Body weight abnormalities are estimated to cost 21% ($190.2 billion) of annual medical spending in the United States [1, 2]. Even if a few policies and interventions [3–5] have been successful in slowing down the prevalence of obesity in recent years [6] the overall picture is still daunting, as a recent study reported that the adult obesity rate in the U.S. increased by more than two percent from 25.5% to 27.7% in just six years [7]. It is well known that an increased awareness of factors affecting weight helps individuals better manage their weight. Simple weight monitoring has been shown to be linked with weight loss both in observational studies [8–10], and randomized control trials [11–13]. In addition, studies have shown that monitoring eating habits [14, 15] and exercise levels [16] are also conducive of weight loss.

The extent of the association between monitoring frequency and amount of weight loss has also been investigated, although less thoroughly. VanWormer [17] ranked 12 weighing frequency studies in terms of methodological quality finding only one (Wing et al. [18]) to be worth an A rating. Burke et al. [19] found the a positive correlation between weight loss and the frequency at which individual monitor weight-related activities. Studies linking monitoring adherence and weight loss usually consider settings in which weight frequency is self-reported retrospectively [20] or it is summarized to a single frequency value that fails to capture temporal variation in self-weighing activity [21]. An exception can be found in the very recent study by Helander et al. [22] that investigates how the temporal variation in frequency of weight monitoring affects changes in body weight, showing that breaks in weight self-monitoring longer than 6 days are associated with weight gain. Their study, however, considered a very small (N = 40) and biased (all Scandinavian, participating in a health promotion program) population and does not analyze tracked activities other than weight monitoring. In the conclusion of their paper, Helander et al. note that it is unclear whether and to what extent their findings generalize to a larger population independently recording multiple activities outside the context of a predetermined program, and envisioned that their own analysis could be repeated on data sourced from a large on-line population of individuals reporting their weight using a connected Wi-Fi scale.

In this paper we analyze millions of recordings automatically collected through digital health trackers for monitoring weight, exercise, and food intake, and surface temporal patterns of interaction with trackers that are associated with weight change over time. The population studied is orders of magnitude larger than what previous studies have considered, consisting of several thousands of individuals observed over more than two years. The inferences we present are derived from real-world usage data and ultimately link observed tracking behavior in a large-scale and distributed setting to weight loss. In addition, the population studied consists of individuals who have not expressed an explicit interest in losing or controlling their weight. The temporal links we uncover between tracking behavior and weight change are on the timescale of days, allowing for prompt intervention strategies.

Our analysis shows that adherence to activity tracking is a good predictor of weight change. The positive association between adherence to activity tracking and weight loss is consistent with past studies on weight self-monitoring [18, 22, 27] and food and workout logging [14–16]. To the best of our knowledge, the temporal relationship between adherence to tracking and weight loss surfaced in our secondary analysis had not been explored before for general activity tracking. When restricted to weight self-monitoring only, our secondary analysis confirms previous findings on the directionality of the association [28, 29] and the high individual variability observed [22].

The primary novelty of our analysis lies in its setting, as our sample sizes are unprecedentedly high for studies of this nature that exist in the literature [19] in terms of number individuals (several thousand, depending on the activity) activity logs (several million) and longitude (more than two years). In addition, the population considered was not proactively enrolled in any weight loss programs or experiments, and the analyzed data was sourced by the users’ connected devices without solicitation—a setting that constitutes a major departure from the controlled environment described in previous research. We note that in a distributed setting such as the one considered, the decoupling between performing an activity and tracking it might be substantial (due to missing data, incorrect reports, etc.), therefore establishing a link between the observed behavior through the lens of the tracking device and the outcome variable does not immediately follow from previous studies in which observations were sourced from a controlled environment.

One limitation of our analysis is self-selection bias of the population under study. The on-line population considered consist of individuals who elected to buy a tracking device, connect them to the reward platform, and log at least a handful of activities. This indicates that the individuals may be on average more motivated to lose weight. We believe that the extra motivation towards weight loss, however present, should be comparable to that of other populations analyzed in previous studies, the majority of which belonged to weight loss programs or enrolled in experiments explicitly advertised as studying weight loss.

Another limitation to generalizing the findings of the present study is that users were receiving monetary incentives for logging activities and the findings might not hold if incentives were to be removed. We argue that is likely not the case as our inclusion/exclusion criteria selected users who had been on the platform for at least a few months, which is a time horizon beyond which it is known that monetary rewards, much like other external motivation strategies, tend to lose efficacy on habit formation [30–32].

We presented an analysis of the association between weight loss and temporal patterns of weight self-monitoring, food logging, and workout tracking. Our findings show that users who track activities more frequently on average tend to lose more weight than their peers who track activities less frequently. We also show that the claim still holds when considered longitudinally: when an individual increases their adherence to tracking they are more likely to lose or maintain their weight.

An overarching conclusion of this study is that data collected from activity trackers can inform real-time interventions [33] designed to adapt to users’ heterogeneous natures and performance, and targeted to improve the health outcome monitored. It is worth noting that such interventions need only rely on the predictive power of the observed variable to the target variable, irrespective of any direct casual link. For example, both increased adherence to activity tracking and weight loss could be the reflection of a person’s internal tendency to engage in health-enhancing behavior, a hypothesis known as “healthy user effect” [34]. In the “healthy user effect” hypothesis, no direct causal link between adherence to activity tracking and weight change is present, however effective interventions could still be designed, triggered by an observed drop in activity-tracking adherence, to nudge an individual into the “healthy state” (e.g., by reminding the user of the importance of healthy habits) thus ultimately affecting their weight.

The opportunity to leverage digital health technologies to extend the notion of precision medicine to preventative care [35] becomes especially relevant now that activity trackers are becoming mainstream, with a market size projected to increase to more than $50 billion by 2018 [36, 37]. However, as argued by Patel et al. in a recent opinion paper [38], the successful use and potential health benefits of digital health technologies does not only depend on penetration and adoption, but on the design of personalized engagement strategies that can cope with the high individual variability of the targeted users.