Self-efficacy and Physical Activity in Overweight and Obese Adults Participating in a Worksite Weight Loss Intervention: Multistate Modeling of Wearable Device Data

Physical activity promotion is a public health priority. Physical activity confers protective health benefits including reduced risks of cancers of the breast, colon, endometrium, bladder, esophagus, kidney, lung, and stomach (1–3). It is also associated with a reduced risk of cardiovascular disease, type II diabetes, and other chronic conditions, and it is important for maintaining a healthy weight (1). Despite the benefits, less than a third of Americans meet nationally recommended physical activity guidelines (1). Overweight and obese individuals are at increased risk for being insufficiently active (4, 5).

Self-efficacy, defined as one's belief about his or her confidence in their ability to engage in a particular behavior (6), is an important predictor of behavior. Self-efficacy has been linked to physical activity in overweight and obese individuals (7), and interventions targeting it have been shown to be efficacious for increasing physical activity in this population (8). Limited research has examined the relationship between physical activity–related self-efficacy and longitudinal patterns of adherence to recommended physical activity guidelines in overweight and obese adults.

The nationally recommended physical activity guidelines were derived from a considerable body of epidemiologic research and provide utility for informing public health interventions. The guidelines include that adults should engage in at least 150 minutes of moderate-intensity, or 75 minutes of vigorous-intensity aerobic physical activity, or some equivalent combination, each week (1). Public health and worksite interventions often categorize participants into either sufficiently or insufficiently active based on this threshold, and focus on helping individuals who are insufficiently active to gradually increase their physical activity levels until they meet the recommended guidelines. Recognizing that long-term adherence to recommended physical activity guidelines introduces challenges beyond physical activity initiation, interventions for individuals who are sufficiently active are often centered on helping individuals maintain their physical activity levels.

Physical activity studies predominately rely on 7-day physical activity assessments that are conducted only a few times (e.g., pre- and postintervention). However, evidence indicates that physical activity levels can vary markedly from 1 week to the next (9, 10). The advent of wearable physical activity trackers may introduce opportunities to supplement behavioral intervention, and allow researchers to take a more dynamic approach to physical activity assessment. Research supports the feasibility of use and acceptability of commercially available activity trackers in a diversity of at-risk populations (11–14). They can facilitate prompt self-monitoring of behavior, which is a behavior change technique associated with increased self-efficacy and physical activity in obese adults (7).

While it is clear that wearable physical activity technologies have a role to play in the future of physical activity promotion efforts, aligning their use with public health initiatives presents challenges. Among them is that their output can be unwieldy. It can be difficult for researchers to relate observed data to existing scientific literature, and to evaluate the statistical reliability of apparent patterns. These issues are compounded by device nonwear, which is common and tends to increase over time (15). Traditional approaches for modeling longitudinal data may not be appropriate for studies with large numbers of individuals and a large number of observations within each person, and are limited in their ability to handle missing data/nonvalid wear. Novel modeling approaches are needed to best utilize the large amounts of data obtained from wearable activity monitors, and to better capture the within-person dynamics of physical activity patterns over time.

Multistate models are flexible extensions of survival models that are broadly applicable to the study of event histories (16–18). They model stochastic processes in time in which at any one point an individual occupies one of a number of prespecified, mutually exclusive states (e.g., sufficiently active, insufficiently active, not wearing device). A state structure defines the possible transitions between states, and multistate modeling can be used to model the instantaneous risks associated with transitioning from one state to another. Multistate modeling may allow researchers to glean practical insights regarding dynamic changes in intervention adherence and physical activity in terms that are concordant with nationally recommended guidelines.

In this study, we conducted a secondary analysis of physical activity data from overweight and obese school district employees participating in a 26-week weight loss intervention. We applied multistate modeling to the intensive longitudinal data from Fitbit physical activity trackers worn by participants to explore how baseline self-efficacy, body mass index (BMI), age, and time predicted patterns in physical activity, controlling for potentially confounding covariates.

