Adding Team-Based Financial Incentives to the Carrot Rewards App Increases Daily Step Count on a Population Scale: A 24-Week Matched Pairs Study

Mobile health applications (mHealth apps) targeting physical inactivity have increased in popularity yet are usually limited by low engagement. This study examined the impact of adding team-based incentives (Step Together Challenges, STCs) to an existing mHealth app (Carrot Rewards) that rewarded individual physical activity achievements. A 24-week quasi-experimental study (retrospective matched pairs design) was conducted in three Canadian provinces (pre-intervention: weeks 1–12; intervention: weeks 13–24). Participants who used Carrot Rewards and STCs (experimental group) were matched with those who used Carrot Rewards only (controls) on age, gender, province and baseline mean daily step count (± 500 steps/d). Carrot Rewards users earned individual-level incentives (worth $0.04 CAD) each day they reached a personalized daily step goal. With a single partner, STC users could earn team incentives ($0.40 CAD) for collaboratively reaching individual daily step goals 10 times in seven days (e.g., Partner A completes four goals and Partner B completes six goals in a week).


Introduction
Page 3 /19 The health bene ts of physical activity (PA) are irrefutable and yet, widespread inactivity persists. 1,2 Interventions that increase PA on population-levels are needed to help tackle this important public health issue. 3 As smartphone ownership increases (approaching 90% in the US), 4 so has the number of mobile health applications (mHealth apps) in the major app stores (over 325,000; 30% are PA apps). 5 Part of the appeal of mHealth apps is their potential to reach large populations at relatively low cost. 6 Their effectiveness, however, is often limited by low user engagement with 90% of mHealth apps being deleted within 30 days. [7][8][9] A still limited number of RCT studies suggest a dose-response relationship exists between engagement and effectiveness, with greater app use associated with larger PA improvements. [10][11][12] Conversely, waning engagement has been linked with declining PA. 10,11,[13][14][15] Recent advances in behavioural science provide a new framework from which to develop practical solutions to address this notorious mHealth app engagement problem.
Behavioural economics, a branch of economics shaped by insights from psychology, [16][17][18] has stimulated renewed interest in nancial health incentive interventions, such as rewarding people to walk more. 19 This increasingly common intervention is grounded in a behavioural economics concept called present bias which describes how individuals tend to place disproportionate emphasis on the present "cost" of a health behaviour (e.g., time) while discounting the future "bene ts" of that behaviour (e.g., increased health). 16 Behavioural economics suggests that providing timely nancial incentives for behaviours with bene ts that are largely delayed (e.g., PA) may encourage individuals to choose to engage in those behaviours rather than put them off. 16,18,20 Individual-level nancial incentives for PA (e.g., incentives for personal PA achievements) have shown promising results. A recent meta-analysis of RCTs concluded that nancial incentives increased PA in the short-term by up to 4,000 steps per day, with some evidence of long-term (six or more months) and sustained (after incentives withdrawn) effects. 21 A number of studies also suggest team-based incentives (e.g., incentives for group achievements) may be e cacious as well. 22 Compared to individual incentives, team incentives have yielded better gym attendance, more PA, and greater weight loss in RCT settings. [23][24][25] Interestingly, Patel et al. (2018) found that combining individual and team (i.e., combined) incentives was more e cacious than individual or team incentives alone. 26 The notion of "aligning the thoughts or behaviours of individuals in a group" is another pertinent behavioural economics concept called herd behaviour. 27 Herd behaviour describes how individuals are more likely to follow others in decision making instead of making independent decisions (e.g., "My friend is going for a walk, so I probably should too."). 21 The tendency for humans to want to behave in ways that are consistent with the people in their social networks may be leveraged in an mHealth context, for instance: (a) by providing feedback on peers' progress, and/or (b) with team-based incentives. Recent evidence, though, suggests that adding a social component to mHealth interventions does not necessarily translate into positive effects. 7,11,28,29 It appears that mHealth features designed to increase social connectivity among participants with no prior relationship do not work as well as those delivered among people with existing relationships (e.g., work colleagues challenge each other in an online walking challenge). 7,26,29,30 Babcock et al. (2015) compared anonymous partners to partners with an existing social connection and found PA incentives were not as effective in the anonymous group, highlighting the importance of leveraging pre-existing social connections in mHealth interventions. 24,26,30 Despite their popularity, very little is known about the effectiveness of commercial PA apps (or their design features) since few have undergone rigorous peer-reviewed evaluation. 10,31 Among the 15 studies included in the recent Petersen et al. (2019) review of PA apps, for example, only ve examined commercially available ones (e.g., Fitbit, 'Zombie, Run!') despite there being several hundred thousand in the major app stores. 31 Among these ve, little consideration was given to the role of engagement as an effect moderator despite suggestions that intervention exposure is imperative and that greater engagement usually yields larger effects. 32 The Carrot Rewards app was a top tier Canadian app (i.e. 1.3 + million downloads, 500,000 + monthly active users (MAUs) as of May 2019) that rewarded users with loyalty points redeemable for consumer goods (e.g., gas, movies) for walking more. It was developed in partnership with the Public Health Agency of Canada as part of its Multi-Sectoral Partnership Approach to Healthy Living and Chronic Disease Prevention. 33 One of the stated objectives of the initiative was to conduct rigorous evaluations of the app intervention, including the impact of new features. 12,34 In March 2018, Carrot Rewards launched a new social feature called 'Step Together Challenges' (STCs) to complement their existing walking program (called 'Steps') where individualized daily step goal achievements were rewarded with very small incentives ($0.04 CAD per day). STCs allowed users to invite a friend from their existing social network to participate in a collaborative walking challenge for bonus incentives ($0.40 CAD per week).
To enhance our understanding of mHealth interventions, and acknowledging how di cult it can be to conduct RCTs in fast paced commercial digital environments, non-RCT alternatives (e.g., quasiexperimental designs) have been recommended. 31,35,36 Quasi-experimental evaluations of "top tier" commercial apps (i.e. the top 2% of apps reporting more than 500,000 MAUs) 5 may provide particularly valuable insight into mHealth app engagement, it's role in promoting health behaviours, and how it can be improved on a population scale. The primary objective of this study, then, was to examine the impact of adding team incentives to the Carrot Rewards app on mean daily step count. An important secondary objective was to determine whether a dose-response relationship existed between app engagement (i.e. STCs completed) and mean daily step count.

