleague and I [Liebowitz] later showed that the data it had used contained thousands of egregious typos, such as loans with negative interest rates. Our study found no evidence of discrimination. . . . Ironically, an enthusiastic Fannie Mae Foundation report singled out one paragon of nondiscriminatory lending, which worked with community activists and followed “the most flexible underwriting criteria permitted.” That lender’s $1 billion commitment to low-income loans in 1992 had grown to $80 billion by 1999 and $600 billion by early 2003. Who was that virtuous lender? Why – Countrywide, the nation’s largest mortgage lender, recently in the headlines as it hurtled toward bankruptcy. . . . This damage was quite predictable: “After the warm and fuzzy glow of ‘flexible underwriting standards’ has worn off, we may discover that they are nothing more than standards that lead to bad loans . . . these policies will have done a disservice to their putative beneficiaries if . . . they are dispossessed from their homes.” I [Liebowitz] wrote that, with Ted Day, in a 1998 academic article.

There’s a lot of interesting stuff here, from the frustration of an academic Cassandra whose 10-year-old warnings have been ignored, to the interplay between economics and politics in setting bank lending rules. Liebowitz presents the story as if it’s obvious that Countrywide etc. were making bad loans (in his words, “Sound crazy? You bet”). It’s hard for me to believe that the president of the Boston Fed and the activists at ACORN were strong enough to muscle the 1995 Congress into passing such a bad law. There must have been some other arguments or interests in the law’s favor. Liebowitz must be oversimplifying the political aspects of his story (although I can see how he might do that in his frustration of having predicted the problem in 1998 and having seen nothing done about it).

It all comes back to statistics (of course)

My real interest in this story (and, I assume, the reason why Aleks sent to me) is statistical. Liebowitz is talking about rules for statistical inference, prediction, and decision making that aren’t allowed to use certain information, even if that information has predictive power.

To remove this from the politically-charged area of racial discrimination, let me give an example from education. Suppose you are giving final grades in a college calculus class, and you can use the following information: homework grades, midterm exam score, and final exam score. The goal of the grade is to assess how capable the student is at calculus. Presumably the best estimate will be some weighted average of homeworks, midterm, and final.

Now suppose you have some students whose final exam scores are missing–for simplicity, imagine they are missing completely at random, and for some reason you can’t have the students retake the exam. For these students, you can estimate what their final exam scores would’ve been, by fitting a regression of finals on midterms and homeworks, using the other students in the class.

OK, fine. Now imagine that you have one more bit of information available on the students–their math SAT scores. And further suppose that this variable adds predictive power: that is, in a regression of final exam scores on midterms, homeworks, and SATs, that the coefficient of SAT is clearly positive. The question is: should you use it? What if you have two students with identical homeworks and midterms, but one got a 500 on his SAT and the other got a 700? Should you impute a higher final exam score to the kid with the 700 SAT? This would be a better predictor, but somehow it doesn’t seem fair. If anything, it almost seems unfair in the other direction, that the overachiever who did so well despite his 500 SAT would get punished. But, really, homework and midterm exam scores are noisy, and more information should be better.

Now you can consider sex and race as additional predictors and you see the problem. In many settings, information is available that you can’t use because of some fairness rule. (The #1 example, maybe, is pre-existing conditions in medical insurance.) No easy answers, but to me it helps to think how it plays out in an apolitical measurement setting.

And now for something completely uninformed

To get back to the mortgage mess . . . my own, completely uniformed, take is that moral hazard has a lot to do with it. I can’t find the article where I read this–I think it was in the Times–but I saw some discussion of how various industry and government people were trying to put together a solution that would allow for some relief, but still allow people to get the profits if house prices went back up again. Hey . . . where’s the money for that coming from??

P.S. Commenter bccheah points to this interesting New York Times article on the topic.

7 thoughts on “Not enough discrimination?”

  1. Moral Hazard has everything to do with this… The way I see it is:

    1) Lax rules and enforcement in the equities market leads to a stock bubble in which employees are receiving huge quantities of options, IPOs are set at low prices and zoom high in the first few days of trading, and the main industry making money is the financial services industry.

