Source-linked AI summary
Online Actions with Offline Impact: How Online Social Networks Influence Online and Offline User Behavior
Tim Althoff, Pranav Jindal, Jure Leskovec
TL;DR
The paper asks whether social-network features causally change user engagement, retention, and offline behavior, beyond attracting users who were already more active. It analyzes five years of activity-tracking data using natural experiments and related observational designs, finding increases in online and offline activity and separating social influence from increased motivation.
Problem
Existing evidence offers limited understanding of whether social-network features influence engagement, retention, and offline behavior rather than merely selecting more active users.
Method
The study analyzes activity-tracking data and uses delayed edge formation as a natural experiment alongside difference-in-difference and matching-based observational studies.
Results
30% higher in-app activity, 17% lower one-year dropout, and 7% higher physical activity follow social-network joining, with 55% of the observed effect attributed to social influence.
Takeaways & Limitations
Online social-network connections are associated with measurable changes in both application behavior and real-world physical activity, and the paper predicts which users are most influenced.
Takeaways & Limitations
The seven-day window for individual edge analyses estimates only short-term effects because longer-term estimates could be confounded by other edges in observational data.
Abstract
from arXiv · showhide
Many of today's most widely used computing applications utilize social networking features and allow users to connect, follow each other, share content, and comment on others' posts. However, despite the widespread adoption of these features, there is little understanding of the consequences that social networking has on user retention, engagement, and online as well as offline behavior. Here, we study how social networks influence user behavior in a physical activity tracking application. We analyze 791 million online and offline actions of 6 million users over the course of 5 years, and show that social networking leads to a significant increase in users' online as well as offline activities. Specifically, we establish a causal effect of how social networks influence user behavior. We show that the creation of new social connections increases user online in-application activity by 30%, user retention by 17%, and user offline real-world physical activity by 7% (about 400 steps per day). By exploiting a natural experiment we distinguish the effect of social influence of new social connections from the simultaneous increase in user's motivation to use the app and take more steps. We show that social influence accounts for 55% of the observed changes in user behavior, while the remaining 45% can be explained by the user's increased motivation to use the app. Further, we show that subsequent, individual edge formations in the social network lead to significant increases in daily steps. These effects diminish with each additional edge and vary based on edge attributes and user demographics. Finally, we utilize these insights to develop a model that accurately predicts which users will be most influenced by the creation of new social network connections.
1. INTRODUCTION
The paper addresses limited evidence about whether online social networks causally change users’ engagement, retention, and offline behavior. Using activity-tracking data and natural experiments, it finds positive effects online and offline while separating social influence from users’ motivation.
- Research gap: Online social networking’s effects on user engagement, retention, and real-world behavior remain poorly understood, including whether observed differences reflect selection or influence.Existing work largely examined online outcomes, while offline effects and counterfactual behavior were harder to measure.
- Approach: The study analyzes online in-app engagement and offline physical activity in a large smartphone activity-tracking dataset.The dataset covers 6 million users over 5 years, with activity posts measuring online engagement and passively collected steps measuring offline activity.
- Causal decomposition: 55% of the observed average behavioral effect is attributed to social influence, while 45% is attributed to increased user motivation.Delayed social-network edge formation provides the natural experiment used to distinguish these components.
- Findings: 30% higher in-app activity, 17% lower one-year dropout, and 7% higher physical activity follow social-network joining relative to a matched control group.The effects are positive for both online and offline behavior and persist for several months, although they diminish over time.
- Contribution: The study combines natural experiments, difference-in-difference models, and matching-based observational studies to separate selection effects from causal social-network effects.Its methods are presented as applicable beyond physical activity and health behaviors to other online and offline activities.
2. DATASET DESCRIPTION
The dataset combines self-reported in-app activity with passively measured steps and social-network interactions. Its delayed introduction and delayed friendship acceptance enable comparisons of online engagement and offline physical activity while helping separate influence from selection.
- Dataset: 6 million individuals across more than 100 countries were observed using the Argus smartphone app over 5 years.The observation period ran from January 2011 to January 2016.
- Measures: 631 million self-reported activity posts and 160 million days of objectively measured steps tracking were recorded.Posts include activities such as running, walking, sleep, heart rate, yoga, cycling, and weight.
