IntroductionPhysical inactivity is a significant public health concern, with recent data showing less than 25% of youth meet the recommended 60 minutes of moderate-to-vigorous physical activity (MVPA) per day [1,2]. Consequences of low MVPA and high sedentary time over the lifespan have significant well-established links to cardiovascular disease, diabetes, poor respiratory health, and cancer [2]. Levels of MVPA decrease as children transition into adolescence, with the majority of individuals failing to meet physical activity guidelines by the age of 13 years [3]. Therefore, interventions targeting increases in MVPA and decreases in sedentary time are critical for adolescents, with long-term benefits that may accrue through adulthood.
Effect sizes for pediatric physical activity interventions are small, producing an increase in objectively measured MVPA that is less than 5 minutes per day on average [4,5]. To improve the efficacy of physical activity interventions, behavioral scientists are leveraging digital technologies to gain insight into key determinants of adolescent physical activity. An innovative perspective involves examining how macrotemporal factors (ie, slow-changing variables over weeks or months) interact with dynamic, microtemporal factors (ie, fast-changing variables over minutes, hours, or days) to influence adolescent physical activity behaviors [6]. Studying physical activity behavior in this manner may increase the understanding of how fast-changing, microtemporal (eg, daily, within-day) and slow-changing, macrotemporal (eg, long-term) variables affect physical activity engagement and could subsequently inform adaptive intervention technologies that are individually tailored to accommodate these microtemporal and macrotemporal variables.
Studies demonstrate that contextual factors, such as the availability of sidewalks and playgrounds, are related to engagement in physical activity [7]. Other macrotemporal variables related to physical activity include person-specific motivational factors, such as self-efficacy for exercise and perceived behavioral control [8-10]. In addition, Cushing et al [11], as well as Dunton et al [12], demonstrated that microtemporal variability in affect (ie, positive and negative affect), physical feeling states (ie, energy and fatigue), and external context (eg, being at home, rainy weather conditions, or time of day) [13] is related to MVPA and sedentary time, and these findings have been extended to adults as well [14]. The field has successfully established broadly that MVPA interventions work; therefore, the critical next phase of intervention optimization is understanding for whom these interventions work and in what context [15].
Framed within the ORBIT (Obesity Relation Behavioral Intervention Trials) model of intervention development [16], we [17] developed and evaluated the NUDGE (Network Underwritten Dynamic Goals Engine) as an unrandomized proof-of-concept study. NUDGE is a mobile health (mHealth) intervention that uses an automated text-message platform to promote adolescent physical activity by prompting users to self-monitor and set daily goals for minutes of MVPA, which was informed by Carver and Scheier’s [18] Cybernetic Control Theory. NUDGE demonstrated feasibility and preliminary efficacy for improving adolescent MVPA (Cohen d=0.83) and sedentary time (d=0.40). Evaluating macro- and microtemporal variables and their influence on NUDGE could be especially informative and provide insight into individual determinants of physical activity engagement for intervention development.
This study is a secondary analysis of NUDGE to guide future intervention development within the “refine phase” of the ORBIT model. Briefly, this phase uses available data or literature to iteratively make changes to a program before conducting a fully powered randomized controlled trial. The goal is to understand how both macro- and microtemporal variables (eg., psychological and environmental variables at both levels) influence the efficacy of a brief mHealth intervention (ie, NUDGE) for physical activity. We expected that both macro- and microtemporal positive affect and energy would be related to larger intervention effects (ie, more MVPA and less sedentary time), whereas negative affect, fatigue, and perception of poor weather would be related to a diminished intervention effect.
MethodsParticipants and ProceduresWe combined participants from 2 separate studies with identical procedures, ecological momentary assessment (EMA) protocols, and objective monitoring of physical activity. The primary difference between study protocols was that one study involved the delivery of the NUDGE program to the intervention group (detailed below), whereas the other study involved an attention control group. The research team recruited participants in the NUDGE group in 2017 and the attention-control group from 2015 to 2016. Recruitment occurred across all 4 seasons of the year. Recruitment methods included flyers, information letters, and social media postings. We excluded adolescents from participating if they had vision loss or low vision, could not read English, or had a medical condition that would limit their physical activity.