Study design and sample
A 24-week retrospective pre-post matched pairs design was used to examine the effect of adding STCs to the Carrot Rewards 'Steps' walking program on mean daily step count. Participants were drawn from the existing Carrot Rewards user base which included Canadians 13 years of age or older living in the three provinces the app was launched (i.e. British Columbia (BC), Newfoundland and Labrador (NL), Ontario (ON)). All participants had to have opted-into the app's 'Steps' walking program to be included in the study. The experimental group included participants using the STC feature for the rst time between March 19 and April 16, 2018 (the rst month STC was available). Control participants were drawn from the cohort of Carrot Rewards users who had enabled the 'Steps' walking program but had not engaged in a STC during the study period. Control participants were matched with existing experimental participants on age (± 1 year), gender, province and baseline step count (± 500 steps/d, so individuals with similar PA levels would be compared). Only one control user was selected to match to experimental users if they met each of the four criteria; therefore, one control user could be matched with multiple experimental users who shared the same age, gender, province and baseline daily step count (for more detail see Additional le 1).
The pre-intervention period was de ned as the 12 weeks preceding experimental users' rst STC (Study Weeks 1-12). The intervention period included the 12 weeks following the initiation of the rst STC (Study Weeks 13-24). Participants were required to have valid pre-intervention and intervention study periods, consisting of a minimum of four weeks of daily step count data in each period-a valid week was operationally de ned as a minimum of four days with step counts between 1,000 and 40,000 inclusive, as previously done. 37 A study ow chart is provided (Additional le 2). Ethical approval for this study was provided by Western University's Human Research Ethics Board (#111252).

Individual and team incentives
Upon downloading the free, commercial Carrot Rewards app, and following a two-week baseline period, Carrot Rewards users earned individual-level incentives in the form of loyalty points each day they reached a personalized daily step goal (worth $0.04 CAD/day, the smallest possible loyalty point increment to maximize program scalability/sustainability). Goals were initially set using the two-week baseline median. While small, incentives were tied to objectively measured PA and were given nearly instantaneously with a push noti cation using smartphone technology (e.g., linking data from native smartphone accelerometer with loyalty program application programming interfaces (APIs)). To boost app engagement and PA, the 'Steps' walking program evolved with re nement of older features, as well as the introduction of new ones. For example, the algorithm used to calculate each user's daily step goal was updated to be more personalized and adaptive.
STCs were introduced in March 2018 to allow users to collaboratively pursue team-based goals with a peer of their choosing for additional rewards (i.e. a pre-existing friend they had already connected with on the app). Users participating in a STC could each earn a bonus incentive worth $0.40 CAD for together reaching 10 individual daily step goals in a seven-day period (e.g., Partner A completes four goals and Partner B completes six goals in a week; Fig. 1). Users could only participate in one STC at a time. The app allowed users to see their partner's daily step progress in real time, as well as their own, though users could not communicate about their shared progress in-app (this needed to be done through other means e.g., text messages, in-person, etc.). Over the course of a 12-month period, participants could earn a maximum of about $25.00 CAD in points, though in general users earned about $5 to $10 CAD. In addition to promoting social support, the STC feature integrated other behaviour change techniques as well including goal setting/review, self-monitoring and demonstration. For more app design detail, completed Mobile App Rating Scale (MARS self-score 4.23/5; for understanding app quality, aesthetics and functional appeal) 38 and App Behavior Change Scale (ABACUS self-score 4.5/5; for measuring potential to change behaviour) 39 are provided (Additional les 3 and 4).