- Measures: In-app posts measure online activity, whereas accelerometer-defined steps measure offline physical activity.The paper uses “activity” for offline physical activity and “posts” for online in-application activity.
- Social network: The social-network feature added friend and follower connections that generated activity notifications and a timeline-like feed.Friend connections required receiver approval, while follower connections did not.
- Identification: The network’s introduction after two years of pre-social observation and delayed friendship acceptance support causal analysis of user behavior.These properties help disentangle social-network influence from users’ pre-existing behavior and selection effects.
3. DISTINGUISHING INTRINSIC MOTIVATION FROM SOCIAL INFLUENCE
The paper separates intrinsic motivation from social influence by comparing users whose friendship requests are accepted immediately with those accepted after more than seven days. This natural experiment shows that friendship formation contributes to increased physical activity beyond the motivation to send a request.
- Motivation versus influence: Users may become more active before creating an edge, so post-connection increases alone cannot distinguish social influence from intrinsic motivation.The paper notes that users adding edges may already be motivated to use the app and increase activity.
- Natural experiment: The natural experiment compares immediate and delayed request acceptance among users who already sent friend requests, varying when the social connection becomes active.Delayed acceptance is treated as occurring more than seven days after the request, whereas direct acceptance occurs within one day.
- Natural experiment: The framework uses seven-day pre- and post-request behavior to estimate intrinsic motivation from delayed accepted edges and total change from directly accepted edges.Delayed edges isolate motivation because the connection is not active during the observation window; direct edges include both motivation and influence.
- Validation: The delayed and direct acceptance groups are treated as comparable because the paper validates balance across relevant request and user covariates.The study defines covariate balance using an absolute standardized mean difference below 0.25 and reports that the mechanism creates well-balanced groups.
- Results: 148 daily steps reflect intrinsic motivation, while directly accepted requests produce 328 additional daily steps, leaving 180 daily steps attributable to social influence.The influence estimate is statistically significant, with a 95% confidence interval of 74–236 steps.
- Results: 55% of total behavior change is attributed to social influence and 45% to elevated intrinsic motivation, establishing a causal effect of a social connection on offline behavior.The proposed mechanism is exposure to friends’ activity and status updates through notifications and the personalized activity feed.
4. HOW JOINING A SOCIAL NETWORK IMPACTS USER BEHAVIOR
Joining the social network increased users’ physical activity, app engagement, and retention relative to matched controls, with effects that persisted for months but diminished over time.
- User Physical Activity Level: The matched difference-in-differences design compared joining users with similar non-joining controls using pre-treatment activity and signup timing.Joining was defined as creating the first accepted friend or follow connection.
- User Physical Activity Level: 406 additional daily steps marked a 7% physical-activity increase after users joined the social network, compared with matched controls.Treatment users averaged 6,454 steps versus 6,048 for controls, and the effect persisted for over three months.
- User Physical Activity Level: The observed activity increase was robust to less stringent observation constraints, which produced a similar 382 daily-step increase.The constrained experiment included 6,076 of 211,383 social-network users.
- Online User Engagement and Retention: 17% higher one-year retention followed social-network joining: 28.0% of treatment users remained active versus 24.0% of controls.The retention difference became clearly noticeable after about three weeks and persisted through one year.
- Online User Engagement and Retention: 30% more daily posts followed social-network joining, with the increase lasting over four months before becoming statistically indistinguishable after one year.Among active users, treatment users averaged 3.95 versus 3.05 posts per day immediately after joining.
5. THE EFFECT OF INDIVIDUAL EDGE FORMATIONS
Individual social-network edges produced temporary increases in daily steps, with effects decreasing across successive connections and varying by edge type, initiator, and user characteristics.
- Method and Scope: The seven-day before-and-after design estimates short-term effects because longer windows could be confounded by other edges and time-varying activity.Edges with request and acceptance more than one day apart were filtered out to reduce short-term-window bias.
- Edge Effects: Up to 1,236 additional daily steps followed the first edge, while increases became smaller with each additional edge.The analysis examined users’ first five edges and compared activity before and after each creation event.
- Edge Effects: Friend edges produced larger effects than follower edges, with sender effects exceeding receiver effects, especially for later edges.For the first and fifth sender edges, friend effects were 1,236 versus 381 and 613 versus 283 daily steps; fifth-edge sender effects were 613 versus 320 for receivers.
