Melinda Thielbar

“Women never pay their bills.”

My friend was a young consultant working on a model for a small-ish company. The company wanted to know which customers were likely to be “problem” customers, i.e. who was likely to not pay their bills for so long their paperwork would be sent to a collections agent. The idea was to skip the two-week process where someone from their own office tries to collect and just send some of the bills straight to an outside collections agency.  The outside agency charged a fee, and so it was important to select the “right” bills to collections and collect on the rest of the bills themselves.

Hence the discussion with the office staff where my friend asked about their process. He was told, in no uncertain terms, that most of the problem customers were women.

Being a young consultant, my friend dutifully put gender into the statistical model—and found it was not important. He explained that it was not important. The office staff didn’t believe him. They showed him a bar chart for the customers who were in collection.

“See? Sixty percent of our problem customers are women.”

My friend stared. “Sixty percent of your total customers are women,” he said.

Then he and the office staff stared at each other for a while, and then my friend went away and found a different client.

Most consultants have this conversation at some point in their careers, and most of us eventually learn how to explain it. The simple explanation is this: If you have two people who are late on their bill, and everything about them is the same except one is female and one is male, there is no difference in the probability that they’ll never pay you. If you have to pick one or the other, you might as well flip a coin.

Business rules like “women never pay their bills,” are what I call Organizational Folklore. Someone said it. It sounded reasonable, so someone else repeated it, and by the next week, everyone believed it. Add in a bar chart to back it up, and suddenly every woman customer who walks in the door is treated like a potential thief. If you’re using Organizational Folklore to run your business, you’re probably not making good decisions.

Continue reading →

Once upon a time, I worked in training evaluation. Specifically, I studied training programs for the poor. You know, those jobs training programs that were supposed to “lift people out of poverty” by “making them self-sufficient”. My research at the time focused on how the data the programs were already collecting could be matched with other data that was already being collected by other organizations. It would have eliminated the need for costly back-end evaluation programs by people like me. If that sounds bleeding edge, I would remind the reader I was a young, starry-eyed graduate student at the time. It was 1998.

I was poised to start a Ph.D. in the subject when I realized something: Nobody cared about what I was doing. Politicians didn’t care. They just wanted to cherry-pick my studies to support whatever they wanted to do next. The people running the training programs didn’t care. They just wanted to show “good” results so they could get more money. My bosses mostly wanted to keep doing research, so there was this perverse disincentive to never solve the problem–and certainly no incentive to automate.

A comment from Gregory Piatetsky on Wednesday’s post made me think of those days, and why I decided to go work for a private company rather than stay in government research. I do think things are different now (I promise, Rick Pack, I really believe that). There’s been too much success with actual data based on real research done by people who care about good work. Better tools mean more output with less input, and even if that if that means there’s more noise, it also means it’s easier for people who can produce signal to be productive.

However, as another commenter pointed out, Congress wasn’t swayed by the results of a seriously flawed economic study. Congress found a seriously-flawed economic study to support what they wanted to do. They used the numbers as a way to say “I’m right and you’re wrong you can argue with me, but you can’t argue with figures so shut up,” instead of as a piece of information that should be considered in an important decision.

And that is what’s wrong with data science right now. Those of us who want to do good work (like that starry-eyed graduate student from 1998) have to be constantly selling the work, guiding decision-makers on how to use it, and guarding against the guy who’s going to pick out one number and yell louder (sorry, folks, it is usually a guy).

A mathematical model doesn’t end a conversation. It starts a conversation. It’s a way to walk yourself through a thought process, not a bigger stick for hitting your competition.  People who engage in lazy thinking, lazy argument, and lazy politics shouldn’t be allowed to use numbers–for the same reasons we don’t let 10-year-olds drive cars.  The world is complicated, and while bullet points are useful, they are not sufficient for reasoning about a complex world.

Particularly not if you’re going to make big decisions for the rest of us. Because no matter what you might think of your God-like powers, none of us is that smart.