1 Introduction

Peer influence and social interaction have been found to have positive health-related effects, such as helping people lose weight and increasing physical activities [1,2,3,4,5]. Recently, increased attention has been paid to promoting health habits through online social interaction in online fitness communities (e.g., running or cycling online groups) [6, 7]. Studies show that social interaction features such as cooperation and competition provide participants with a group of peers and help motivate them to reach their fitness or health goals [2, 8]. In this work, we study the relationship between fitness behaviors and online social behavior via a subscription service where users can post and follow peers’ activity feeds.

Traditionally, studies on physical activity and social interaction have relied on information that is self-reported (e.g., diary studies) or measured via (expensive) wearable sensors, which have been limited in terms of scale, granularity of activity and duration of observation period [9, 10]. Recent developments in smart phone GPS tracking and accessibility has provided for an increasing large adoption of mobile devices which track everyday physical activities. These resultant behavioral trace data allow for precise measurement of individual’s activities and online social action in scale and cost that traditional survey-based methods for collecting data on physical activities cannot match, and provide a good alternative to classic sensor studies [11, 12]. Our work leverages these new data sources (e.g., behavioral trace data) to explore the relationship between fitness behaviors and online social interaction over time. We employ this novel data set along with event history methods to understand the relationship between online social interaction and activity levels within one of these app-based activity communities. Specifically, we focus on the large app-based community known as Strava where users have covered over 12 billion miles worldwide [13]. In this work we focus on two comparable major metropolitan areas in the US.

This work analyzes the dynamics of online fitness behaviors and network subscription as well as the relationship between them. We ask the following research questions: (1) how do users’ activity levels of exercise change over time? (2) how do exercise activeness and network subscription vary among users? (3) how subscription magnitude is associated with activity occurrence? We find clear seasonal patterns of users’ fitness behaviors and discuss the implications for fitness application designs and health prevention. We also show that paid-plan users exercise more actively and attract more followers than free-plan users. Last, our analysis shows a positive relationship between social subscription and physical activities, supporting the claim –“stay connected, keep motivated”.

The remainder of this paper is organized as follows. We start by reviewing existing studies on online fitness communities and roles of social interaction in physical activities in online fitness communities. Next, we describe the behavioral trace data collected from Strava for this work. We then describe the methodology for this work, describing our analysis on activity levels and online social interaction via network subscription, as well as the methodology used to model physical activities given user characteristics and network subscription. We finally discuss and summarize our findings.

2 Related Work

With advances in pervasive technologies, activity tracking applications and online fitness communities such as Strava, RunKeeper, MapMyRun, etc. are attracting more and more users around the world. Online fitness communities usually have features of both activity logging and social networking [9]. Online fitness communities log activity-related data and help users analyze their performance. Online fitness communities also serve as an activity-based network that connect users and provide users with a series of social interactions features meant to encourage behavioral change and healthy life-style promotion. For example, on Strava, users can follow both recreational and professional athletes, view their activities and interact with them by making comments, giving kudos, etc.

Behavioral trace data archived by online fitness communities record large amount of data that is generated by users throughout their physical activity (e.g., running/cycling) and through their social interaction online. One can compare this data to the more often collected self-report surveys which acquire individual’s perception of their workout routine and social engagement, rather than the behavioral trace data which records their exact physical activity, timing, distance, etc. and precise social interaction (e.g., running in groups or liking someones run activity). This data source also shows advantages in terms of scale, granularity and observation duration against data collected via expensive sensors [10]. Thus, it provides researchers with new opportunities for understanding the relationship between fitness behavior and social interaction.

Online fitness communities are increasingly attracting researchers from a variety of disciplines. Some studies revolve around incentives and interventions. For example, studying potential of health devices and applications for health-related behavioral change [2, 14, 15]. Another major body of work focuses on the technical potential of wearable sensors and human-computer interaction aspects of these technologies. For example, examining specific features in designs of fitness applications [16,17,18]. However, there are few studies on how online social network structure influence activity engagement, leaving a gap in our current understanding of the social dynamics in these settings.

