Longitudinal patterns of physical activity, sedentary behavior and sleep in urban South African adolescents, Birth-To-Twenty Plus cohort

We analyzed data from participants in the Birth-to-Twenty Plus Cohort (Bt20+), a longitudinal study of singleton children born between April and June 1990 in Soweto and in Johannesburg, South Africa. Detailed information on the cohort is provided elsewhere [28, 29]. Singleton children born in public hospitals to Soweto or Johannesburg residents and residing in the area for at least 6 months after the birth were eligible for inclusion. Of all 5449 births notified in the study area, 3,273 met the residency eligibility criterion and were recruited. Most attrition occurred in the first few years of the study, primarily due to migration out of the study area [30]. At inception the cohort was 78% Black African, with approximately equal numbers of males (49%) and females (51%); attrition was higher for White cohort members [31]. Ethical approval was obtained from the University of the Witwatersrand Committee for Research on Human Subjects (approval ID #M010556). Written participant assent and caregiver permission to participate in the study were obtained at each wave.

Physical activity and sedentary behavior were self-reported at ages 12, 13, 14, 15, 16 and 17/18 years using a questionnaire developed for South African children and validated in a South African population (Additional file 1) [32]. Respondents were asked to report about frequency and duration of informal physical activity, physical education at school, school sports, club sports and active transportation on a typical week over the past 12 months. Informal physical activity includes physical activity during school breaks or outside of school and not part of a sports team or club, such as skipping, traditional games and playing with a ball; respondents were asked to report on the duration of the three most frequent activities. Due to changes in the wording of the questionnaire between age 12 and age 13, we did not include informal physical activity at age 12 in our analyses. Physical education was defined as any exercise class supervised by a teacher during school time. School sport was defined as any extramural sport organized by the school; club sport was defined as any private extramural sport. Active transportation was defined as walking or biking to and from school; as few participants reported using bicycles, we included only time spent walking in our analyses. Questions regarding sedentary behavior included time spent watching television and videos; reading, drawing and homework; playing a musical instrument; playing video/computer games, and internet surfing, before and after school. At ages 13, 14, 15, 16 and 17/18 years, respondents reported the times they went to bed and woke up on school nights and weekends/holidays, from which we calculated hours of sleep. No information was solicited on quality of sleep.

We calculated informal physical activity minutes by summing the minutes spent per week in each of the three most frequent activities. We calculated physical education minutes by multiplying the reported number of classes per week by the reported number of minutes per class. We created a composite measure of organized sport by combining reported minutes per week of school sports and club sports. We inferred duration of competition for each sport based on sport-specific rules (e.g., rugby has two 40-min halves). If a sport had no rules governing the duration of competition (e.g., gymnastics or tennis), we inferred a duration of 60 min. We then summed the reported number of minutes per week across all reported sports. To better represent moderate-vigorous intensity physical activity (MVPA) participation, only sports with a metabolic equivalent (MET) value ≥3 based on the Compendium of Physical Activity for Youth were included [33]. Walking minutes per week was calculated as the sum of minutes per day spent walking to and from school, multiplied by 5 days per week. Intensity of walking was not reported, and we did not include this walking in our estimate of MVPA. We calculated total sedentary behavior minutes per week by summing the reported time spent per day in each of the activities.

We selected potential predictors a priori, including factors at the individual, maternal and household level [34‐37]. All these variables were collected by questionnaire administered to the mother at recruitment into the cohort or at child age 7/8 y. Covariates include socio-economic status (SES) at child age 7/8 years, mother’s schooling at child age 7/8 y (matriculated from secondary school, did not matriculate), mother’s marital/union status at child age 7/8 y (in union, not in union) and whether the child was first-born or not. These variables have been identified as correlates of physical activity and sedentary behavior in childhood in a subset of Bt20+ participants and in adolescence in rural South Africa [38, 39]. Household physical assets at age 7/8 were used as a proxy for socio-economic status (SES). A validated standardized questionnaire based on the Demographic and Health survey for developing countries was used to calculate an asset score [40]. Mothers were asked about home type, home ownership, electricity in the home, and ownership of a television, car, refrigerator, washing machine, and phone. The resulting score was categorized into quintiles.

We restricted analyses to participants who were Black African due to low numbers of the other ethnic groups. Analyses were further restricted to participants who had physical activity, sedentary behavior and sleep data for at least two data points between the ages of 12 and 17/18 years to ensure model stability.

