The last-mile problem: How data science and behavioral science can work together

07 May 2015

If we want to act on data to get fit or reduce heating bills, we need to understand not just analytics, but how we make decisions. Press Room host Tanya Ott talks with Jim Guszcza about the art and science of applying the findings from masses of data in the context of how people actually behave.


The Press Room
Subscribe on iTunes
Listen to more podcasts
Read the full article.


Tanya Ott: This is The Press Room, Deloitte University Press’s podcast on the issues and ideas that matter to your business today. I’m Tanya Ott, and I’ll be talking to thought leaders and change makers about trends in emerging technologies, CFO strategy, risk and security, social impact, and more. Today we’re considering how business analytics intersects with the science of behavioral insights. In other words, how can business leaders translate mountains of analytical insight into productive and beneficial behavioral change?

Jim Guszcza: The combination of big data and behavioral nudge tactics are two new arrows in a quiver. So they might not have all the answers, but they might be useful tools to make improvements.

Tanya Ott: To tackle this topic, I’m talking to Jim Guszcza, chief data scientist for Deloitte Consulting LLP and author of the recent Deloitte Review article The Last-Mile Problem.

Let’s start by defining our terms. Data analytics – well, everyone has at least a vague idea of what this is, thanks to Brad Pitt.   Yes – Brad Pitt!

In the movie Moneyball, Pitt plays Billy Bean, manager of the Oakland A’s. The team is struggling, has no money to get better, and they’re up against much richer teams who can afford to buy high-profile players. The A’s need to find a competitive edge.

Jim Guszcza:   What Billy Bean discovered when he started using data analytics to better guide baseball scouting decisions was that he could actually compete against richer teams by using data to suggest which players would make good team members on his team on which occasions. He was able to hire these players for a comparatively low salary because other teams were not using analytics in the same way. They were making decisions based on the professional judgment of baseball scouts.

Tanya Ott: In this situation, and in many others, a sophisticated data model and a lot of computing power can do a better job at weighing options than we do on our own. That’s because no matter what we think about our brainpower, we’re not driven by pure, dispassionate logic

Jim Guszcza:   We’re bad at weighing probabilities. We overgeneralize from personal experience. Our estimate of the probability of something is a matter of how vivid it is or how familiar it is. We often confuse familiarity or ease of processing with truth. We tend to be overconfident in our abilities to weigh evidence and make good judgments. And these all have profound implications in the business world because if there’s one common denominator for all areas of business it’s about people making decisions day in and day out.   Models compensate for bounded cognition, irrationality. And once the model’s done its thing, if you find the right way to combine expert knowledge and the model and so on, the right decision to make is kind of straight forward. So in the Money Ball case, the right decision was, okay if the indication tells me this is going to be a great undervalued baseball player, the decision is, well, scout that player. Fine!   Or in insurance, if an actuarial model tells me that Jim is likely to be a terrible driver then maybe you want to raise Jim’s rates to a higher premium threshold or maybe just not even sell Jim insurance in the first place. Those are pretty, in principle, straightforward economical decisions.

Tanya Ott: But a lot of decisions aren’t as straightforward. What’s more, we’re not straightforward. We’re very good at discounting what’s good for us, at giving in to short-term impulses, at ignoring the things we know to be true.

Jim Guszcza:   Predictive modeling adds science to this process of combining information to arrive at a good decision. But right now, what I call The Last Mile Problem is that in many cases the gap behind the model indication and what you really want, which is the changed behavior, that’s still a matter for the gut feel, the unaided judgment of a caseworker in the field.

Tanya Ott: That’s where behavioral science comes in. Now, behavioral science doesn’t have an Oscar-nominated movie yet, so it might need a little more explanation. For that, we’ll turn to the book Nudge by Richard Thaler and Cass Sunstein.

Jim Guszcza:   The way we’ve talked about predictive models, those are sort of correctives to flawed unaided judgments.   The whole idea of nudges is a little different. That says that rather than try to correct for cognitive biases, let’s try to create policies and arrangements and choices that go with the grain rather than against the grain of human psychology.

Tanya Ott: As an example, consider saving for retirement. Say I’m just out of college, I’ve just started my first real job, and I’m confronted with the prospect of enrolling in my company’s 401K. And I KNOW I should be saving as much as possible, but there are choices to make, and forms to fill out, and a lot of research I should probably do, and… ugh, you know what, I’ll get to it later.

But if we look at something called choice architecture, we can make my tendency toward inertia work in my favor.