Recent work have studied the relationship between social interaction and physical activity using fitness applications. Social interaction in online fitness communities may include cooperation and competition and sharing physical activities. Studies by [2, 19,20,21] suggest that social interaction is essential to motivate users to perform physical activities. As mentioned above, online fitness communities enable users to connect and interact with a group of peers online. In the case of Strava, users can follow other athletes, view their profile and activities, and receive activity feed once their peer post a completed activity. They may also compare workout and network-based stats with each other, “like” others’ posted activities and make comments under posts. The work by [9] examines how Strava users’ social motives predict perceived usefulness of the platform based on survey responses collected from 394 Strava users. Three aspects of social motives are considered: staying informed on friends’ activities, viewing progress made by friends, and receiving support from others via kudos and comments. The results show that social motives influence habitual Strava use directly, and when compared to novice users social motives are more important for experienced users.

3 Data

This work utilizes behavioral trace data collected from the activity-based network, Strava. Strava sits at the intersection of social media and activity tracking applications and is known colloquially as the “Facebook” of activity-based apps; users have the ability to not only track and log their activities, but also connect to and interact with a group of peers online. The platform continues to grow in popularity among cyclists and runners in recent years around the world.

Table 1. Data summary
Full size table

In this work, we study 2,605,147 cycling and running activities from 11,245 anonymized users from an activity-based tracking platform known as Strava. Our data includes but is not limit to the following three main components: (1) user profile information, including gender, date of birth, location, user account status (e.g. free or paid plan), sign-up date, etc.; (2) logs of posted activity for users in the sample over time, including activity timestamp, location, type, performance stats, etc.; and (3) social subscription: who followed whom and the corresponding timestamp.

We focus on two major metro areas within the continental US, which have a large active set of Strava users. We have chosen to focus on San Francisco City/County, CA which is where Strava started and continues to be headquartered, and Boston/Suffolk County, MA. Both cities represent similar size metropolitan areas within the US. Boston metro is ranked 10th with about 4.8 million residents, and San Francisco metro is ranked 11th with about 4.7 million residents. Thus, these two areas represent comparable cities on the coasts within the US context, but with wildly different weather patterns. For example, Boston has an average high of 36\(^\circ \) Fahrenheit in January, and San Francisco has an average high of 58\(^\circ \) Fahrenheit in same month. Our analysis examines and compares fitness behaviors and social subscription behaviors of users from these two counties. The key difference in these two areas is their weather patterns, and so we expect differences in community activities to stem from these seasonal differences.

Table 1 presents the summary of our data in terms of user group and activities. Male users are proportionally greater than female users in San Francisco and Suffolk County, and gender proportion is relatively similar in both locations. There exist more cyclists than runners in both groups – this is to be expected as Strava was originally developed by a group of cyclists. San Francisco users have 69.5% of self-reported cyclists and 30.5% of runners, while the proportion of cyclists and runners of Suffolk County users are 62.3% and 37.7%, respectively. Lastly, we observe that the major age group of users is between 18 and 35 for both locations, and the second majority age group is between 36–49.

4 Methods

First, we examine seasonal patterns of workouts done by users in our sample. Specifically, we look at total numbers of posted activities across all users ranging from September, 2009 to April, 2017. Theoretically, seasonality of physical activities may vary by activity type and location. Hence, this analysis focuses on the two major activity types - cycling and running, which accounts for 53% and 40% of total activities, respectively. Moreover, we do so for users from San Francisco County and Suffolk County separately in order to have a simple control setting for weather in the analysis.

As we are interested in the relationship between fitness behaviors and online social interaction over time, we examine users’ activity level and network subscription by different user groups. We begin with a simple metric activity level measuring to what extent a user engages Strava to track/log exercise actively:

$$\text {Activity level}=\frac{\#\, \text {of activities by a user}}{\# \, \text {of days a user in Strava service}}$$

We focus on how users’ activity level and number of their followers varies by gender, age group, training plan enrolled (i.e. free or paid plan) and athlete type (i.e. cyclist or runner). For gender differences in physical activity, prior work argues that physical inactivity is more prevalent among female. Indeed, male users are proportionally greater than female users in Strava. However, we are also interested in exploring gender differences in activeness in the online fitness community. Next, we observe a disproportionate distribution of users age in our sample data (i.e., the majority of Strava users aged between 18 and 49). Therefore, we want to examine how actively each user age group engages these activities. Further, we are interested in finding whether users from paid plan workout more than free-plan users. We are also interested understanding how workout activity differ by athlete type.