We performed secondary analysis on data from the Vibrant Lives Plus program. Vibrant Lives Plus was a weight loss program conducted from November 2017 to May 2018 in a school district in the area of Houston, TX. It was designed to help school district employees be more physically active, eat smaller portions, and consume a healthy diet. The physical activity component of this intervention emphasized increasing moderate-to-vigorous intensity aerobic physical activity (MVPA). Program content was comprised of 16 lessons that participants received by email or mail over the course of 26 weeks. Participants received a Fitbit Flex 2 (Fitbit Inc.), an Aria Wi-Fi–connected scale, and 5–10 text messages per week that provided brief reminders about the lesson content. Participants received additional support if they needed help with using the technology or if they were having trouble meeting weight loss goals. Participants who lost less than 3% of their body weight after 12 weeks in the program received an additional three telephone coaching sessions. Questionnaires were administered at baseline and postintervention, which collected information related to participant characteristics and relevant psychosocial constructs. We used Fitabase (Small Steps Labs) to gather participants' Fitbit data from the Fitbit server. The University of Texas MD Anderson Cancer Center institutional review board approved the analysis of these data.

Participants wore their Fitbit device on their wrist. Proprietary Fitbit algorithms yield estimates of users' daily “Very active” and “Fairly active” minutes. To align with nationally recommended guidelines, we summed and collapsed these data to the week level. Previous research has indicated that this method of estimating weekly MVPA demonstrates concurrent validity with research-grade accelerometry (19). We defined nonvalid wear days as days in which the Fitbit device recorded fewer than 1,500 steps, as has been done in other studies (20, 21). We defined nonvalid wear weeks as weeks in which participants had less than four valid wear days (22).

We created a variable reflecting five states of physical activity. Each state corresponded to a combination of physical activity categories over the previous 2 weeks. The weekly categories were (i) meeting recommended guidelines (≥150 MVPA minutes), (ii) not meeting recommended guidelines, and (iii) nonvalid wear week. The physical activity states used for multistate modeling analysis represented transitions between these categories from 1 week to the next. They were (i) stable active (i.e., meeting physical activity recommendations for 2 weeks), (ii) stable insufficiently active, (iii) stable nonvalid wear, (iv) favorable transition (i.e., improvement in physical activity category), and (v) unfavorable transition (see Fig. 1 for an illustrative example). This state structure permitted 10 logically reasonable transition types (Fig. 2), which may have occurred over the course of 25 transition opportunities (due to the 26-week study design). We included both stable states and transition states as this distinction may be useful for informing just-in-time, adaptive intervention (23, 24); we specified this particular state structure to strike a balance between model parsimony and clinical utility.

We assessed state transition probabilities by individual-level covariates, including baseline physical activity-related self-efficacy, BMI, age, ethnicity, education, and marital status. For self-efficacy we created a variable based on a survey item that asked, “How confident are you in your ability to increase physical activity over the next 6 months?” Responses were presented on a 5-point Likert-type scale. We collapsed responses to reflect higher (i.e., “very” or “completely confident”) versus lower self-efficacy (i.e., “not at all,” “somewhat,” or “moderately confident”). For BMI category, we included a covariate reflecting whether participants were overweight versus obese. For age, we created a variable based on participants' median split (43 years). For ethnicity, we included a variable indicating whether or not individuals identified as Hispanic. For education, we included a variable indicating whether or not individuals had attended at least some college or had a 2-year degree. For marital status, we included a variable indicating whether or not individuals were married or living with a significant other. In all models, we included a time-varying covariate to compare state transition probabilities in the first half of the intervention period to those in the second half; we chose this time point based on literature highlighting novelty effects for physical activity–related wearable technologies that may tend to persist for up to three months (15).

We fit Cox proportional hazards regression models to evaluate the multistate model (25). We used the Andersen and Gill formulation of the Cox proportional hazards regression model to account for the count-level data associated with multiple state transitions per participant (26). Multistate models are commonly based on the Markovian assumption that future transition intensities depend only on the present state. Anticipating that the amount of time spent in a state may have bearing on subsequent transition probabilities, we evaluated a Cox semi-Markov model, also known as a clock reset model (16, 17). This approach makes the assumption that future states are predicted both by one's current state and the time spent in that state. Thus, we evaluated Cox proportional hazard regression models in which event probabilities corresponding to state transitions were regressed on predictor covariates, taking into account the time spent in the preceding state (16). We conducted a likelihood ratio test to evaluate the global statistical significance of the model, and calculated HRs and corresponding 95% CIs for covariates. We set the nominal alpha to 0.05, and conducted analyses using the mstate and survivial packages in R version 3.5.3.