    2) People holding these stocks eventually figure out that they're going out of business, and dump them.

    3) The resulting liquidity crunch leads the Fed to bail out financial services by pumping money into the economy.

    4) The resulting "free money" and the fact that financial services industries have made out like bandits leads them to write as many home loans as possible.

    5) HGTV becomes hugely popular.

    6) Interest rates reset and government does what the financial services companies expected… bails them out again.

    If there is ANY industry that is actively seeking out an edge and will produce quantitative estimates of the correct moral hazard level they should take on, it is the financial services industry. And, btw, that moral hazard level is very HIGH based on historical government actions.

  2. "these policies will have done a disservice to their putative beneficiaries if . . . they are dispossessed from their homes"

    There can also be a disservice for some of those who hold on to their bad investment and continue to attempt to service their loans. There is also a disservice to everyone who wisely decided not to become a speculator, take on an ARM, or state their income was more than it actually was. Those who stayed on the sidelines will be paying for those who gambled.

    http://www.voiceofsandiego.org/articles/2008/04/2…

  3. You and the author of the article are talking about two different approaches to race, and neither does much to explain the subprime mess.

    There is a race-blind approach that simply insists that the race variable be censored, and there is an affirmative-action type approach that insists that race be counted and acted upon.

    Your censored math scores are the race-blind approach, but I just can't see how the race-blind approach could cause a credit crisis. If a bank behaves as normal on credit scores and financial history and forces itself to omit the race variable from consideration, subprime loans still won't happen.

    No, for an antidiscrimination law to cause a mortgage crisis, we can't just have censoring of one variable, but the forced lending to people with bad credit because they are of a certain race. The article you link to claims that this is what the 1977 Community Reinvestment Act did, and that caused the subprime mess. It's a common theme; I've heard it myself a dozen times. But there's no evidence: the letter of the law is race-blind, not affirmative-action, and what evidence there is of the implementation doesn't find much link between the CRA and lending to bad credit. Here's a further discussion of the data on and polemic uses of the Community Reinvestment Act.

    So your argument that the race variable was censored doesn't address why banks developed lax lending standards; the article's affirmative-action argument just doesn't correlate with reality. Race is just not enough to matter.

    Banks, mortgage brokers, S&Ls, &c. started lending to people with bad credit and thus induced this mess because they found ways to minimize, profit from, shift, or otherwise deal with the risk. There's just no need to resort to stories of government coercion to explain all this. I mean, how many mortgage brokers did you hear five years ago kvetching about how the government forces them to take on risky loans with high commissions?

  4. At least constitutionally speaking, the rationale for not allowing some of those additional predictors (such as race) to be used is that they *ought* not be predictive of anything once the other (non-prohibited) predictors have been factored in. One could argue that prohibiting their use will/should, in the long run, reduce their predictive power relative to what it would have been had they been allowed to continue to be used (think about the social/economic/whatever status of African Americans in the South before and after Jim Crow). If we think in these terms, then enshrining a protected class X in the constitution amounts to a societal commitment to drive hat{eta}_X toward zero over time.

  5. A solution to the problem of legal prohibition of such "phobic" variables as race and sex would be to use surrogates for them (or build models to predict them and use the predictions).

    For example, in a predicting recidivism of parolees, sex is a strong predictor but can not be used by the authorities because of legal restrictions. Since most prisons house only s single sex, one could get around this by adding prison name as a nominal predictor. Of course one would have to give up several more degrees of freedom, but, hey, there was plenty of data.

    I would certainly hope this is what the medical insurance companies Prof. Gelman mentions are doing, otherwise I may be paying too little/much for my insurance.

  7. Bond ratings offer about 16 levels, and about 3 major ratings company.

    Even then, one can simply look at their performance by looking to see if bond ratings are smoothly distributed. They are not.

    Bad estimation is all over the place. But I would start with the ratings companies, and the foolish banks who weren't watching the ratings spread.

Comments are closed.