We conducted phone screenings with adolescents 13‐18 years of age and their caregivers to determine their eligibility for the study and scheduled eligible adolescents for a baseline visit. At that visit, we obtained informed consent and assent for caregivers and adolescents, respectively. We provided participants with an accelerometer to wear on their nondominant wrist for the 20-day study period and trained them on how to respond to the EMA surveys via a study smartphone. To maximize adherence to the protocol [19], we allowed participants to select 4 times per day (2 morning and 2 afternoon and evening) to receive EMA surveys, with each survey delivered at least 2 hours apart. Participants in the NUDGE intervention chose an additional time in the evening to receive text messages to set physical activity goals and a time in the morning to receive a goal reminder text message. Participants received up to US $40 for completing study procedures.
NUDGE InterventionAs described elsewhere [17], the NUDGE intervention is an automated chatbot. Users are asked to set a daily physical activity goal in minutes of MVPA, which is stored by the system. Next, users are prompted to self-monitor their MVPA at the end of the day and are provided with feedback based on their goal attainment. This process repeats daily for the 20-day study period. A goal is set based on a user’s self-reported exercise from the prior day. Users are allowed to set slightly more or less ambitious goals depending on their preference, but overall, the system encourages increasingly strenuous goals until users achieve 60 minutes of MVPA in a single day.
Matched Attention Control ConditionWe drew a matched sample of 21 participants from a 100-person dataset of individuals who did not complete the intervention to create the matched attention control. Attention control participants were matched to NUDGE participants on sex, age, and race and ethnicity, with priority given in the order the variables are listed. Participants who did not receive the NUDGE intervention were given the same study experience. They wore an accelerometer on their nondominant wrist, answered EMA surveys at 4 self-selected times per day for 20 days, and completed the same baseline and exit questionnaires. Research staff explained that the study was an assessment of physical activity and psychological variables, and the reimbursement structure was the same as the NUDGE group.
MeasuresDemographicsAdolescents completed self-reported measures of demographic variables (eg, sex as a biological variable and race and ethnicity) during the baseline visit.
Physical Activity and Sedentary TimeWe used wrist-worn ActiGraph GT3X-BT accelerometers (ActiGraph LLC) to measure physical activity across 20 consecutive intervention days. We initialized accelerometers to detect movements at 30 Hz. We processed the accelerometer data in Actigraph Actilife software v6.10.2 and converted raw measurements into 60-second epochs. We removed sleep and nonwear periods using the Sadeh et al [20] and Troiano et al [21] algorithms, respectively. We used the Chandler algorithm [22], a validated algorithm for measuring MVPA in youth using wrist-worn accelerometers, to define cut points for sedentary behavior, light physical activity, and MVPA.
Positive and Negative AffectWe measured positive and negative affect 4 times per day via EMA using the Positive and Negative Affect Schedule for Children (PANAS-C; [23,24]). The PANAS-C has strong psychometric properties for describing positive and negative affect and also demonstrates good test-retest reliability and convergent and discriminant validity [24,25]. Previous studies show delivering the PANAS-C on a mobile phone is feasible, reliable, and successful in gathering EMA responses to measure affect in real time [11,12,26,27]. The positive affective domain included the following mood states: joyful, cheerful, happy, lively, and proud. The negative affective domain included the following mood states: miserable, mad, afraid, scared, and sad. Participants rated the extent to which they experienced each mood state on a 5-point Likert scale from 1, “not at all,” to 5, “extremely.” We summed item responses for the positive and negative affective domains for each survey completion. We then mean-centered these values so that all regression weights were relative to no positive or negative affect. Finally, we averaged these values across the day to produce estimates of positive and negative affect for each study day.
Energy and FatigueEnergy and fatigue were measured using questions from the Profile of Mood States (POMS; [28]) 4 times a day via EMA survey. For energy, 3 items with the highest factor loading from the vigor-activity subscale were used (ie, energetic, full of pep, and vigorous). For fatigue, 3 items with the highest factor loading from the fatigue-inertia subscale were used (ie, fatigued, exhausted, and worn out). Participants rated the extent to which they experienced each mood state on a 5-point Likert scale from 1, “not at all,” to 5, “extremely.” We summed item responses for the energy and fatigue scales for each survey completion. We also mean-centered these values so that all regression weights were relative to no energy or fatigue. We then averaged these values across the day to produce daily estimates of energy and fatigue.
