Study sample

This paper describes secondary analyses of the GoActive study data [12], a randomised controlled trial to increase moderate-to-vigorous physical activity among adolescents. The results of the GoActive trial showed that adolescents become more physically inactive over time with no difference found between control and intervention participants’ physical activity [13], therefore the cohort for this study included all participants from control and intervention schools. Methods of the GoActive trial have been described in detail elsewhere [12]; a brief summary of the methods relevant to the current analysis is provided below.

Recruitment and ethics

Baseline data collection took place September 2016-January 2017 with follow up April-July 2018. Participants (all year 9 students, 13-14years) were recruited from 16 non-fee-paying, co-educational secondary schools across Cambridgeshire and Essex, United Kingdom. Ethical approval for the GoActive study was obtained from the Cambridge Psychology Research Ethics committee (PRE.126.2016). The GoActive study was prospectively registered (ISRCTN31583496).

Outcome measure

MVPA was measured by wrist-worn accelerometers (Axivity AX3, United Kingdom). Participants were asked to wear the accelerometers for 7 days continuously on their non-dominant wrist. Trained research assistants fitted the devices during data collection visits in schools and participants were encouraged to return the accelerometer to their tutor group (homeroom) teachers at the end of the 7 days. Participants were also given a freepost envelope to return accelerometers to the study team. A novel data processing approach was developed [13] in which data were processed by first separating days of possible wear into quadrants: morning (6am-12pm), afternoon (12pm-6pm), evening (6pm-midnight), and night (midnight-6am). Participants were included if over six hours of wear time spread over at least two days was recorded in each of the first three quadrants (i.e. ≥6 hours from 7 possible mornings, ≥6 hours from 7 possible afternoons, and ≥6 hours from 7 possible evenings). The night quadrant was considered sleep time. An individual hour was included for analyses when at least 70% of possible wear time was recorded. A diurnal adjustment was used to reduce bias arising from the 24-hour wear protocol. Accelerometer output was processed to provide average daily minutes of MVPA at baseline and follow up which is equivalent to ≥2000 ActiGraph counts per minute [12]. Mean wear time was … Change in average daily minutes of MVPA was then calculated via the following equation:

Change in average daily minutes of MVPA = Follow up MVPA–Baseline MVPA

Change in average daily minutes of MVPA was selected as the outcome measure–as opposed to change in school time MVPA–because certain school environment features have a theoretical basis to influence physical activity both during and outside of school time. For example, schools may allow access to school sports facilities during weekdays after school or support extracurricular activities at weekends [14]

Exposures

A full list of exposure variables can be found in S1 Table. School level policy, social and physical environment features (n = 98)–including break time length, physical activity opportunities at school and hours of PE per week–were self-reported by contact teachers at all schools in a survey based on a questionnaire previously used in schools in the East of England as part of the Speedy-3 Study [15]. The survey used to collect these features can be found in S1 File. Individual level (n = 3) friendship support for physical activity was derived using three items from the European Youth Heart Study [16] self-reported by participants in baseline questionnaires. Additional school level exposure variables (n = 167)–including Office for Standards in Education (Ofsted) rating and school funding levels) were collected from publicly available data based on previous evidence [9] and after discussion with school-based physical activity promotion stakeholders from policy and practice (GoActive Final Report, 2020). School funding levels and Ofsted overall effectiveness rating–which assesses schools on factors including motivational climate, school connectedness, teacher support of pupils, teacher leadership behaviours and teacher and pupil skill building [17] may provide insight into the challenging-to-measure school social environment. These data were obtained from the following website: https://schools-financial-benchmarking.service.gov.uk/.

Descriptive data and covariates

Sociodemographic data (age, sex, ethnicity, language spoken at home, religious affiliation, family structure, family SES and BMI) were self-reported by participants at baseline. Participants reported their ethnicity from 20 response options and values were recoded to the following categories: White; Mixed ethnicity; Asian; Black; Other ethnicity. Religious affiliation was reported from eight response options and values were recoded to the following categories: No religion; Any religion. Language spoken at home was reported from 26 response options and values were recoded to the following categories: English only; English and other language(s); Other language(s) not English). Participants reported one or two main care-givers from eight response options and family structure was recoded to the following categories: Birth Mother and Father; Any other family structure. Participants reported six items from the Family Affluence Scale [18] (family car ownership, holidays, computers, availability of bathrooms, dishwasher ownership and having their own bedroom) and values were summed (possible range 0–13) and used as a proxy measure of family SES (affluence: low = 0–6, medium = 7–9, high = 10–13) [19]. Height and weight were recorded by research staff trained in anthropometric assessment and BMI z-score was calculated from height, weight, age, and sex [20].

