1. Introduction
Physical activity in children is associated with positive health outcomes, including a lower risk of cardiometabolic diseases and improved psychological well-being [
1,
2]. It is recommended that children over the age of 5 years engage in an average of 60 min of moderate-to-vigorous physical activity (MVPA) per day [
3,
4,
5], yet evidence suggests that many children do not achieve this level of physical activity [
6]. Strategies to increase children’s physical activity participation through targeting individual, interpersonal and school-based characteristics have had limited effects [
7,
8]. Thus, there is a need for a greater understanding of the factors that are associated with children’s physical activity.
Physical activity is a complex behaviour influenced by many individual, inter-personal, environmental and policy level factors [
9]. These different levels are consistent with the socio-ecological model, which suggests that each of the different levels impact on behavior, and that the different levels interact [
10]. It is noticeable that most studies have focused on the individual level, with factors including sex, age, perceived barriers to activity and motivation consistently associated with physical activity levels in children [
11]. Parental factors, including a higher socio-economic position and parental support of physical activity, such as encouragement or the provision of transport, have also been shown to be associated with child MVPA [
12]. Children’s physical activity participation may be influenced by the activity levels of their close friends [
13,
14], and several studies have shown that, during childhood and adolescence, friends engage in similar levels of activity [
15,
16]. Studies using social network analysis, where the full friendship network is modelled, show weak-to-moderate associations between children’s self-reported [
17,
18,
19,
20,
21,
22] and objectively-assessed MVPA [
23,
24,
25,
26], and there is some evidence indicating potential gender differences [
22,
26]. At the broader school-level, factors such as support for active transportation, physical activity policies, and facilities, have been shown to be associated with physical activity both in and outside of school time [
27,
28].
Multilevel models have been increasingly used to explicitly model complex hierarchical structures, where children within households are nested within schools and/or neighbourhoods. This type of model is especially common with school-based cluster study designs, where children are nested within schools, as it adjusts standard errors to account for clustering. However, such models also allow description and comparison of the amount of variability at the different levels. A multilevel model combines fixed effects, which are known covariates at different levels that explain physical activity, and random effects that partition the residual variation (that is, the variability due to unknown factors) into variability at the different levels.
Evidence using multi-level modelling consistently shows that children’s physical activity levels vary between schools, with the school accounting for between 2% and 10% of the total variance in daily MVPA [
27,
28,
29,
30,
31,
32,
33,
34,
35], indicating that the school environment can influence children’s physical activity. Most studies have used self-reported physical activity outcomes, but where objective measures of MVPA have been used, the between-school variance is higher, at around 6%–18% [
27,
31]. Some studies have looked simultaneously at school- and class-level variation to determine the proportion of variability attributable to each level. These studies have shown that variation in physical activity is seen at both the school- and class-level, but that there is greater variation at the class-level for physical activity both within and outside of school [
31,
35]. These differences remain after adjusting for individual demographics, such as age, gender, ethnicity and parent education level [
35], indicating that specific aspects of the school or classroom environment have the potential to influence physical activity participation. While the evidence cited earlier suggests that both peer groups and neighbourhoods vary, we are aware of no studies that have explicitly described between-friend or between-neighbourhood variations. It is also not clear how known factors and unknown variability at multiple levels are related, for example, whether differences between schools are due to school-level factors, demographics, structures such as classes or friendship groups within the school, or a combination of all of these.
The aim of this paper is to explore the variability in children’s accelerometer-measured MVPA at five different levels of the hierarchical levels of socio-ecological model: individual, friendship groups (at two levels), school and neighbourhood, both before and after accounting for child, parent and household factors. We identify the proportion of total variability that can be attributed to each level. These analyses are intended to enable a better understanding of how different factors explain variability at the different levels, and where the residual variability (due to unknown factors) lies. This will enable us to identify the levels within the hierarchical structure that are most important to the physical activity patterns of primary school aged children and help to inform the development of complex interventions targeted at these different levels to increase children’s physical activity at a population level.
3. Results
There was considerable variability in MVPA between schools (
Table 1), with average MVPA more than doubling between the lowest and highest schools on a weekday (38.6 versus 89.4 min per day) with comparable findings for weekend days. Missing data are described in
Table 2. MVPA missing data ranged from 12%–13% for weekdays and 22%–25% for weekends, and parent characteristics and accelerometer missing data ranged from 11%–23%.
