Second Uber Science Symposium: Exploring Advances in Behavioral Science
May 20, 2019 / GlobalOn May 3, 2019, Uber’s Applied Behavioral Science team hosted the Behavioral Science Track of the company’s Second Uber Science Symposium at our San Francisco office. The program featured a full day of presentations delivered by leading researchers in the behavioral, social, and cognitive sciences from technology companies, consulting firms, and universities. The track was attended by members of academia and industry who enjoyed deep dive talks on topics ranging from smartphones and happiness to psychologically-informed machine learning models, as well as conversation and networking with fellow attendees and invited speakers.
Our Behavioral Science Track aimed to facilitate the sharing of information and best practices, foster discussion, and form bridges between research communities. To this end, we were joined by twelve speakers from the greater behavioral science community and over 400 attendees (more than 150 in person) from diverse perspectives and organizations, including tech, consulting, business schools, and psychology departments. We also accepted submissions from attendees for our “data blitz,” a session comprised of five-minute lightning talks, selecting a dynamic lineup of presentations on theoretical, applied, and methodological developments.
This Uber Science Symposium was the second in a series designed to bring together researchers and practitioners from different fields. In addition to Behavioral Science, the Second Uber Science Symposium included two parallel tracks with full-day programs, Programming Systems and Tools and Bayesian Optimization. The First Uber Science Symposium, held on November 28, 2018, focused on Deep Learning, Reinforcement Learning, Natural Language Processing/Conversational AI.
Below, we share some program highlights from the Behavioral Science Track of our second Uber Science Symposium.
Melanie Brucks, Assistant Professor of Marketing at Columbia Graduate School of Business, shared her research on technology-mediated innovation. Previous studies of collaborative innovation have shown that unfocused (divergent) thinking works well for idea generation, while focused (convergent) thinking is good for selecting the best idea. For centuries, both of these processes took place in person, but over the last few decades, technology has allowed groups to collaborate across long distances, particularly via video link. Melanie’s research asks, how does this technology affect innovation?
Using laboratory experiments and field studies, Melanie investigated idea generation and idea selection under in-person and virtual video communication conditions. Across studies, in-person conditions were better for idea generation, but virtual conditions were better for idea selection. To understand this disparity, she measured visual and cognitive focus via eye-tracking and memory tests, and found that virtual communication led to more focus on the screen and less focus on the environment. These findings suggest that groups should do idea generation in in-person meetings but idea selection in virtual meetings.
A Geometric Approach to Characterizing our Experiences and How We Remember Them
Jeremy Manning, Assistant Professor of Psychological and Brain Sciences at Dartmouth College, presented his research on the dynamics of human thoughts. Thoughts are comprised of layers of overlapping content from both external and internal sources that change over different timescales, such as hearing the alarm clock (external, short timescale), needing coffee (internal, medium timescale), commuting to work (external, medium timescale), or being in San Francisco (external, long timescale). Later in the day, remembering your alarm going off will activate other thoughts that occurred at the same time, like needing coffee, as well as temporally adjacent thoughts, like commuting to work.
Jeremy’s lab used behavioral experiments and functional brain imaging to track how thoughts unfold over time. Video content was carefully coded and projected into multidimensional space, with each point representing a moment in time, resulting in maps of how the video progressed through “thought space.” Then, experiment participants watched these videos and retold what happened in the videos while undergoing functional brain scans. Just like with the videos, multidimensional maps were created of how people’s retellings and their patterns of brain activity evolved over time. By comparing those cognitive and neural thought maps to the video’s thought map, as well as to the thought maps of other people, Jeremy’s lab measured how thoughts and memories converge or diverge from objective experience and from the group. The lab is beginning to apply this technique to online learning videos to identify moments to intervene in order to optimize learning.
Using Behavioral Science to Improve Financial Wellbeing
Wendy De La Rosa, co-founder of Common Cents Lab, shared how behavioral science can help understand and address financial insecurity in moderate to low-income individuals. Her research shows that, while people want to be financially secure and understand how to improve their financial security, they don’t actually do it. Increasing information alone doesn’t lead to behavioral change; in fact, financial education accounts for only 0.1 percent of financial behavior change. One line of experiments investigated the effect of online banking tools designed to reduce behavioral barriers to saving or increase behavioral barriers to spending.
