Using data science to increase economic mobility, reduce bias
on Balance of Power podcast [36:28] June 12, 2020
DAVID WESTIN: US jobless claims came down a bit but remain over 1.5 millio. The Federal Reserve sees unemployment at or near 10% the end of the year, begging the question — what more can we do?
Give us a sense of what we could be doing better in this really dire situation with employment, to get people back to the right jobs?
MICHAEL ROSENBAUM: Thanks, David, for having me. Obviously, Coronavirus over the last 90 days has had a devastating impact on the economy. But some of the dynamics that are at play have really been going on for a long time. The issue of lack of economic mobility. The issue of matching folks to jobs where they can thrive and find dignity. We’ve been talking about the future of work in this context for years now.
Arena focuses on one piece of this, which is the informational asymmetry attached to it. The labor market is not very efficient. We like to hire people like ourselves. Generally, we’re not very good at hiring. As individual job applicants, we generally aren’t incredibly good at figuring out where we’re going to thrive.
Arena is built to predict the likelihood that an individual will achieve an outcome at a job, without relying on a resume. In doing so, it gives an employer the ability to hire more accurately. And it gives individuals the ability to find somewhere where we would thrive.
The entire premise behind Arena is that resumes are poor predictors of outcomes
DAVID WESTIN: So anyone who’s run any significant institution that does a lot of hiring – and I have done that – knows exactly what you’re saying. There’s an innate tendency to hire people who are like you. You feel more comfortable with it because you know them. You know how to evaluate their background. How does Arena get around that? Because the employer has to know something about the background of the person, don’t they?
MICHAEL ROSENBAUM: Yes, but the entire premise behind Arena is that resumes are poor predictors of outcome. Resumes correlate specifically with class, race, and gender. But they’re generally not great predictors of success at a job. There may be data points in a resume that can tell you something. But there’s a lot of other data you can collect on someone using modern machine learning technologies that allows you to make those predictions much more effectively — such as, keystroke data.
You think about Amazon and Netflix, they use keystroke data in a variety of ways to predict consumer outcome. What’s the likelihood that I’m going to buy something on an Amazon site? One of the data points on that is, how long I hover over a review and what words are in that. Netflix uses a whole range of data related to what we look at in the catalog of movies to figure out what we’re likely to want to watch.
So we’re using essentially the same technologies to use data outside a resume to predict the likelihood that someone will achieve an outcome. The result of that, you can have a really substantial impact on reducing the biases. Some biases are malicious — hopefully implicitly malicious — but others are not malicious. “I’d rather hire someone who went to a fancy four year college than to a community college into a job.” Even though it may be that the person with the community college degree would be better at the job.
DAVID WESTIN: Any of us who’ve been involved in hiring people know exactly what you’re talking about. That is exactly right. And even if we try to screen ourselves out for it, we have a tough time doing it. So if it’s not a resume, that’s a poor predictor of how you’ll do at a job, what’s a good predictor?
MICHAEL ROSENBAUM: There are things that might come up in a resume – around a resume – that might be relevant.
Arena today is deployed primarily into healthcare, also into restaurants. In healthcare, Arena is deployed into organizations that process 3.5 million unique job applicants per year. So about 17% of the US healthcare workforce. I’ll give you an example of ‘predictors’ that could tell us something:
A small hospital system, 7500 employee hospital system – has a hospital on one side of the street and a long-term care facility on the other side of the street. They hire Certified Nursing Assistants into both places. It turns out that if someone’s been a leader in a community organization they are slightly more likely to thrive in a CNA job in the hospital, and slightly more likely not to be happy and to quit the job in the long-term care facility. It only predicts less than 1% of the result, but it does tell you something.
The other piece really is Keystroke Data. Again, if you ask someone a question, do they open another browser, do they skip it, do they hover over something? That sort of data tends to be very significant in being able to predict whether or not someone’s going to stay in a job, show up to work on time in the job, report being engaged in a job, or be involved in an incident.
DAVID WESTIN: Applying your technique to Arena itself, what’s your track record, how do you know that you’re succeeding?
MICHAEL ROSENBAUM: Our clients typically always start with employee retention – that’s the metric they are looking to optimize for. Arena has a 100% success rate across all its clients at reducing employee turnover, [DAVID WESTIN: Wow] That doesn’t mean every individual, that means the enterprise. Our average, impact 12 months after going live is a reduction of turnover by 21%. Our average impact 24 months after going live is a reduction of turnover by 43%.
DAVID WESTIN: That’s really substantial.
MICHAEL ROSENBAUM: I would add as part of that, in every case, the workforce of our clients become more diverse across a wide range of axes, when using this technology. By reducing the bias we have — malicious or not malicious — we can have a more diverse workforce and, from an enterprise perspective, better operating numbers.
DAVID WESTIN: We could use both of those right now. That’s really fascinating. Thank you.