We stratified all analyses by sex due to well-established sex differences in adolescent physical activity levels [11, 39, 41, 42]. We conducted a sensitivity analysis by comparing SES asset quintile, mother’s education, mother’s marital /union status, mother’s age at delivery, and parity, between participants who met inclusion criteria (Black African ethnicity and ≥ 2 data points) and participants who did not meet inclusion criteria, using chi-square tests. Descriptive summary data are presented as 25th percentile (Q1), median (Q2) and 75th percentile (Q3) for time spent in physical activity, sedentary behavior and sleep (Table 1).

We used Latent Class Growth Analysis (LCGA) to group participants into distinct classes based on common longitudinal trajectories of informal physical activity, physical education, organized sports, walking to and from school, sedentary behavior, school-night sleep or weekend sleep. LCGA is a type of Growth Mixture Modeling where it is assumed that the estimated variance and covariance of the growth factors within each class are fixed to zero [43]. We used full information maximum likelihood estimation to integrate all available information based on missing at random assumptions [44].

We determined the optimal number of latent classes using Bayesian Information Criterion (BIC), entropy, the Lo-Medell-Rubin-Likelihood Ratio Test (LMR-LRT) and the Bootstrap Likelihood Ratio Test (BLRT) [43]. The best model was considered the one with the lowest BIC value, highest entropy significant LMR-LRT, and significant BLRT. In addition to model fit indices, we also considered the percentage of participants per class, parsimony, theoretical justification and interpretability [43]. We implemented a restriction requiring at least 5% per class to avoid small group sizes. After establishing the number of trajectory classes, we assigned participants to a class based on the individual’s maximum posterior probability. Because we determined there to be only one physical education trajectory class in both males and females, we did not include this domain in subsequent analyses. We explored the relationships between individual trajectory classes across domains using chi-square tests. We identified predictors of trajectory class membership for each movement behavior using multinomial logistic regression. We conducted the LCGA in MPlus 7.3 [44] and the logistic regression in Stata version 14. We set significance at p < 0.05.

To describe overall movement behavior patterns, we performed multi-trajectory modeling. This modeling technique is designed to identify latent clusters, or multi-trajectory “groups,” of individuals following similar trajectories over multiple indicators of an outcome [45]. We examined two sets of multi-trajectory models: (1) overall physical activity - informal physical activity, organized sports and walking to and from school; and (2) overall movement behavior defined by the 3 physical activity domains, sedentary behavior and school-night sleep. For both overall physical activity and overall movement behavior, we chose the preferred multi-trajectory model using a two-stage model selection process. In the first stage, we determined the number of multi-trajectory groups based on BIC model fit statistics, practical interest and ability to distinguish between groups. In the second stage, we determined the preferred order of the polynomial defining the shape of each trajectory, given the number of groups selected in the first stage [46]. We compared characteristics of the multi-trajectory groups using chi-square tests, and identified predictors of multi-trajectory group membership using multinomial logistic regression. We conducted the multi-trajectory analysis using a plug-in for Stata version 14 [45].

Duration of all forms of physical activity decreased over adolescence, while sedentary behavior increased and duration of weekend sleep remained stable, in both males and females (Table 2). School night sleep decreased across adolescence in males but remained stable in females.

We identified 2 trajectories for informal physical activity for males, with one (93% of the sample) showing a steady decrease in physical activity from age 13–18 years and the other showing a steady increase (Fig. 1). In females, we also identified two informal physical activity trajectories with one (95% of the sample) showing a steady decrease and the other showing a decrease followed by a sharp increase at age 16 years. For organized sports, we identified 3 trajectories for males: consistently low (82%), increasing (11%) and decreasing (8%) and two for females, representing no participation (89%) or some participation in organized sports from age 12. We identified 2 trajectories of walking to and from school, lower (representing 82 and 75% of males and females, respectively) and higher.

For sedentary behavior, we identified 3 trajectories in males, showing steady lower (78%), steady higher (10%) and increasing (12%) and 2 trajectories in females, increasing (92%) and steady higher (Fig. 2). For school-night sleep, we identified 4 trajectories in males and 3 trajectories in females. For weekend sleep, we identified 3 trajectories for males and 4 trajectories for females. Model fit statistics of individual trajectories are provided in Additional file 2: Tables S1 and S2.

In males, the distribution of the sedentary behavior trajectories was independent of the distributions of all the physical activity domain trajectories (Additional file 2: Table S3). In females, the distribution of the sedentary behavior trajectories differed across the distribution of the walking to and from school trajectories (p = 0.04) but was independent of the distributions of the informal physical activity and organized sport trajectories (Additional file 2: Table S4). In both males and females, the distributions of school-night sleep trajectories and weekend sleep trajectories differed across the distribution of sedentary behavior trajectories (p < 0.001 and p = 0.002). The distribution of the school-night sleep trajectories differed by the distribution of the walking to and from school trajectory in females but not in males (p = 0.001) (Additional file 2: Tables S5 and S6).