Outcomes
The primary outcome was mean daily step count as measured by built-in smartphone accelerometers. In recent validation studies, the iPhone step counting feature, as well as those for Android smartphones were accurate in laboratory and eld conditions. 40-43 Self-reported demographics (i.e. age, gender, province) and number of STCs completed were also collected. Number of STCs completed was de ned as any STC that was started and nished within the seven-day window, irrespective of whether the challenge was completed successfully or not. To nish the challenge, a user simply needed to open the app to facilitate app vs. smartphone data synchronization.

Statistical analyses
Chi-square and independent t-tests were conducted to examine group equivalency on demographic measures. Controlling for pre-intervention mean daily step count, ANCOVA was performed to examine group differences in intervention period mean daily step count. Data were expressed in estimated marginal means (95% CI). A pairwise t-test was also used to examine the mean daily step count change over time (pre-intervention vs. intervention) for each group. Sensitivity analyses were performed on users with complete data sets only (valid step count data for all 24 study weeks), and users with a 1:1 control to experimental participant matching ratio. Finally, linear regression was performed to determine the relationship between the number of STCs completed and intervention period mean daily step count.

Sample characteristics
Study sample characteristics, mean baseline daily step count, and mean number of valid weeks in the pre-intervention and intervention period are in Table 1. Group differences in age, gender and province were detected likely due to the large sample size, though effect sizes were very small. As well, the mean baseline daily step count for our study sample was higher than for Carrot Rewards users in general possibly because they were a more engaged sub-group in general.

Group differences
Controlling for pre-intervention mean daily step counts, ANCOVA showed a signi cant difference in intervention mean daily step count (F(1, 61 167) = 1,515.97, p < 0.0001), favouring the experimental over the control group with a small effect (η p 2 = 0.024; Table 2). An estimated marginal means difference of 537 steps per day favoured the experimental group.  Table 3). The experimental group showed an increase of 504 mean steps per day more than the control group from the pre-intervention to intervention periods.  Fig. 2 and Additional le 7). Although the increase in intervention mean steps per day appears somewhat exponential when inspecting increases in step count for users completing 15 and 16 STCs in particular, these means actually represent a very small proportion of users (n = 270 vs. n = 39,355 experimental users in total; see Additional le 7).

Main ndings
In this large quasi-experimental evaluation of the Carrot Rewards app we found that adding team-based incentives increased mean daily step count. Speci cally, our experimental group had an increase of 1,143 steps whereas the control group saw an increase of 606 steps during the study period -a 537 adjusted mean daily step count difference, or 3,759 additional steps per week (roughly equivalent to about 40 minutes of walking). This may be attributed to the application of 'herd behaviour', the behavioural economics principle describing the tendency for individuals to follow others' behaviours instead of making independent decisions. Carrot Rewards exploited this predictable human tendency by providing real-time feedback on peer progress, as well as by rewarding users only if both achieved at least a few daily goals throughout the Challenge. In doing so, we speculate that STCs may have also served to increase feelings of social connectedness within the small teams, which according to self-determination theory (a global theory of human motivation) promotes quality health behaviour change. 47 A dose-response relationship was also observed with app engagement positively associated with mean daily step count. Each new STC corresponded to roughly 200 more steps per day.

Implications
The clinical implications of daily step count increases of these magnitudes are important. For instance, an increase of 1,000 steps per day has been associated with signi cant weight loss in adults and better glycemic control (i.e. lower A1C) in individuals living with type 2 diabetes. 15,48 Higher step counts in general are also associated with improved mood and overall health ratings, and are inversely related to systolic blood pressure. 15,49 From a public health perspective, a 1% reduction in the number of Canadians classi ed as "physically inactive" (fewer than 5,000 daily steps) would yield annual healthcare savings of $2.1 billion CAD. 50 With nearly half of the Carrot Rewards users in general accumulating less than 5,000 steps per day, a mere 500 to 1,000 step increase from baseline values could have broad implications. 12 In fact, a recent 12 month analysis of the Carrot Rewards app suggests approximately 100,000 Canadians (or 0.3% of the Canadian population) moved up from the "physically inactive" category to the "moderately active" (5,000 to 7,500 steps per day). 34