Perceived Weather BarriersWe additionally measured participants’ perception of weather during EMA surveys with one item 4 times per day: “What is the weather like outside?” In this variable, the intent is to capture high or low perceptions of weather as a barrier and not actual weather conditions. Participants rated their perception of weather as 1, “too hot,” 2, “too cold,” 3, “just right,” 4, “too rainy,” or 5, “too windy.” Our app was limited and only allowed a single response to this prompt (eg, participants could not say it was both too hot and too windy). For this reason, we recoded responses to indicate endorsements of perceived optimal weather and perceived weather barriers (eg, 0=“just right” and 1=all other responses, respectively). Finally, a daily weather barrier score was generated to capture each of the 4 EMA responses by summing these responses across the day, where lower scores represented fewer perceived weather barriers to MVPA. The final daily weather barrier variable ranged from 0 to 4, where 0 represents a response of just right at each EMA survey and 4 represents an endorsement of a perceived weather barrier at each EMA survey. Higher scores indicate more cumulative perceived weather barriers within a day.
Data Analysis PlanWe first divided each hypothesized moderator (eg, positive and negative affect, energy, fatigue, and perceived weather) into between-person and within-person components. The 4 daily assessments were averaged into day-level variables, and then person-mean centering was used to calculate the within-person value [29]. For example, each participant received a personal between-person value of their positive affect, indicating their average level of positive affect during the study, as well as a within-person value of positive affect for each of the 20 days, indicating their daily positive affect relative to their average. Because within-day scores are aggregated at the day level, the within-person effects can be interpreted as day-to-day fluctuations.
Using SAS (SAS Institute) PROC MIXED, we used first-order autoregressive (AR1) multilevel models for the day-level physical activity and sedentary behavior dependent variables. Hypothesized moderators were specified as interactions with group participation (ie, NUDGE intervention vs attention control) on MVPA and sedentary time. All models were estimated using the full information maximum likelihood estimation [29], and separate models were run for each hypothesized moderator. Based on our prior work, we used a random linear model for time at the day level (with the first day of the study coded 0) in the current analyses [17]. For models with significant interactions, we conducted post hoc probing at the mean, as well as ±1 SD above or below the mean of the moderator to determine simple slopes and regions of significance. Each model is conditioned such that regression coefficients can be interpreted as minutes of MVPA or sedentary time per day.
Ethical ConsiderationsAll study procedures were approved by the local institutional review board at the University of Kansas (STUDY00140405). Informed consent and assent were obtained for data collected in this project and the study team has followed institutional policies for maintaining privacy and confidentiality. Data in the current manuscript are deidentified.
DiscussionPrincipal FindingsThis study aimed to determine how macro- and microtemporal variables influence the effects of our mHealth intervention targeting adolescent physical activity. All of the macrotemporal study variables moderated the effect of NUDGE on MVPA, but only macrotemporal energy moderated the effect of NUDGE on sedentary time. With the exception of macrotemporal perceived weather barriers on MVPA, all of our significant interactions were in the hypothesized direction. In other words, adolescents with higher levels of macrotemporal positive affect and energy experienced larger intervention effects (ie, more MVPA and less sedentary time). Similarly, those with lower levels of negative affect and fatigue experienced larger intervention effects. Microtemporal associations with MVPA identified in other studies were not as important in this study as might have been indicated by other associational work [10,12,30]. This study indicates that while microtemporal findings do likely hold potential for personalization, the most immediate opportunity may be at the macrotemporal level. Momentary affect is difficult to capture using passive monitoring; therefore, fortunately, macrotemporal assessment is both easier and more suitable [31,32].
