Introduction

Physical activity (PA), diet, and sedentary behavior (SB) plays an important role in obesity and healthy-related outcomes in adolescents [1, 2] and may contribute to prevention of noncommunicable diseases in adulthood [3]. Considering a more integrated approach to promote health, researches on behavioral clustering has increased in recent times [4, 5]. Indeed, clustering is a widely used technique in data-driven applications for identifying patterns in datasets. Specifically, two-step cluster analysis is a statistical method that can automatically identify similar clusters of people within data sets using k-means clustering algorithm [6]. Previous reviews [6,7,8] have presented different cluster behaviors among adolescents, however, few studies consider identify how these synergic obesogenic behaviors cluster in different stages of adolescence; since these, modifiable behaviors can change with age increases [9] or to persist into adulthood [10, 11].

Is known that PA, diet, and SB in its unhealthy form tend to cluster in certain subgroups [4, 6]. Studies have been reported some types clusters according to age, like as in Australian children three clusters (“most healthy”, “energy dense consumer who watch TV” and “high SB with low moderate to vigorous PA”) were found in all age groups (5–6 and 10–12 years old) [6]. Two similar lifestyles patterns, “snacking and sedentary” and “varied food and PA”, were identified in French children aged 3–6 and 7–11 years old [12]. However, older adolescents were significantly more likely to be in clusters characterized with low levels of PA and SB [13] since this, younger adolescents (12.5–14.99 years old) were more likely to be in healthier cluster behaviors as compared to older adolescents (15–17.5 years old) [14]. Understand deeper how these behaviors interact differ according to age may assist the targeting of public health initiatives and then as such interventions that stimulate a positive pattern to produce health benefits.

Besides behavior clusters in different ways in adolescents its can differ according to sociodemographic indicators (e.g., sex and maternal education) [6, 15]. For instance, younger adolescent boys and girls were allocated in healthier clusters compared to the older ones [16, 17]. Studies comparing clusters between sexes, have shown that a higher proportion of boys were observed in clusters with higher levels of PA [16, 18, 19], in contrast to girls [13, 18, 19]. On the other hand, girls appeared in a greater proportion than boys in clusters characterized with better quality of diet [14, 20] and high time spent on screen activities [21]. In addition, previous studies observed an association between clusters and low education level of adolescents’ parents [6, 22, 23].

Furthermore, clusters considering PA, diet, and SB have been associated according to social, economic, and cultural aspects, which affect individual behaviors differently, however are a lack of evidence regarding factors that influence clustering behaviors [9]. Thus, to identify adolescents clusters according to age groups in low and upper-middle income countries, like Brazil, is needed [7]. A study pointed that Brazilian adolescents with higher maternal education levels were more likely to be in the healthiest clusters (e.g., Health-promoting PA and diet) [24]. Thus, this study aims to examine clusters of PA, diet, and SB according to age groups and its association with sociodemographic indicators in a representative sample of Brazilian adolescents.

Methods

Study design and population

Data were derived from the National Adolescents School-based Health Survey (PeNSE), a cross-sectional study conducted in 2015, whose main objective was to obtain lifestyle data of adolescents enrolled in private and public schools in Brazil. The aims, methods, sample, and ethics procedure have been described in detail previously [25]. Briefly, PeNSE is part of Brazil’s surveillance of risk and protective factors for chronic diseases. It was supported through partnerships with Brazilian Institute of Geography and Statistics (IBGE), Brazilian Ministry of Education, Brazilian Ministry of Health, Health Surveillance Secretariat, World Health Organization, Centers for Disease Control and Prevention of the United States, and Pan-American Health Organization. This study was submitted and approved by the National Research Ethics Commission (CONEP) number 1.006.467/2015.

This study used PeNSE sample two (adolescents enrolled from the 6th grade of elementary school to the 3rd grade of high school). Each geographic stratum (five macro-regions) was one unit; and in each macro-region, schools and classrooms were dimensioned and selected in advance as conglomerates. The size of the sample was calculated to provide estimates in each of the geographical strata. Further details on the sampling design and weighting can be found elsewhere [25]. All adolescents present in class were invited to participate in the study. Once they agreed to participate, they were asked to respond to an electronic questionnaire on a smartphone during regular school hours. To monitor main indicators, the questionnaire was divided into steps that included household and school contexts; eating habits; PA and body image; use of psychoactive substances; violence; accidents; and sexual behavior. All procedures were carried out by previously trained IBGE technicians for adjustments and adequacy. Briefly, IBGE state supervisors were trained so they could replicate the same training course to the interviewers, in each federative unit. The main objective of this strategy was to ensure the quality and standardization of the procedures for data collection [25]. The measures of interest were previously validated [26].

Cluster variables

Physical activity

Leisure time PA was accessed using the question: “In the past 7 days, excluding the physical education class, how many days did you partake in physical activities like sports, dance, gym exercises, combat sports, or any other activity? The answers ranged from none to seven days in a week.