Table 3 and
Table 4 report the percentage of the total variation in MVPA attributable to known covariates (fixed effects) and random variation at the different levels for age 9 and age 11, respectively. For all models, most of the variation in MVPA was at the individual level, that is, it was not explained by covariates and did not exhibit clustering at any level of the hierarchical structure. Model 1 explored the amount of total variation at each level, and showed patterns differing between weekdays and weekends and between age 9 and age 11. Between-school variation accounted for around 12%–13% of the total MVPA on weekdays, with less clustering on weekends at age 9 (6%). There was very little between-friendship group variation at age 9, but more clustering within friendship groups at age 11, especially on weekdays. On weekdays, the majority of this between-friend variation was clustering within dyads, but on weekends it was dominated by triads. There was very little between-neighbourhood variation. At age 11, more of the total variation was attributable to clustering within the hierarchical model structure, especially at the friendship level.
In Model 2, including gender as a fixed effect explained some of the variance at all levels, especially clustering within friendships, but there were strong differences in the patterns of remaining variation by gender. The DIC indicated a large improvement in the model fit when allowing variation to differ by gender. However, as with Model 1, the largest proportion of residual variability in MVPA remained at the individual level, accounting for around a third to half of the total for boys (36%–51%) and over two thirds for girls (60%–78%), with the remainder clustering at other levels of the hierarchical structure. At age 9, both genders showed similar clustering at the school level within the week (14%–16%), but between-school variation at weekends was over twice as much for boys than for girls (19% vs. 6%). A larger proportion of the variability was attributed to friendships for boys than girls, especially at weekends, accounting for 23% of the total on weekdays (combining dyads and triads) and 30% on weekends, compared to 12% and 11%, respectively, for girls. At age 11, both genders showed similar clustering at school level, with a higher proportion of between-school variability on weekdays compared to weekends (16% vs. 10% for boys and 16% vs. 9% for girls).
A larger proportion of the weekday variability was at the friendship group level for both boys and girls (40% and 20% respectively), compared to age 9. On weekdays, clustering was stronger between friendship dyads, but on weekends slightly more variability was attributed to friendship triads. At age 11, neighbourhood accounted for 16% of the total variation in boys’ weekend average MVPA, but for all other age, gender and day combinations, neighbourhood-level variation accounted for relatively small amounts of the total (3%–6%).
To explore the impact of missing covariate data, we reran Models 1 and 2 on the complete data (
Tables S1 and S2, Supplementary Material). This suggested that, at age 9, the percentage of variation at the school and neighbourhood levels was slightly overestimated in the presence of the missing data, while, at age 11, the percentage of variation at the dyad friendship level was slightly overestimated. Thus, the residual individual variation reported in
Table 3 and
Table 4 might, therefore, be higher than estimated.
Model 3 included child, parent and neighbourhood covariates. The DIC indicated a slight improvement in model fit, and both child and parental characteristics were associated with MVPA. However, the covariates explained only 3%–13% of the total variation, and this was mostly individual-level variation.
Figure 2 shows the percentage of total variation explained by the covariates, and residual variation at each level for girls and boys, based on Model 3. Estimates for the fixed effects are given in
Tables S3 and S4 (Supplementary Material).
4. Discussion
A summary of the main findings from this paper is presented in
Table 5, which highlights the important additions to the knowledge base of children’s physical activity and how they relate to the existing evidence base. Central among these findings is the new evidence that the largest source of variation in children’s physical activity operates at the individual level, accounting for two thirds of the variability for girls and half of the total residual variability for boys. Thus, the key finding of our study is that there are still important individual-level characteristics associated with children’s physical activity that we have not identified in this study, despite including a large number of variables at different levels. There have been a large number of interventions designed to increase children’s physical activity, yet the majority show, at best, only small increases, equivalent to an increase of around 4 min of MVPA per day [
52,
53,
54]. Our findings suggest that one reason for this could be that we simply do not understand enough about the factors that influence physical activity. This indicates a need for further research to better identify and understand such factors before we are able to develop more effective interventions [
55]. This finding suggests that, in addition to the current zeitgeist of focusing on structural and system level influences [
9], there is also an urgent need to enhance our understanding of individual level influences on children’s physical activity and that strategies to understand and then change all levels are needed to help more children to be physically active.