One-hundred participants were recruited for Vibrant Lives Plus. Of those, four did not set up their Fitbit devices. Thus, there were 96 participants and 2,496 person-week observations for the 26-week study period. Forty-one percent of the analytic sample identified as Hispanic and all participants were female. The average age of participants was 43.6 years (SD = 9.51) and 70 (73%) of participants were obese (Table 1).

Over the course of the study, the average prevalence of individuals in the stable active, insufficiently active, and nonvalid wear states was 13%, 44%, and 16%, respectively (Fig. 3). The average prevalence of individuals in the favorable transition and unfavorable transition states was 12% and 14%, respectively. Once a participant reached a stable active, insufficiently active, and nonvalid wear state, on average, they stayed in that state for 3.5, 6.4, and 5.9 weeks, respectively, and had a 27%, 43%, and 48% chance of staying in that state for 4 weeks or more.

Over the course of the study, 29% of participants in a favorable transition state next made it to a stable active state. However, 50% next descended to an unfavorable transition state, and 21% next entered a stable insufficiently active state (Table 2). Meanwhile, 42% of participants in an unfavorable transition state next ascended to a favorable transition state, while 37% and 21% of participants next entered a stable insufficiently active and stable nonvalid wear state, respectively.

Individuals with higher self-efficacy at baseline were more likely to reach a stable active state than individuals with lower self-efficacy over the course of the intervention (Table 3). Adjusting for covariates, individuals with lower self-efficacy who were in a favorable transition state were 49% less likely to transition to a stable active state than individuals with higher self-efficacy [HR = 0.51; 95% CI (0.29–0.88)]. The prevalence percentage of being in a stable active state over the course of the study was 18% for individuals with high self-efficacy, compared with 6% for individuals with low self-efficacy.

Individuals with an obese status at baseline were less likely to make a favorable transition out of an inactive state over the course of the intervention. Adjusting for covariates, obese individuals who were in a stable insufficiently active state were 39% less likely to transition to a favorable transition state than nonobese individuals [HR = 0.61; 95% CI (0.40–0.91)]. The overall prevalence percentage of being in a stable insufficiently active state was 47% for obese individuals, compared with 37% for individuals who were not obese.

There was better adherence to wearing the Fitbit in the relatively older individuals (i.e., older than 43 years). The HRs for two transitions highlight this tendency. Adjusting for covariates, older individuals who were in a stable insufficiently active state were 42% less likely to move to an unfavorable transition state than younger individuals [HR = 0.58; 95% CI (0.35–0.95)]. In turn, older individuals who were in an unfavorable transition state were 47% less likely to move to a stable nonvalid wear state [HR = 0.53; 95% CI (0.30–0.93)]. The overall prevalence percentage of being in a stable nonvalid wear state was 25% for younger individuals, compared with 8% for older individuals.

Adjusting for covariates, individuals who were not married or living with a significant other were 84% more likely to transition from a stable insufficiently active state to an unfavorable transition state, compared with individuals who were married or living with a significant other (HR = 1.84; 95% CI (1.10–3.07)]. Adjusting for covariates, neither ethnicity nor education evidenced statistically significant associations with state transitions.

In the second half of the intervention, inactive participants became less likely to wear their Fitbit device. Adjusting for covariates, study participants were 73% more likely to transition from a stable insufficiently active state to an unfavorable transition state in the second half of the intervention period, compared with the first [HR = 1.73; 95% CI (1.07–2.81)]. The overall prevalence percentage of being in a stable nonvalid wear state was 23% for the second half of the intervention period, compared with 9% for the first half.

In this study, we found the percentage of overweight/obese individuals in an active state was, on average, 13%. This is concordant with surveillance studies that have highlighted a low prevalence of meeting recommended physical activity guidelines in middle-aged women in the United States (27). School district employees may face an increased risk of inactivity and obesity due to long work hours and low occupational physical activity (28, 29). Physical activity–related self-efficacy appears to be a protective factor. The finding that self-efficacy was positively associated with reliably meeting aerobic physical activity guidelines is in accord with previous research that has pointed to the utility of self-efficacy for predicting physical activity (30). This is encouraging, given that interventions facilitating health self-management behaviors, such as providing prompt feedback, have been shown to increase self-efficacy (31). Furthermore, intervention content focused on increasing self-efficacy has been shown to lead to increases in physical activity in overweight and obese individuals (7). Our findings extend previous literature by analyzing longitudinal physical activity patterns, rather than treating physical activity as a static construct.

