Authors:
(1) Muhammad Zia Hydari, Katz Graduate School of Business, University of Pittsburgh and Corresponding author;
(2) Idris Adjerid, Pamplin College of Business;
(3) AAaron D. Striegel, Department of Computer Science and Engineering, University of Notre Dame.
Table of Links
Abstract and 1 Introduction
2. Background and 2.1. Leaderboards
3. Effect of Leaderboards on Healthful Physical Activity and 3.1. Competition
3.2. Social Influence
3.3. Moderating Effects of Prior Activity Levels and Leaderboard Size
4. Data and Model
4.1. Data
4.2. Model
5. Estimation and Robustness of the Main Effects of Leaderboards
5.2. Robustness Check for Leaderboard Initiation
5.3. Fitbit Compliance
5.4. Fitbit Attrition, Leaderboard De-Adoption, and Additional Robustness Checks
6. Heterogeneous Effect of Leaderboards
6.1. Heterogeneity by Prior Activity Levels
6.2. Interaction of Leaderboard Size, Rank, and Prior Activity Levels
6.3. Summary of Findings from Heterogeneous Effect Analysis
7. Conclusions and Discussion, Endnotes, and References
4. Data and Model
4.1. Data
We use a unique panel data set comprised of 516 undergraduates at a U.S. university from October 2015 to September 2017.[11] This data set consists of granular wearable device data and periodic survey data. With respect to wearable device data, the students were offered Fitbit Charge HR devices, which were then used to record their physical activity. We access three types of Fitbit data: (i) step count, accessed on a daily basis; (ii) leaderboard data, which captures if a focal student has a leaderboard and, if so, the seven-day average step count of other leaderboard participants for the determination of participants’ leaderboard rankings; and (iii) minute-by-minute heart rate data. Students synchronize their data with the Fitbit platform either through a dongle and a desktop application or a smartphone application. We implemented a client application that invoked the Fitbit application programming interface (API) to download the synchronized student activity data and store it locally in a secure database. The client application was a set of scripts that ran automatically every night. All study participants explicitly authorized our client application to allow access to their data via the Fitbit APIs.
Step measurements only occur if students wear their Fitbit devices regularly. We will use the term compliance to refer to the regularity with which students wear their Fitbit device. We calculate compliance from the heart rate data by assuming that a student is wearing their Fitbit during a particular minute of the day if the reported heart rate is nonzero. Students were paid $20 for maintaining at least 40% compliance and synchronizing their data regularly to Fitbit servers. Fitbit Charge HR could store up to seven days of data locally, so synchronizing beyond a seven-day interval would result in lost data and lower compliance.
Fitbit Charge HR’s step measurements, which we use as the outcome in this study, are fairly accurate. Validation studies in laboratory and natural settings have found Fitbit Charge HR’s mean absolute percent error (MAPE) for step count to be less than 10%, except for very light activity (Wahl et al. 2017, Bai et al. 2018). Bai et al. (2018) also found Fitbit Charge HR’s heart rate measure to have an MAPE of ≈ 10%, although other studies have found mixed results. Even if the MAPE for heart rate were higher, our study is not likely to be negatively impacted. We use heart rate only for measuring compliance such that any nonzero heart rate measurement is construed as the device being used by the participant during that minute.
Participants were also asked to complete an intensive survey at the start of study and were further asked to take shorter surveys in six-month waves to refresh key measures. These surveys notably provided data on demographics (gender, religious affiliation, parent’s income, etc.), psychological attributes using validated scales (personality, self-regulation), social interaction and ability (trust, anxiety, etc.), technology use (social media use, mobile app usage, etc.), and health state (body mass index, satisfaction with health, etc.). Although most students took the survey, there was some nonresponse as these surveys were not mandatory. On average, students completed three waves of survey data (approximately six months apart). We use these data in two ways. Primarily, we use relevant survey data to model the propensity for opting into a leaderboard and, in conjunction with advanced weighting approaches, construct a weighted sample that achieves covariate balance between leaderboard adopters and nonadopters. Secondarily, we use a subset of the survey data to generate controls that capture time-varying features of individuals that may relate to both leaderboard adoption and physical activity, and check the robustness of our main results.12 Table 1 provides descriptive statistics about the outcome, treatment, and some demographic variables, whereas Online Appendix Table A.1 describes the relevant portions of the survey.