Last, we model user workout frequency over time with a particular interest in examining whether a user who is followed by more peers tend to exercise more. In this work, we perform an event history analysis in order to characterize the occurrence of the repeated events - physical activities given a time-dependent variable - follower count along with time-independent variables - gender, age, training plan enrolled and athlete type (i.e. cyclist or runner). To control for seasonality of location we include a dummy variable to indicate whether a user comes from San Francisco County (indicator = 0) or Suffolk County (indicator = 1). We use the popular cox proportional hazards model where the unique effect of a unit increase in a covariate is multiplicative with respect to the hazard rate. In this model, the dependent variable is h(t) - the hazard of an event at time t. Roughly speaking, h(t) can be interpreted as the instantaneous probability that an event will occur at time t. The hazard function follows the form:

$$\lambda (t|X_i) = \lambda _0(t)\text {exp}(\beta _1X_{i1}+\dots +\beta _pX_{ip}+\beta _{p+1}Z_i(t))$$

where \(X_{ip},...,X_{ip}\) are the \(1^{st}\) to the \(p^{th}\) time-independent covariate and \(Z_i(t)\) is the time-dependent covariate at time t for the observation i. Users who churned the Strava service and hence deleted the account are censored in this analysis.

5 Results

5.1 Seasonality of Physical Activity

Figure 1 shows a time series of activity frequency in Strava ranging from 2009 to 2017. For both runs and rides, we first observe a clear seasonal pattern for both locations; number of activities done per year usually peak at late spring or summer and then drop to the local lowest point at winter. When comparing between the two locations, we observe Suffolk users’ activity behaviors are more greatly influenced by the time of year. Unsurprisingly, this indicates that weather is indeed a great factor for exercise activeness especially for outdoor exercise. Further, it is interesting to notice that cycling is more subject to seasonal change with larger difference in frequency during different seasons when compared to running. Moreover, it is important to note an overall increasing trend for both running and cycling, which indicates an increasing popularity of exercise within the Strava app over time. This analysis suggests that it is important to consider temporal dimension in the analysis and modeling of user fitness behaviors.

Fig. 1.
figure 1

Seasonal pattern of physical activity frequency.

Full size image

5.2 Exercise Activeness and Network Subscription

Figures 2 and 3 show how users’ activity level and follower count vary in terms of gender, training plan, athlete type and age. We compare the two locations with the entire set to see if there is any spatial difference. Overall, we find that distributions of activity level and follower count follow similar patterns for both locations.

One significant difference is between users with free plan and users with paid plan. We observe that paid-plan users tend to exercise more actively and have more followers. It is worth pointing out that on average female paid-plan users tend to have more followers than their male paid-plan counterparts, while we do not see significant gender difference in follower count in free-plan user group.

We find that the activity level of cyclists is on average higher than runners, and male cyclists and runners tend to be more active than their female counterpart. Future study might examine users and their activities from other areas and countries for a generalization of these findings.

Last, we observe that activity level differs by age group and gender. Men who are over 50 tend to be more active than men in other age groups, and the pattern seems consistent across two locations. However, for female users in Suffolk County differences in activity level by age group is larger; women who are between 36 and 49 tend to be more active than women who are younger or older. For gender differences, we find that men tend to be exercise more than women across all age groups. While mid-aged or older-aged users tend to exercise more actively, the results show that younger users tend to receive more follower counts. However, note that the age group of 0–17 contains only a small number of observations and may not be representative of the larger sub-population.