The data presented in this paper show that the between school variability was around 14%–16% of the total residual variation on weekdays. The between-friend variability increases between ages 9 and 11 from around 23%–30% to 29%–40% for boys and 11%–12% to 16%–20% for girls. These findings highlight the critical role that friends play on physical activity and how the importance of friends and friendship groups increases as children age. Collectively, these findings suggest that friend- and peer-focused strategies for behaviour change have considerable potential. Such work is ongoing, and a 2012 review highlighted 23 peer and friend-based physical activity studies. However, the majority of these were descriptive, cross-sectional designs that used self-reported measures of physical activity [
56]. As such, this is an area that would benefit from greater attention, using more robust measures of physical activity, and could provide a fruitful means for developing new behaviour change interventions.
There were substantial differences between girls and boys in terms of the amount of total variability, the level at which it clustered and the amount explained by known factors. The analyses also showed that the important predictors of physical activity differed with age and for weekday versus weekend days. There was more variability in MVPA among boys than girls, but more of this variability was explained by known factors (i.e., in this study, BMI, active travel, participation in sports clubs, parental physical activity, age, BMI and parental support), and a larger proportion of the residual variation clustered within the hierarchical structure. These results therefore suggest that the factors studied in this study explain more about the physical activity of boys than girls. Studies consistently show that boys are more active than girls at all ages [
6,
36,
38,
57], but there is little evidence that the effect of physical activity behaviour change interventions differs for boys and girls [
52], with generally limited impacts for both genders. It may therefore be the case that a greater understanding of the factors that influence children’s physical activity is required to promote effective behaviour change programs and, as such, the data from this paper highlight a particular need to look more closely at the individual level predictors of girls’ physical activity.
Weekends were different to weekdays, with nearly twice as much total variability at weekends, a smaller proportion explained by known factors, and less clustering of the residual variation. We have previously found [
57] that those children who are most inactive tend to be even more so at weekends, highlighting the importance of weekend physical activity. Current research in this area has been dominated by school-based interventions [
8,
52] with little attention on weekend or other non-school periods, such as school holidays. This is likely due to the relative ease of recruitment in schools and difficulties recruiting representative samples outside of schools. This finding highlights a need to develop strategies to understand the factors that impact physical activity at weekends and during school holidays, and to use this information to develop behaviour change programs for these settings.
Methodological Strengths and Limitations
The analysis conducted in this paper used a multi-level model to examine the impact of different types of influences on children’s physical activity. This approach facilitated an assessment of what we know and, perhaps more importantly, what we do not know about children’s physical activity. The information on what we do not know is particularly useful, as it can help to guide future exploratory research. We did not, however, assess psychosocial constructs, such as motivation and competence, that have been shown to be important in other studies of children’s physical activity [
58,
59]. It is important, however, to recognise that there were a number of technical challenges with this type of analysis, and these limit the potential to fully explore the impact of different sources of variation in physical activity. Key among these challenges is missing data. We are not aware of any routine methods to implement multiple imputation of covariates for a complex multilevel model that includes both cross-classified and multiple-membership levels. As a result, we have had to restrict our analysis to complete cases only. This is an important limitation for social network analyses, as omitting one observation potentially omits multiple friendship ties, which results in an overly sparse friendship network. There is currently no consensus about how to deal with missing data in such networks, or how much bias this may introduce in estimates of either fixed or random effects. Sensitivity analyses in the current study suggested that variation at the neighbourhood, school and friendship levels was slightly overestimated in the presence of the missing data. We were also unable to look at longitudinal change between ages 9 and 11. While, in theory, the MMMC model could be extended to a longitudinal model, in practice there was too little overlap between friendship networks at the two timepoints to facilitate these analyses. While some of this is due to changing friendships, a large portion is once again due to missing data. Finally, this model relies on Bayesian MCMC methods, with associated issues of convergence and computational resources. Future work could take advantage of approximation methods, such as integrated nested Laplace approximations [
60].