Statistical analysis

As a large number of school environment features plausibly relate to physical activity, the least absolute shrinkage and selection operator (LASSO) variable selection method was used to determine the features of the school environment that are most relevant to change in MVPA [21]. The tuning penalty applied to the LASSO–which determines the amount of shrinkage applied by capping the sum of the absolute coefficients—was the Extended Bayesian information criterion which accounts for collinearity [22]. School environment features selected by the LASSO were added to a multiple linear regression model to identify features significantly associated with change in physical activity. Change in MVPA from 13y to 15y (average follow up = 18 months) acted as the dependent variable and the school policy, social and physical environment features served as independent variables. Estimates of change in minutes per day of MVPA per 1-unit change in each independent variable were reported. As GoActive was a cluster-randomized controlled trial and this study used data from the whole cohort, treatment group (intervention and control) was added to each LASSO, along with adjustments for baseline physical activity, baseline BMI and sex. This process was repeated separately by sex (Male, and Female) and SES (Low, Medium, and High) to stratify the analyses, looking for differences in effect. To account for the non-independence of participants (clustering within schools and SES), robust standard errors were calculated. LASSO variable selection has previously been used to identify the most meaningful correlates of adolescent physical activity using self-report physical activity data [23]. The LASSO method is different from traditional approaches to variable selection such as the stepwise approach which uses p-values to determine which variables to include. Instead, the LASSO shrinks the absolute value of the magnitude of regression coefficients of unrelated school environment features to zero by regularization. The main driver for using this method was that a large number of school environment features plausibly related to change in MVPA. As this was an exploratory analysis, intended to test many potential predictors of change as a hypothesis generating exercise, LASSO variable selection was used to allow only the most relevant school environment variables to physical activity to be included in the model. This was in order to produce a robust and parsimonious model via minimizing prediction error and improving interpretability [24]. The number of independent variables that regression analysis is capable of dealing with is a subject of contention [25]. The LASSO approach is justified to counteract the rise in biases (e.g., overfitting, multicollinearity) when there are greater than five independent variables in a multiple regression equation [26] To improve our understanding of these findings, post hoc analyses were performed to explore the association between change in the variables selected by the LASSO and change in MVPA. Statistical analyses was performed in STATA version 15.1 [27, 28] (using the lassopack (v1.2) installation package to run the LASSO.

Main analyses

A total of 267 school environment features were included in the LASSO variable selection technique, a full list of variables included in the LASSO is shown in S1 Table. Estimates of effects for the variables selected by the LASSO and included in the main analyses are shown in Table 2. No school policy or physical environment features were selected by the LASSO as potential predictors of change in MVPA. Only 3 variables (baseline MVPA, having a friend ask participants to do physical activities or play sports with them, and a participant asking a friend to do physical activities or play sports with them) were selected and added to the regression model when running the analyses with all participants. In this regression analysis, only baseline MVPA was found to be independently negatively associated with change in MVPA.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Estimated effects of the variables selected by the LASSO on change in MVPA during adolescence.

https://doi.org/10.1371/journal.pone.0249328.t002

Post hoc analyses

Longitudinal associations between change in friendship support for physical activity and change in MVPA are shown in Table 3. For all participants, both friendship support for physical activity variables were positively associated with change in MVPA. For every unit increase in change in participants asking friends to do physical activity, a change in MVPA of 2.78 (1.55, 4.02) minutes per day is predicted. For every unit increase in change in friends asking participants to do physical activity, a change in MVPA of 1.80 (0.48, 3.11) minutes is predicted.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Estimate effects of change in the variables selected by the LASSO and change in MVPA.

https://doi.org/10.1371/journal.pone.0249328.t003

In sub-group analyses by sex, for boys only, positive associations were found for both measures of friendship support for physical activity; whereas, for girls only, no associations were found for either measure of friendship support for physical activity. In sub-group analyses by SES, positive associations were found for participants with a medium and High SES, but no associations were found for participants with a low SES.