Fig. 2.
figure 2

Distributions of activity level of workout by gender, training plan, athlete type and age

Full size image
Fig. 3.
figure 3

Distributions of user follower count by gender, training plan, athlete type and age

Full size image

5.3 Modeling Activity Occurrence

Overall, in the Figs. 2 and 3 we observe that activity level does not always align with social subscription level. For example, users who are over 50 exercise quite actively but have far fewer followers compared to other age groups. This suggests that there might exist a more complex relationship between activity level and online social interaction. Therefore, we move on to discuss the results of modeling users’ physical activities over time, given variables of interest that are explored in our previous analysis.

Table 2 presents the results of this event history analysis. To recall, the variables in the model are follower count which is time dependent as well as age, gender, athlete type and location which are time independent. Time independent variables except for age are categorical variables in this analysis. The results of cox model show that gender of male, runner and paid plan are significantly related to increase in possibility of activity occurrence, supporting the findings from our previous exploratory data analysis.

Moreover, our cox model reveals that every one unit increase in follower count results in an increased 2.1% probability of an exercise occurrence \((P-value<\)0.001). For instance, the model suggests that a user who has 50 followers has an approximately 100% increase in the probability of performing a physical activity. Therefore, even though 2.1% appears to be a modest boost for activity occurrence, this could be a relatively large boost given that the followers is a count variable. This suggests that a greater follower count that a user has is thus correlated to a higher probability that the user exercises. Strava users who have more followers are experiencing more exposure of their posted activities to their followers and likely receiving more social feedback (i.e. comments, kudos) from them.

Table 2. Modeling activities using cox hazard models
Full size table

6 Discussion

In online fitness community such as Strava, a rich set of social interaction features starts with following other athletes and hence building users’ activity-based social network. Therefore, our work aims to analyze the dynamics of online fitness behaviors and network subscription as well as the relationship between them. Specifically, we ask how users’ activity levels of exercise change over time, how exercise activeness and network subscription vary among users, and how subscription magnitude is associated with activity occurrence. We utilize behavioral trace data from the online community Strava to answer these research questions. Data focus on two major U.S. metro areas that have a large number of active Strava users - San Francisco, CA and Suffolk, MA; data contain profile information of sampled users, user activity logs as well as network subscription logs.

We find that users’ fitness behaviors display clear seasonal patterns. In general, late spring and summer are more attractive seasons for rides and runs, whereas winter appears to be less attractive. We also observe that compared to running, cycling is more sensitive to seasonality. Although strong seasonal patterns of physical activities (especially outdoors activities) are unsurprising in human behavior, results demonstrate that individual physical inactivity is likely to be aligned with seasonality in a systematic way. For designers of fitness applications, an implication of the analysis may be to take into account both individual exercise preference and optimal seasons for certain activity types. For example, Strava supports a great variety of activity type, but current practices in using and advertising the application are limited to outdoors activities, (mostly cycling and running).

Paid-plan users exercise more actively and attract more followers than free-plan users. We also observe significant gender differences in follower counts among paid-plan users; while activity levels of paid-plan users do not vary much by gender, active female users tend to have more followers than male users do. However, reasons behind the findings require a further investigation. It could be that active female users tend to connect to more users and hence receive more followers in return. Future work might also examine gender differences in the way that networks are structured in terms of symmetric and asymmetric ties for free-plan and paid-plan users.

The results demonstrate a positive relationship between social subscription and activity occurrence. Modeling individual activity occurrence using event history analysis enables us to quantify the “power” of gaining one follower for users to exercise more. In this work, we focus on characteristics of egos (eg. gender, training plan, age group, etc.). One analysis that may be worth to perform next is to take into account nodal covariates for both egos and alters. For example, users who have many active followers versus users who have many inactive followers; or users who are mostly followed by the same gender versus users who are mostly followed by users whose gender differs from them. Also, built upon the findings of this work, future work may further compare one-way connections with mutual connections to see which type of connections has a stronger association with user activity levels of exercise.

7 Conclusion

Our work analyzes the dynamics of online fitness behaviors and network subscription as well as the relationship between them. We utilize a large-scale behavioral trace data set from an online fitness community Strava. Our results indicate that fitness activity levels not only has seasonal variations, but also vary by user group. The results of event history analysis suggest that individual activity levels are significantly associated with how well users are connected in an online fitness community. The implications of these results for studies on network-based health and design of application features for health promotion are also discussed.