- Explaining Decreasing Influence: Activity oscillations before and after edge creation help explain why the estimated influence decreases across successive edges.Post-edge activity varies less while pre-edge activity rises and falls across edge numbers.
- User Susceptibility: Behavior changes were larger for older users, users with higher BMI, and users taking more steps before edge creation.The largest reported age-group change was 935 steps for 30–40-year-old edge senders.
6. PREDICTING BEHAVIOR CHANGE
The paper predicts which users increase or decrease activity after new social-network connections, using edge, demographic, and prior-activity features. Combining these features achieves strong predictive performance.
- Prediction task: The task predicts whether users increase or decrease activity during the 7 days after adding a friend or follower.The dataset contains 432,133 measurable edge creations, with users increasing activity 55.4% of the time.
- Predictive features: AUC=0.715 for pre-edge physical activity, making it the strongest single predictor.Demographics reach AUC=0.685, while edge type, initiator, and edge number reach AUC=0.574.
- Predictive performance: AUC=0.785 results when combining all feature groups.The full model combines prior response, edge characteristics, demographics, and physical activity level.
- Predictive heterogeneity: Behavior change after friend edges is less predictable than after follow edges, while sender responses are more predictable than receiver responses.Demographics and prior activity are particularly predictive for follow edges.
- Conclusion: The proposed models predict which users increase activity after edge creation with AUC above 78%.Prediction varies substantially over time even within the same user.
7. RELATED WORK
Prior work examined social-network structure and online influence, but offline behaviors were often studied through online proxies. This paper extends that literature by directly measuring physical activity at scale and around new connections.
- Prior research: Prior research studied network structure, growth, communities, information diffusion, influence maximization, social capital, and social influence.The cited literature covers fundamental properties and influence processes in online social networks.
- Online and offline outcomes: Existing social-influence studies mainly measured online outcomes such as app adoption, content downloads, voting, and resharing.Important offline outcomes include physical activity, food intake, mental health, obesity, and smoking.
- Paper contribution: The paper studies offline physical activity, uses network introduction and growth to identify changes near connection formation, and analyzes 6 million users with 631 million activity posts.Physical activity is presented as behavior relevant to human health.
8. CONCLUSION
The paper presents causal and observational analyses of how social networks affect online engagement, retention, and offline physical activity. It also identifies predictable variation in responses and motivates future intervention research.
- Conclusion: Natural experiments and observational studies examine how social-network features influence online and offline behavior in a large activity-tracking dataset.The study focuses on engagement, retention, and real-world physical activity.
- Conclusion: Social influence accounts for 55% of the observed effects after delayed friendship formation.The natural experiment separates social influence from simultaneous intrinsic motivation.
- Conclusion: User susceptibility to behavior change can be predicted with significant accuracy after new social connections.The paper proposes future work on incentives and contextual interventions for healthy behavior change.
10. APPENDIX
Appendix analyses reinforce the causal interpretation of social-network effects and examine how activity and tracking responses vary across connection types, initiators, and successive edges.
- Network timing: 84% of social-network edges are created within 7 days of the previous edge.The distributions include bidirectional friend and unidirectional follower connections and are shown as unnormalized CCDFs.
- Joining experiment: 382 daily steps is the treatment increase under a weaker four-week post-joining tracking requirement.The effect is practically identical to earlier estimates.
- Joining experiment: Joining users show a significant activity boost, whereas matched control users show no difference.The treatment effect is practically identical to the effect in Figure 4.
- Intrinsic motivation and social influence: Pending edge requests also produce significant activity increases, supporting the causal analysis based on delayed accepted requests.The repeated analysis yields very similar effect-size estimates.
- Intrinsic motivation and social influence: 127 daily steps follow matched pending requests versus 448 after accepted requests.The estimates imply 28% from elevated intrinsic motivation and 72% attributed to social influence.
- Individual edge formations: Individual edge effects decrease with edge number, are larger for friend than follower edges, and are larger for senders than receivers.The same broad pattern appears for both activity levels and steps tracking, although follower edges can reduce tracking propensity.
- Steps tracking: Follower-edge senders and receivers take more steps but are slightly less likely to track steps on a given day.Friend-edge tracking baselines exceed follower-edge baselines because the edge types tend to occur at different points in users’ lifetimes.