Informal physical activity: Being in a higher SES quintile was associated with higher likelihood of being in the steady increasing trajectory compared to the steady decreasing trajectory among males (Relative Risk Ratio (RRR) =1.48 95% CI: 1.11–1.98) (Table 3).

Walking to and from school: Being in a higher SES quintile and having a mother who matriculated from secondary school were each associated with lower likelihood of being in the higher trajectory compared to the lower trajectory in males (RRR = 0.81 95% CI: 0.66–0.98 and RRR = 0.55 95% CI: 0.33–0.99) and females (RRR = 0.68 95% CI 0.58–0.81 and RRR = 0.47 95% CI: 0.28–0.80).

Sedentary behaviors: Among males, having a mother who matriculated from secondary school was associated with higher likelihood of being in the increasing trajectory compared to the steady lower trajectory (RRR = 1.82 95% CI: 1.02–3.25).

Sleep: In adjusted models, having a mother who matriculated from secondary school was associated with over two times the likelihood of being in the 8 h per night trajectory and over three times the likelihood of being in the 7 h per night trajectory compared to the 9 h per night trajectory in males and lower likelihood of being in the 9 h per night trajectory compared to the 8 h per night trajectory (RRR = 0.82 95% CI: 0.70–0.95 and RRR = 0.46 95% CI: 0.30–0.72) in females.

None of these variables were associated with informal physical activity or sedentary behavior trajectories in females or with organized sport and weekend sleep trajectories in either males or females.

Overall physical activity: We identified 3 multi-trajectory groups for males and for females (Additional file 2: Figure S2). In both male and female multi-trajectory groups 1 and 3, informal physical activity and organized sport trajectories declined throughout adolescence. The main distinguishing characteristic between groups 1 and 3 was the level of walking to and from school. Group 2 maintained higher levels of both informal physical activity and organized sports than groups 1 and 3 and had a walking to and from school trajectory similar to group 3. Group 1, representing 23% of the sample in males and 32% of the sample in females, respectively, can be described as “Decreasing Activity, non-walkers”, group 2 (29 and 17%, respectively) as “Active, walkers” and group 3 (48 and 51%, respectively) as “Decreasing activity, walkers.” Model fit statistics of the physical activity multi-trajectory groups are shown in Additional file 2: Table S7.

Higher SES quintile and having a mother who matriculated from secondary school were each associated with increased likelihood of being in group 1 (“Decreasing Activity, non-walkers”) compared to group 3 (“Active, walkers”) but were not associated with risk of being in group 3 (“Decreasing activity, walkers” compared to group 2 in either males or females (Table 4).

Overall movement behaviors: We identified 4 multi-trajectory groups for males and 2 for females (Additional file 2: Figure S3). Model fit statistics are shown in Additional file 2: Table S8. In males, the most notable distinctions between the multi-trajectory groups were seen among the 3 physical activity trajectories, with the sedentary behavior and sleep trajectories in each group being nearly identical. In females, the primary distinction between multi-trajectory groups related to walking to and from school, with the other trajectories showing minimal differences between the two groups.

A major strength of our study was the use of longitudinal data throughout adolescence. Another strength of our study was the use of LCGA to group participants into common longitudinal trajectories of physical activity domains, sedentary behavior and sleep. The goal of LCGA is to model individual trajectories and is recommended when a single growth curve may not describe the variables of interest in the population adequately.

The use of LCGA comes with some limitations. In LCGA, class membership is inferred using a series of exploratory analyses to determine the best fitting model. The validity of LCGA may be questioned because model fit statistics may not adequately differentiate between a model with multiple latent classes and a single-class model with non-normal outcomes [65]. However, using multiple indicators of model fit strengthens the validity of our findings [66]. Both LCGA and multi-trajectory modeling assign class/group membership based on estimated posterior probabilities from a maximum likelihood analysis. Not all participants follow a trajectory/group perfectly and misclassification can occur [43, 45]. Our attempt to generate multi-trajectory groups that included physical activity, sedentary behavior and sleep had limited success, as there was not enough heterogeneity in the linkages between these behaviors in this population to identify clusters of individuals in terms of these outcomes together. The questionnaires changed slightly over time. All the measures we analyzed were self-reported, and we do not have data on intensity of informal physical activity or walking or on quality of sleep.