In one experiment, Wendy’s team worked with an online financial firm and asked people what percent of their tax refund they wanted to save, either before they received the refund or after. People who were asked before they received the refund saved at a rate of 27 percent, compared to 17 percent among people who were asked afterwards. In a second experiment, Wendy’s team worked with online banks to help people in the U.S. government’s Supplemental Nutrition Assistance Program (SNAP) better plan their spending during the recent government shutdown, when March benefits were disbursed two weeks early. When people had the option of hiding a portion of their early disbursement, spending rates decreased and smoothed over time, whereas without this option, spending spiked right after the benefits were disbursed, making it more difficult for recipients to stretch the March disbursement until the end of the month.
Personalizing In-app Subscriptions
Julian Runge, a computational social scientist and data scientist in mobile gaming, showed how behavioral science and machine learning can be combined for superior business impact. In mobile gaming, where freemium is the predominant pricing model and the conversion rate to paid is low, pricing algorithms are designed to optimize for conversion and revenue. Behavioral science research on the optimal pricing of experience goods suggests a “skim-then-penetrate” approach, where prices start off high to capture non-price sensitive users but drop for people who don’t convert at the higher price point. Personalization algorithms can help learn where to set that initial price point and the subsequent drops for different segments.
In an experiment testing each of these approaches separately, as well as a third treatment combining them, Julian’s team found that the combination of behavioral science and machine learning led to a large, statistically significant lift in conversion and revenue compared to a randomized hold-out receiving the best non-personalized treatment.
Harnessing Technology to Increase Social Connection and Wellbeing
Elizabeth Dunn, Professor of Psychology at the University of British Columbia, presented research from her lab on how social connections make people happier, as well as the role of technology in social connection, and shared new insights on how to increase social connectedness. Liz’s research has shown that tiny, seemingly insignificant social interactions, like with your barista or rideshare driver, have a big impact on a person’s happiness. But lately, technology such as smartphones supplants much of this casual social interaction–everyone stares at their phones instead of talking to each other. Liz asks, why don’t people talk to strangers?
In a brand new experiment, Liz’s lab tested the hypothesis that people don’t talk to strangers because they underestimate others’ willingness to talk. During a 20-minute group lunch, participants were asked to wear a green wristband if they would be interested in talking with someone new, or a red wristband if they’d prefer to keep to themselves; these wristbands served as explicit social signals about willingness to talk. In a control lunch group, no prompts were given about social interaction preferences and everyone wore white wristbands.
During the social signal group lunch, people with green wristbands interacted with other participants about three times more than people with red wristbands. Notably, in the control group, participants spent about the same amount of time talking with others as the red wristband group in the experimental condition. This suggests that, in our society, we act as though we are all wearing red wristbands, i.e., not interested in talking with others.
Are Firms Loss-Averse? Pre-payments, Bonus Claw-backs, and Sales Performance in the Auto Industry
Charlotte Blank, Chief Behavioral Officer at Maritz, shared an example of how behavioral science is highly contextual. Maritz partners with car manufacturers to incentivize dealerships with bonuses based on sales targets. Previous behavioral science research on bonuses in another context–school teachers–showed that giving teachers their performance bonus upfront, but clawing money back at the end of the year if they didn’t reach performance goals, was more effective for improving student outcomes than waiting until the end of the year to give the bonus. This effect is consistent with the behavioral concept that people are loss averse: the threat of losing money you already have is more motivating than the prospect of gaining that same amount of money.
Maritz wanted to test whether this same principle would apply to car dealerships. In an experiment, car dealerships were randomly assigned to either the up-front bonus with clawback or end-of-year bonus condition. Unlike in the teacher experiment, the up-front bonus condition actually hurt dealership sales, highlighting the context-dependence of loss aversion and the importance of running experiments to test behavioral science ideas in practice.
Adding the Self into Choice Architecture
Hal Hershfield, Associate Professor of Marketing, Behavioral Decision Making, and Psychology at UCLA, presented his research on how to help people make choices that benefit themselves in the future. Hal shared behavioral and neuroscientific studies suggesting that people see their future self as another person that they’re not particularly close to emotionally. Lacking this closeness, there is low motivation to do things, such as saving money, that would benefit this future self. Hal asked, how then do we help people feel closer to their future self?
In a series of experiments, Hal and colleagues tested different interventions aimed at increasing financial saving. In one experiment, people who saw an aged representation of themselves were more likely to contribute to a retirement plan. In another experiment, merely asking people to think about the future was sufficient to increase savings. In a third experiment, even manipulating the perceived closeness of “the future” in a visual representation of time led to a savings increase. Hal also highlighted another approach that makes saving feel less painful for the present self. For example, framing a savings plan as $5/day had a higher opt-in rate than the same plan framed as $35/week or $150/month. Together, these studies show that small interventions can have a large impact on savings decisions.