Main analyses findings

Longitudinal posthoc findings

A positive association between change in friendship support for physical activity and change in MVPA during adolescence was found. This finding is in line with the totality of the literature surrounding friendship support for physical activity. It has consistently been demonstrated that a positive association exists between friendship support for physical activity and physical activity during adolescence [31, 41]. Our findings lend support to the hypotheses that friendship support for physical activity may be an important mechanism for continued participation in physical activity throughout adolescence [43]. However, it is not possible to rule out that effects were not attributable to reverse causation, where an increase in physical activity prompted an increase in participants asking friends or being asked by friends to do physical activity. Further research is required to help explain how friendship influences physical activity during adolescence.

Friendships during adolescence may be positively or negatively influential on physical activity [44]. It has previously been demonstrated that girls receive less friendship support for physical activity than boys [45]. Friendship support for physical activity may have a greater influence on different types of physical activity, as stronger and more consistent associations were found between friendship support for physical activity and vigorous physical activities and competitive sports which may be preferred, at the population level, by boys [42]. Adolescents often report having a higher number of same-sex friendships [46] and tend to rely upon their peers to define acceptable and desirable behaviours [47, 48]. A gendered culture within British secondary schools, where boys and girls are encouraged to participate in different physical activities; may bring rise to gendered social norms and explain differences in gender peer modelling, as previously hypothesized by Reimers, Schmidt [45]. Gender stereotyping of physical activities, where social norms consider certain physical activities to be masculine or feminine, can negatively impact physical activity participation in girls [49] as adolescents may be more likely to imitate the physical activity of their same sex peers [46] and girls are less likely to do sufficient physical activity [50]. Female adolescents may have fewer active friends (relative to boys) on which they can model desirable physical activity behaviours. Further investigation is needed to confirm these hypotheses in order to aid the development of interventions designed to overcome inequalities in physical activity between boys and girls.

An alternative explanation for the finding of a positive association for boys (and not girls), is differences in motivation for physical activity between sexes. Motivation is a significant predictor of physical activity during adolescence [51]. However, correlates and determinants of physical activity often differ by sex and variations in adolescent’s motivation to be physically active have been found between boys and girls [52]. A difference in intrinsic motivation to participate in physical activity between boys and girls may also explain why a longitudinal, positive association was found between social support and change in MVPA in boys only. For example, boys may have higher intrinsic motivation to be active than girls and might seek more opportunities to participate in physical activity with friends.

Strengths and limitations

The main limitation of this study was the subjective quantification of school environment features which may have brought rise to self-report biases. A further limitation is that we were only able to analyse the social environment features longitudinally and it was not possible to analyse the school policy, and physical environment features longitudinally. These analyses were exploratory, rather than confirmatory. Selection/homophily or influence processes in adolescent friendship networks (e.g. whether individuals become friends and then friends influence physical activity during adolescence or whether friends select friends based on similarity in physical activity) is also a highly debated topic with mixed findings [53]. Therefore, selection bias within friendship groups may have impacted findings and identified variables should be considered potential correlates rather than causal mechanisms. It is also possible that effects were not found in participants with a low SES due to a small sample size. There is also currently no consensus in the existing literature on how to classify intensities of physical activity which brings rise to the possibility of a degree of misclassification. Strengths were the longitudinal, device-measured physical activity measurements and relatively large representative sample size. These longitudinal associations are more likely to reveal causal relationships since they account for fixed propensities to do activity that vary across individuals.

Conclusion

This study was an exploratory analysis, intended to test many potential predictors of change in MVPA within the school policy, social, and physical environment as a hypothesis generating exercise. No policy and social environment features were selected by the LASSO as potential predictors of change in MVPA. Two social environment features were identified by the LASSO but were not associated with change in MVPA in our main analyses. Post-hoc analyses suggested friendship support for physical activity could be harnessed to increase adolescent physical activity levels. Making changes in the school environment to promote friendship support for physical activity may increase adolescent physical activity [54]. Future research to better understand the school social environment, in particular exploring how friendships during adolescence influence physical activity via social network analysis, may provide insight into how to prevent the decline in physical activity during adolescence.