A statistical test used to detect discrimination by comparing outcomes across groups, like loan repayment rates for white versus Black applicants. It's popular in lending, criminal justice, and elections, but the speaker explains why it has serious problems.
White applicants or white loan recipients and they got paying back their loans, but 90% of black loan recipients ended up paying back their loans, then it suggests a double standard. It suggests that the lender, the bank was only issuing loans to extremely credit worthy black applicants who was issuing loans kind of willy nilly to the white applicants. And here, at least on the surface, it feels like this is not subject to the same problem that we had before. And we're not talking about the distribution of credit scores, we're not talking about any of that. We're just saying some people got a loan and some people didn't get a loan. Now look at the people who got the loan. If the repayment rates were higher for one group than another, it suggests that that group was subject to a higher standard. We're only lending to them if they were very, very. I just want to pause for a second. Does that argument make sense? So see people avoiding eye contact, we're all 100% in agreement. So this is, you know, this is very, very simple. And in fact, you know the argument, it's like it's in some ways exactly the type of test we want. You want something that feels like it's easy to understand. Like you can say this in front of a judge, you can say this in front of a juror and you'd be like, look, this is what the test. And it's relatively easy to compute. Like you just need a couple summary statistics, often with data that you have. So it's great except for the fact doesn't quite work. And this is what I want to talk about. It's like, you know, this is very compelling intuition, but it doesn't quite work at the same time. It's still one of the most widely used tasks across domains. So not only in lending and the original examples that was considered here, but also criminal justice and police employment, economic qualification, political elections, there was gender bias looking at the success of elected candidates. And so in all of these domains it's a very, very common test. And again, I think that the popularity of these tests stems from the simplicity of computing the statistic. It's something you can do in an Excel spreadsheet. You don't really need anything, even just like summary reports. These statistics are just out there and so it's easy to do and the intuition feels very compelling. So now what is the problem with this text? And this is an observation that's been made, it's been made for a long time.