Jim Guszcza:   Rather than try to make people more ‘rational’, let’s try to simply design their choices in a way that kind of prompts them to make the choices they would make if they had more time and information. Richard Thaler, who’s the founder of behavioral finance, and one of his students at UCLA, what they found is that if companies set up their retirement and savings programs so that people are automatically enrolled and have the option to opt out, people are much more likely to save for retirement than if they are not automatically enrolled and have to opt in.   So that’s a really simple finding. It’s like something as momentous as what you save for retirement is based on a simple flip in the way the alternatives are arranged.

Tanya Ott: OK. So we’ve got data analytics – moneyball – on the one hand. We’ve got behavioral science – nudges – on the other. And in his article the Last Mile Problem, Guszcza starts thinking about what would happen if we put these two practices together.

Jim Guszcza:   The way I look at this is predictive models, and data analytics and the behavior nudge movement are two complimentary reactions to the same set of discoveries about human nature. That we do tend to kind of deviate from what classical economists call the Standards of Rationality. So it’s a very simple idea. Predictive models help us combine information in a better, more principled way than unaided judgment. And behavioral economics, in these cases where the ultimate goal is behavior change, helps us go from the model implication to the improved behaviors.

Tanya Ott: In other words – a model can tell us what needs to change, but behavioral science can tell us how to make that change happen.

Jim Guszcza:   A model tells you that this person is likely to succumb to a chronic disease. Or that this person is likely to fall behind on child support payments. A model tells you that this student is at high risk of dropping out of high school. Well, what are the right interventions to prompt those people to change their diet and exercise behavior? Become more financially responsible? Stay in school? What are the right interventions? And I think these are all areas where behavioral economics won’t have all the answers, but it’s a natural place to look for an enriched set of policy options.

Tanya Ott: It’s such a simple idea that it’s kind of amazing to think it’s not already happening. But Guszcza says it’s rarer than you’d expect.

Jim Guszcza:   I think a few people have begun to combine them but there are only a few isolated examples that I’m aware of out there.

Tanya Ott: One of them comes from President Obama’s campaign organization in 2012

Jim Guszcza:   Predictive models were used to identify which voters were likely to be persuadable. If I can knock on Sarah’s door I can persuade her to vote for Barrack Obama. She might not vote for Obama if I don’t knock on her door, but if I knock on her door and give her the right spiel, yeah, she’s much more likely to vote for Obama. So the last mile is, well, how do I actually prompt that behavior change on Sarah’s part? How do I actually prompt Sarah to go out on Election Day and vote? And the strategy the Obama campaign used was commitment cards. So it was simply a card with Barack Obama’s picture on it and a “Yes, We Can” slogan, but also a place where Sarah could write down her specific plan for going to this specific polling location at this specific time on voting day. Psychologists have found that when people articulate a goal, they’re more likely to follow through on what they say.

Tanya Ott: In the article in the Deloitte Review, Guszcza talks about how this nudge could be applied in the public policy realm.

Jim Guszcza:   my colleagues and I have built models to help child support enforcement officers to proactively non-custodial parents who are at risk of falling behind on child support payments. But it occurred to me that this could perhaps be improved if we could test various behavioral nudge tactics to help parents stay current with their child support payments. So maybe non-custodial parents filling out commitment cards saying that they will write the check on a certain day, maybe that will help them stay current with their child support payments.

Tanya Ott: Commitment contracts are a particularly fascinating nudge, because you can make a contract with yourself – and give it some teeth.

Jim Guszcza:   It’s my present self, setting up a contract that will actually nudge my future self to go to the gym. If I want to lose weight, if I want to stay on an exercise regime, what I can do is I can agree in advance to donate $1,000 to a charity of my choice if I don’t go to the gym if I don’t go to the gym a certain number of times in the next several weeks. So if I fail to go swimming, say, six times in the next two weeks I will donate $1,000 to a charitable organization. Or if I want to make it even more pressing, I’ll donate $1,000 to an organization that I really loathe! So my present self knows that it’s important for me to exercise. This is my long-term goal: to lose weight, to stay healthy. But I also know that on any given moment, on any given morning, on any given bit of free time, I’ll probably give into temptation and more likely than not fail to go to the gym. This happens to people all the time.

Tanya Ott: And if you want to make that commitment contract even more likely to be honored, you get a friend involved.

Jim Guszcza:   People who sign commitment cards against their future selves to go to the gym or to diet, those commitments contracts are more likely to result improved behavior if a friend of theirs is monitoring their behavior. You know, they don’t want to let their friend down.

Tanya Ott: And that insight sparked another idea for Guszcza – what if predictive models could help pick the RIGHT outside monitor to prompt that desired behavior.