Similar studies
Findings from our quasi-experimental examination in a real-world commercial context compliment those from traditional RCT studies examining the impact of team-based incentives on PA or weight loss, in digital and non-digital settings. 15, 24-26 Babcock et al. (2015) found that the number of gym visits was 9-17% higher in the team incentive compared to the individual incentive conditions. In particular, team incentives for teams where members knew each other were more effective than those for anonymous teams. 24 Patel et al. (2016) examined a PA intervention delivered through a research-based smartphone app, combining the social aspects of team incentives with individual-level rewards, much like the incentive scheme used in the present study. 26 In comparing a control condition to individual, team and combined (individual plus team) incentives groups,  found participants in the combined incentives group had a signi cantly higher mean daily step counts compared to controls (1446 daily step count group difference), whereas the team and individual incentive groups did not outperform controls. 26 Smith-McLallen et al., (2017) also compared a digitally delivered standard walking program to an enhanced program including incentives, feedback and competitive challenges and found that the enhanced group improved their mean daily step count by 726 more than the standard program group over nine months. 15 Finally, a year-long evaluation of the standard Carrot Rewards 'Steps' walking program (i.e. before STCs were introduced) found an average increase of 448 and 884 steps per day from baseline for 'regular' and 'committed' users (i.e. users engaging with the app on 26 to 51 out of 52 weeks, or all 52 weeks, respectively). 12 It is interesting to note that these pre-post daily step count differences are similar in magnitude to the ones reported here, and that in both studies greater engagement yielded larger effects. This aligns with a growing but still limited (to a small number of RCTs) evidence base suggesting that app exposure is paramount and that greater engagement may produce greater effects. [10][11][12]32 Limitations This study was not without limitations. The quasi-experimental design, for instance, makes it di cult to conclude with con dence that the Carrot Rewards STC feature caused an increase in PA. To increase internal validity in this real-world public health intervention context, we matched experimental participants with similar controls, used three main analytic approaches to address our primary objective and conducted two sensitivity analyses. Self-selection bias, especially with regard to engagement (i.e. controls did not engage in a STC during study period, thus may have been less engaged a priori) may have confounded our results despite our best efforts to match experimental users with similar controls. Although we were limited in our ability to match users to a few demographic variables, we were able to match based on baseline daily step count and control for baseline daily step count discrepancies in our analyses. Changing seasons (the study started in the typically cold Canadian Winter and ended in the warmer Spring/Summer months) may also have impacted our results. It is reasonable to assume though that this potentially confounding seasonality effect in uenced PA step count similarly in both study groups. In addition, smartphone wear time may have differed between study groups. Device adherence was uniformly high before and during the intervention, as seen with similar numbers of valid weeks in the pre-intervention and intervention periods. This was likely optimized with the presence of the individuallevel PA incentives, possibly shrinking the gap between measured and actual steps (i.e. assessment error) for both groups. 40 Lastly, our study did not evaluate the long-term (six or more months the theoretical threshold of behaviour maintenance) effects of adding team incentives to an existing walking program rewarding users with individual incentives. 51 Future research should evaluate whether combined incentives drive PA improvements over the long-term.

Conclusion
Given the persisting physical inactivity pandemic, there is an urgent need for scalable and effective digital PA interventions. 52 Examinations of real-world effectiveness have been repeatedly called for in the literature highlighting the important contributions made here. We have shown that incorporating concepts from behavioural economics (e.g., herd behaviour) in the design of mHealth app features in a way that leverages pre-existing social networks has the potential to improve user engagement and PA behaviours. Speci cally, participants using the Carrot Rewards app with team-based incentives accumulated more steps per day during a 12-week intervention period compared to matched controls. The more the STC feature was used, the more participants walked. While the effect of STCs on long-term behaviour change is not clear, their role in increasing feelings of social relatedness (even in a digital environment), and thus the potential for sustained change, would be an interesting line of future inquiry. Taken together, much can be learned from this large-scale evaluation of a top-tier commercial PA app. In particular, more high quality quasi-experimental designs are needed to examine real-world effectiveness in a fast paced mHealth context that does not necessarily lend itself to more carefully controlled RCTs that prioritize internal over external validity.

Declarations
Ethics approval and consent to participate Ethical approval for this study was provided by Western University's Human Research Ethics Board (#111252). This study involved the secondary use of de-identi ed data. There was no consent for this secondary data analysis. However, app users were informed of and had to accept the app's privacy policy describing how de-identi ed data may be used for reporting purposes and presented in aggregate.

Consent for publication
Not applicable.
Availability of data and materials The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Chronic Disease Prevention, as well as the Government of Ontario. The funding bodies did not contribute to any aspect of the study.

Authors' contributions
All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by EP, HP, CH, and MM. The rst draft of the manuscript was written by EP and MM and all authors commented on previous versions of the manuscript. All authors read and approved the submitted manuscript and have agreed to be personally accountable for their contribution.