Dynamic Identity Processes and their Implications for Motivating Attitude and Behavior Change
Neil Lewis Jr., Assistant Professor of Communication and Social Behavior at Cornell University, shared research on how identity influences attitudes and behaviors in social contexts. If you want to persuade people, you need to figure out who says what to whom with what effect, and an important part of doing this is “knowing your audience.” Audience segmentation assumes that (1) motivation and behavior are identity-based, (2) particular identities are especially salient, and (3) identities are static. Neil discussed how empirical evidence does not strongly support (2) and (3). In particular, when considering an identity, e.g., race, there are multiple dimensions, for instance norms, social status, and skin color, and different dimensions might matter in different contexts. Additionally, identity processes are dynamic, and social-contextual factors influence when a particular identity is important.
In a field experiment, Neil and his colleagues measured how much time patients in a clinic spent paying attention to a public service health announcement playing on a TV in the waiting room. In the experimental condition, the message was about HIV prevention; in the control condition, the message was about flu prevention. The study also recorded the number and identity (e.g. race, gender) of other people in the clinic at the same time. One striking finding was that, when even one other black person was in the waiting room, the amount of time spent looking at the HIV message by black patients dropped to zero, but this effect was not seen in the flu message condition. These results emphasize the complex relationship between identity, context, social interactions, attitudes, and behaviors.
Leveraging Behavioral Science to Generate Deeper Insights to Inform and Activate Marketing Strategy
Namika Sagara, President of the Behavioral Science Center at Ipsos, leads a behavioral science group that does applied research and consulting. She presented two case studies of successful behavioral science-informed marketing strategies. In one example, her team applied a science-based habit formation framework to increase how often owners gave their dogs wet dog food. They identified a new contextual cue, or use case, for feeding wet dog food that was concrete, unique, embedded, and salient (“include your pup in your special Sunday family dinner”), and reduced the barriers or negative experiences associated with feeding wet dog food by making the package easy to open and reseal without touching the food.
In another example, Namika’s team applied regulatory fit theory. This theory posits that people tend to be in either a promotion mindset where they are trying to make good things happen or a prevention mindset where they are trying to prevent bad things from happening. Marketing content is more effective when it is aligned with a person’s mindset. In their consulting role, Namika’s team suggested that an oral care company branch out from their prevention-focused messaging on preventing gum disease and high-cost oral treatment and create separate marketing campaigns with promotion-focused messaging on fresh breath and fast, thorough mouth cleaning.
Data blitz
The Behavioral Science Track program concluded with a fast-paced “data blitz” of five-minute talks from audience members. Prior to the symposium, all participants were invited to submit short abstracts of their presentations. Seventeen abstracts were submitted, and talks were selected on the basis of scientific merit and topic breadth.
From Behavioral Experiments to Computational Behavioral Science
Steve Wendel, Head of Behavioral Science at Morningstar, showed that by leveraging behavioral simulations, his team can predict the long-term effects of behavioral interventions where only short-term impact measurement is possible.
A Psychologically-Informed Methodology for Modeling and Evaluating Users’ Similarity Judgments
Amy Winecoff, Senior Data Scientist at True Fit, shared how she applied theories and methods from cognitive psychology and decision-making science to build a superior machine learning model for predicting users’ fashion similarity judgments.
Ideas about Liking Predict Situation Selection at a Distance
Andre Wang, PhD Student in the Psychology Department at UC Davis, presented his research on the disconnect between what people say they like and what they actually like in an online dating context.
Marshmallows and Smartphones: Effective Learning in the Real World
Iain Harlow, VP of Science at Cerego, an online learning platform, shared data on contextual factors that affect real-world learning, including session duration, time of day, and intrinsic vs. extrinsic motivation.
Interested in tackling science at scale with Uber? Consider applying for a role on our team!
Acknowledgements
We want to thank our engaging speakers and energetic attendees for joining us, connecting with each other, and helping to make the Behavioral Science Track of the Uber Science Symposium a success.
Laura Libby
Laura Libby is senior data scientist on Uber Labs, the Applied Behavioral Science team.
Joshua Morris
Joshua Morris is a data scientist on Uber Labs, the Applied Behavioral Science team.
Candice Hogan
Candice Hogan is a data science manager leading Uber Labs, the Applied Behavioral Science team.
Posted by Laura Libby, Joshua Morris, Candice Hogan
Related articles
Most popular
Horacio’s story: gaining mobility independence through innovative transportation solutions
Streamlining Financial Precision: Uber’s Advanced Settlement Accounting System
Uber, Unplugged: insights from 6 transit leaders on the future of how we move
Enabling Infinite Retention for Upsert Tables in Apache Pinot
Products
Company