The literature indicates that mHealth interventions produce significant positive health outcomes in youth. The majority of these interventions have used standardized protocols to deliver the same intervention content to users, usually in the form of text messages [33,34]. However, recent developments in mHealth technologies show that mHealth interventions can leverage personalization to promote behavior change [10,35]. In the next iteration of the NUDGE protocol, it may be possible to use baseline information on macrotemporal factors to optimize the intervention for each participant. For instance, participants with low levels of positive affect or high levels of fatigue at baseline may need a preliminary intervention that is more complex than simple text message prompts to set and keep track of their physical activity goals. In both instances, a brief behavioral activation intervention might be warranted to help adolescents acknowledge the importance of setting and meeting goals, even when mood or fatigue are barriers [36].
There are other tailoring opportunities based on how NUDGE interacts with macrotemporal variables assessed at baseline. For example, NUDGE produced a desirable effect for all participants regardless of macrotemporal negative affect or perceived weather. However, it was more effective for those with lower negative affect. For individuals experiencing high negative affect, it may make sense to address this barrier first before enrolling participants in the NUDGE intervention or to use MVPA promotion strategies that are unlikely to be affected by affect (eg, scheduled family exercise). Similarly, the intervention was particularly effective for participants who demonstrated higher than average positive affect. For this group, additional prompts might be wasted contacts that could decrease engagement or lead to habituation [37]. Our results on fatigue suggest that those with higher-than-average fatigue may need a sequencing approach to improve their fatigue before starting on the activity intervention. Finally, the relationship with weather demonstrates that those who perceive more weather barriers derive more benefit from NUDGE, which may indicate the importance of goal-setting and review even in the face of environmental barriers. All of the intervention strategies discussed above fit nicely in phase 2 of the ORBIT model for refining behavioral interventions before proceeding to fully randomized control trials [16].
Within the ORBIT model, NUDGE was tested in a nonrandomized, proof-of-concept study. Initial findings were promising and indicate randomized pilot testing is warranted. However, the ORBIT model encourages behavioral scientists to consider how an intervention can be optimized and refined to deliver more efficient interventions. Rather than making clinical judgments or relying on anecdotal evidence to guide intervention refinement, we can use data from this study to inform the development of future versions of NUDGE. One iteration might use responsive, context-dependent intervention support while another might test tailored decision points for goal setting based on trait-level characteristics (eg, high negative affect and fatigue). This process of using empirical evidence for optimization represents a step forward in the development and refinement of technology-based behavioral interventions [38].
Limitations and Future DirectionsResults from this study should be considered in light of several limitations. Most importantly, participants were not randomized. In addition, the sample was relatively small, participants were predominantly female and Caucasian, and only adolescents from a specific geographic location were included. Future studies should recruit a larger sample of youth to test the generalizability of effects found in this study and provide opportunities for fully powered analyses. In addition, a larger study may be able to identify microtemporal moderators that were not present in this study. While not a major contributor in this study, there is a growing body of literature establishing that microtemporal relationships do affect the health behavior of children and adolescents [30]. Another limitation of this study was our methodology for measuring psychological feeling states. While EMA allowed us to measure daily fluctuations in these states, our protocol instructed adolescents to answer several survey questions 4 times throughout the day for almost 3 weeks. Future studies may benefit from using smart devices (eg, smartwatches) to reduce participant burden by making survey prompts easier to complete and to capture physiological data related to energy and mood (eg, heart rate and skin conductance; [39]).
Findings from this study add further support for the effects of psychological variables and perceived weather barriers on adolescent MVPA and sedentary time. Results highlight the importance of assessing macrotemporal factors (eg, positive and negative affect, energy, fatigue, and perceived weather barriers) in the context of a physical activity intervention for adolescents. While our initial examination of NUDGE showed an overall increase in MVPA, the results from this study indicate that adolescents may respond differently to the NUDGE intervention based on their macrotemporal characteristics, such as positive and negative affect, energy, fatigue, and perceived weather barriers. Future mHealth intervention trials should test the impact of tailored content on adolescents’ MVPA and sedentary time to further refine adaptive physical activity interventions.
ConclusionOur secondary data analysis of the NUDGE intervention suggests that patterns of variables, such as affect, energy, fatigue, and weather may moderate the effectiveness of text-message interventions for physical activity. Future iterations of NUDGE may benefit from using tailoring variables to modify the intervention experience of each participant depending on their level of the variables explored in this paper.