The finding that obese individuals in an inactive state were less likely to make a favorable transition out of that state is concordant with previous literature that has identified a reciprocal relationship between obesity and physical inactivity (4). Intervention content targeting physical activity-related self-efficacy, such as facilitating mastery experiences with meeting recommended guidelines, may be particularly useful. Given the high prevalence of obesity in the millions of people employed by U.S. public school districts, the school environment presents an opportunity for public health impact. Furthermore, interventions targeting school staff may have secondary benefits for school children, because teachers play a particularly important role in modeling health-related behaviors (32).

Relatively older participants (range 44–62 years) had better adherence to wearing the Fitbit than younger participants (23–43 years). This is somewhat surprising, given that age has been associated with barriers to new technology use (33). Although previous research has demonstrated high wear and acceptability of wearable physical activity technologies in middle-aged and older women (9, 34, 35), limited research has examined adherence to device wear by age in overweight and obese women. Results also indicated that those who were not married or living with a significant other were more likely to stop wearing the device if in an inactive state. The Fitbit device used in this study was a wrist-worn device. It may be that participants' perceptions and motivation to wear this highly observable device differed by these demographics. Future research should investigate how to best engage different populations. Future research may also investigate the effect of one's spouse or significant other wearing a physical activity tracker; this may have implications for dyadic approaches to physical activity intervention.

Our results indicated that the proportion of nonvalid wear increased throughout the course of the 6-month intervention, and that it tended to be individuals who were insufficiently active who exhibited device nonwear. It may be that the intrinsic motivation associated with the new device faded over time. Indeed, the novelty effect of wearable physical activity trackers has been shown to last for up to 3 months, and its attenuation may be associated with discontinued use if other sources of motivation do not supplant it (15, 36). The value that users derive from wearable physical activity trackers influences their long-term use and is a function of the interplay between the information the device provides, associated motivations (e.g., from health-related concerns, athletic inclinations, gamifications), and social norms (15). Behavioral interventions centered on the provision of wearable technologies should anticipate this novelty effect and couple device feedback with content related to participant's values, or other enduring forms of motivation (37).

Our application of multistate models to physical activity data provided insights regarding how baseline factors predicted longitudinal patterns in the transitions between physical activity states corresponding to nationally recommended guidelines and device nonwear. This is the first such application of multistate modeling. Multistate models have been applied to model processes pertaining to addictive behaviors (e.g., abstinent, partial lapse, complete relapse; refs. 38–40), but research pertaining to lifestyle physical activity patterns is limited. In addition to evaluating the predictive relationships of baseline covariates, multistate modeling could be used for statistical inference regarding group assignment in prospective studies. In addition, the use of time-varying variables could also allow for the conduct of natural experiments. For example, in the context of the present dataset, it may be possible to assess how standardized testing dates or school holidays affect teachers' physical activity levels. Weeks 2, 8, and 18 of the study corresponded to Thanksgiving break, New Year, and Spring break, respectively. In Fig. 3, each of these periods can be seen to correspond to a distinct uptick in the percentage of participants in a favorable transition state.

Multistate modeling of physical activity has potential to serve as the basis for just-in-time, adaptive intervention (JITAI; refs. 23, 41). As demonstrated in this study, multistate models may be used to categorize participants' real-time physical activity in ways that are congruent with public health recommendations. In future studies, a participant's state status could be used to determine intervention content in real time. For example, an individual in a stable insufficiently active state may receive messaging with techniques centered on facilitating physical activity initiation. If and when that person enters a favorable transition state, intervention content may then provide prompt positive reinforcement. Individuals in a stable active state may receive intervention content centered on physical activity maintenance, while alternate strategies may be used to reengage individuals who stop wearing the device. Highlighting individuals who are in a state of transition may enable interventionists to devote resources to deliver appropriately tailored content as participants cross clinically meaningful thresholds. Moreover, in addition to the estimation features utilized in this study, multistate modeling enables prediction of future event probabilities (16). Tailoring of JITAI content may thus be further bolstered by the prediction of an individual's future states, given their unique baseline factors and time spent in their current state. Researchers may be able to anticipate when participants face an increased probability of crossing a clinically meaningful threshold, and provide appropriately tailored intervention content. Calculation of the transition probabilities that facilitate prediction are done directly in multistate models that make the Markovian assumption, and can be readily simulated in semi-Markov models such as the one utilized in this study (16). Future research should explore the potential of multistate modeling to inform just-in-time, adaptive intervention.