Jim Guszcza:   I got this idea from a wonderful article by Atul Gawande it was published in the New Yorker a few years ago. And he reported success with using health coaches. And he actually reported a very poignant story about a woman who had multiple co-morbidities. She was diabetic. She was obese. Her husband nudging, wouldn’t get her to change her behavior. But when she was assigned a certain health coach she actually did start doing yoga classes, doing standard medications and so on. And (insert name) asked her, you know, why did you change your behavior? Why did you listen to this health coach and not your husband, she responded, “Well, she talked to me like my mother used to talk to me.” And that got me thinking, well, yeah, maybe analytics could actually be used to do these nudges in the right way.   You know, what are the kinds of people who are well-equipped to be health coaches? It’s not necessarily a matter of education or training. It’s more a matter of being a certain kind of person with a certain kind of social perception, a certain kind of empathy, a certain kind of patience. And maybe a certain kind of cultural background that resonates with the cultural background of the person you’re trying to coach. And so it occurred to me that a matchmaking service might be just the right way to select health coaches and match health coaches to the right patient.

Tanya Ott: The combination of moneyball and nudge can go further than that, especially when it comes to health care – because there’s so much more data available now. In fact, you may be collecting it right now.

Jim Guszcza:   Behavioral data are these little digital breadcrumbs that we leave behind as we go about our day to day activities. Self-track devices are a great example. You know these little bracelets that we wear that track how we slept, how many paces are we talking each day and so on.

Tanya Ott: All that data can be especially powerful when it’s put in context – and it incorporates a little positive peer pressure. That’s something one energy company is already using to great effect.

Jim Guszcza:   Opower is very famous for giving people utility feedback reports saying, ‘Hey Sarah, did you know that you used 70% more energy than most of your neighbors with similar households. And what they’ve found is that those kinds of reports will prompt people to try to actually figure out what about their houses, what about their lifestyle is causing them to consume a little bit more energy and will actually prompt them to save energy. So it turns out that being compared to your peers is much more effective than economic arguments or even arguments about the environment in getting people to change their energy utilization behavior.

Tanya Ott: Another example of how data and behavioral sciences can work together comes from the realm of insurance.

Jim Guszcza:   All the data that insurance companies collect for its pricing activities have been traditionally used in this kind of actuarial way. Can we use this new data set to do a better job of segmenting risks, doing a better of coming up with the best price for every individual risk. These fine grain data about the way I accelerate or the way I change lanes, the way I speed up or do I speed in residential areas, absolutely can and should be used to do a better job of rating insurance and setting appropriate price to appropriate risks. But now that we’re in this age of cloud computing and digital mobile devices, we can also think about using this data to actually create new products and services. The general idea is that if these risk scores are good for the insurance company, why aren’t they good for the drivers too? So that the data can be used to give fine-grained feedback reports back to the drivers so they can improve their driving behaviors.

Tanya Ott: That’s because we tend to think we’re better at driving than we actually are – and seeing evidence of our ineptitude can make a difference.

Jim Guszcza:   People are famously overconfident about their driving ability. Going back to overconfidence bias. Most drivers think they’re better than average drivers. Which can’t possibly be true if you think about it for just a second. But if you’re actually given feedback saying (bad edit), ‘no, you’re worse! That’s likely to get a lot of people to change their driving behavior. If the data is sort of reconceived to be used not only for actuarial purposes or predictive modeling purposes, but also for to nudge better driving behavior – well, that’s a win for the insurance company, that’s a win for society and it’s also a win for the individual driver.

Tanya Ott: Right now, tools like this are all voluntary. And Guszcza says may be a consideration on how business products that combine data analytics and behavioral science are received.

Jim Guszcza:   I do think that this is sort of like a technology. It could be used for undesirable ends. But it’s also a set of tools that can help people make decisions that better conform to their long-term goals if used in a proper way and if it’s used in a way that’s upfront and people agree with.

Tanya Ott: You don’t want your nudge to be perceived as a shove, or for your data collection to look like snooping. But approached the right way?

Jim Guszcza:   the combination of big data and behavioral nudge tactics are two new arrows in a quiver. So they might not have all the answers, but they might be useful tools to make improvements.

Tanya Ott: Now these are just some of the ideas Jim Guszcza proposed in his article The Last Mile Problem. Check it out. You’ll find out what colored paper can do for compliance. Why your business communication should read like it’s written by a human being. And how a networked soap dispenser could save thousands of lives a year. You can find the article on  I’m Tanya Ott for The Press Room, a production of Deloitte University Press. Thanks for listening and join us again soon for another edition of the Press Room.

This podcast is provided by Deloitte LLP and is subsidiaries and is intended to provide general information only. This podcast is not intended to constitute professional advice or services of any kind. For additional information about Deloitte LLP and its subsidiaries, please go to