A limitation of this study is the use convenience sampling, which may have tended to identify especially motivated individuals. Related to this, although males were not explicitly excluded, all participants in this study's analytic sample happened to be female. This limits the generalizability of study findings. Furthermore, this sample primarily included non-Hispanic and Hispanic whites. Other racial/ethnic groups, including those identifying as African American or Black, were underrepresented. Larger, more representative samples are needed to more fully investigate whether the patterns observed in this study are generalizable. In keeping with other literature (19, 34, 42–44), we assessed MVPA by combining Fitbit's estimates of “Very active” and “Fairly active” minutes. This method of physical activity assessment does not differentiate between moderate- and vigorous-intensity physical activities. Thus, there may have been some misclassification with respect to whether participants met nationally recommended guidelines each week (i.e., 75 minutes of vigorous-intensity activity). Nonetheless, a strength of this study is the use of device-measured, longitudinal physical activity data that was assessed in a way that has demonstrated concurrent validity with accelerometry (19). Another limitation is the use of a single item to measure self-efficacy. While single-item assessments may provide predictive validity in some circumstances, multiple item assessments are generally superior and should be used in future studies (45, 46). The measurement of self-efficacy only at baseline is another limitation. Self-efficacy may have changed for individuals during the course of the study. This limitation is particularly relevant given our commentary on the use of multistate models for JITAI. Although findings revealed that this construct had predictive utility for subsequent physical activity patterns, it is likely that such prediction could be substantially improved by repeatedly measuring self-efficacy in parallel with physical activity assessment. Indeed, previous research has found self-efficacy in the morning to predict subsequent physical activity patterns (47, 48). Self-efficacy is a dynamic, modifiable construct that is highly predictive of physical activity behaviors (24); those designing JITAIs for physical activity promotion in overweight and obese individuals should consider prioritizing longitudinal assessment of this and other key psychosocial determinants of physical activity. This notwithstanding, it can be helpful to be able to use baseline characteristics, which can be collected with less participant burden, to tailor overall intervention approaches. For example, it may be prudent to provide a more intensive intervention for participants who are obese or have low self-efficacy at baseline.

We found that baseline self-efficacy was positively associated with entering a stable, physically active state. Obesity status was negatively associated with making an upward transition out of a stable insufficiently active state. Relatively older participants were less likely to stop wearing the physical activity tracker than the younger participants, and device nonwear tended to increase for insufficiently active participants over the course of the study. Multistate modeling can provide useful insights for intensive longitudinal physical activity data, including: descriptive information of trends over time, explicit modeling of nonvalid wear, and the ability to provide statistical inference for covariates. Future applications of multistate models may permit dynamic prediction of an individual's future states, which could inform just-in-time, adaptive behavioral intervention for cancer prevention and control.

No potential conflicts of interest were disclosed.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

Conception and design: M.C. Robertson, K.M. Basen-Engquist

Development of methodology: M.C. Robertson

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K.M. Basen-Engquist

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.C. Robertson, C.E. Green, Y. Liao, C.P. Durand

Writing, review, and/or revision of the manuscript: M.C. Robertson, Y. Liao, C.P. Durand, K.M. Basen-Engquist

Study supervision: K.M. Basen-Engquist

M.C. Robertson was supported by the NCI of the NIH under award number F31 CA236433. Y. Liao was supported by a faculty fellowship from the Duncan Family Institute for Cancer Prevention and Risk Assessment at The University of Texas MD Anderson Cancer Center. This work was supported by the NIH through MD Anderson's Cancer Center Support Grant (NCI grant P30 CA016672), Assessment, Intervention and Measurement (AIM) Shared Resource. This research was made possible by the Center for Energy Balance in Cancer Prevention and Survivorship, and the Duncan Family Institute for Cancer Prevention and Risk Assessment at The University of Texas MD Anderson Cancer Center. Vibrant Lives is supported by the Pasadena Vibrant Community. The Pasadena Vibrant Community is an initiative of The University of Texas MD Anderson Cancer Center made possible by an investment from and collaboration with Shell Oil Company.