Sedentary behavior

SB was accessed with the question “in a regular day, how much time do you spend watching television, playing videogames, talking with friends, or others sitting activities?” The responses ranged from one to nine hours a day.

Dietary-patterns

Seven questions were used to assess diet. Adolescents reported how often (one to seven days) during the past week they had eaten fruits, green salads or vegetables, deep-fried empanadas, candies, soda, ultra-processed food and fast foods. To identify dietary patterns Principal Component Analysis (PCA) was performed. The Kaiser–Meyer–Olkin (KMO) test of sampling adequacy and Bartlett's Test of Sphericity were used to assess the suitability of the data for PCA, considering values > 0.60 and p < 0.05, respectively [27]. The number of components were decided with the Kaiser criterion and the screen plot (eigenvalues > 1) [27]. Factor loadings represent the correlations of each predictor indicator with the dietary pattern score. Higher absolute values of factor loadings indicate that the predictor contributes the most to the construction of a particular component. The dietary components (patterns) were named according to the characteristics of the behaviors retained in each of them (Supplementary file 1). Two components were identified: 1) F&V (fruits and green salads and vegetables); and 2) Snacks, salt and beverages (SSB) (deep-fried empanadas, candies, soda, ultra-processed food and fast foods). The food items in each component were scored and divided by the number of variables present in each component (2 and 5) resulting in two dietary patterns F&V and SSB.

Sociodemographic variables

Sex and maternal education were accessed through a questionnaire. Maternal education was accessed with the question “What is your mother’s level of education?” and the possible answers were categorized as low (non-educated, elementary school); medium (high school); high (graduated, higher education) and unknown. Maternal education was chosen since is the indicator most present in clusters studies when made comparisons with sociodemographic variables [13].

Statistical analysis

Only adolescents with complete data in all the variables of the analyses were considered eligible for this secondary data analysis. The descriptive data of the sociodemographic characteristics were presented in absolute and relative frequencies, with respective confidence intervals (95% CI). Two-step cluster analysis was performed using SPSS for Windows (version 23 SPSS Inc.; Chicago, IL, USA) for each age groups (11–12, 13–15, 16–17, and 18–19 years). The two-step cluster analysis is a scalable cluster analysis algorithm that was designed to manage large datasets. The analysis has two steps: pre-clustering and clustering [28]. In the pre-clustering step, all the cases in the data are scanned and the log-likelihood distance between them is measured to determine whether they are going to form pre-clusters based on some threshold distance criterion. In the second step, the subclusters resulting from the pre-clustering step are clustered into the optimal number of clusters using an agglomerative clustering algorithm. The subclusters are regarded as single cases and merged into one cluster by satisfying the minimum distance in (1) until all data are placed within a single cluster [28].

Four variables were considered in cluster analyses: (1) number of days for leisure-time PA; (2) number of days a week when F&V were eaten; (3) number of days a week when SSB were eaten; and (4) number of hours spent in SB. The best combination of a low value in Schwarz’s Bayesian Information Criterion (BIC) were used to define the number of clusters (Supplementary file 2). The log-likelihood ratio was used to measure the distance between the individual observations in the grouping variables. Leisure-time PA, SB, F&V, and SSB were analyzed by one‐way ANOVA to compare between the clusters (Actives, All-day sitters, Inactive 1, and Inactive (2) and the age groups (11–12, 13–15, 16–17, and 18–19 years old). To examine the associations between sociodemographic variables and clusters the Rao-Scott Chi square test, performed in Stata statistical software, version 15 (Stata Inc., College Station, TX, United States of America) was used. Partial eta-square effect sizes were also reported, considering small (≤ 0.05), moderate (0.06–1.3), and high (≥ 1.4) values of effect size [29]. Methods for complex analysis was used to incorporate strata, conglomerates, and sample weight. A significant level of p < 0.05 was assumed.

Results

Sample characteristics

A total of 16,522 adolescents were evaluated. Boys and girls presented almost even split (boys 50.1%); and 43% of the adolescents were in the 13 to 15 age group. More than 27% had mothers with low educational levels (Table 1).

Table 1 Sample characteristics, National Adolescents School-based Survey—PeNSE (sample 2), Brazil, 2015 (n = 16,522)
Full size table

Cluster solutions

A total of 16,336 adolescents were eligible for cluster analysis, while 220 with incomplete data were excluded. Information about the best fit models are presented in Supplementary file 2 and comparison of the four age groups according to the clusters solution in Table 2. Four clusters in adolescents were identified to 11–12 age group; and three clusters comprising the 13–15, 16–17, and 18–19 age groups. Clusters were named according to the most accentuated/highlight behaviors, for example, cluster Actives were characterized with high number of days per week practicing PA. Highlight behaviors can be easily observed in Fig. 1. Silhouette coefficient indicating cohesion and separation were used to identify the quality of cluster solutions in each of the age groups and all values were 0.40, indicating fair models.

Table 2 Comparison of the four age groups according to the clusters solution for physical activity, diet, and sedentary behavior in Brazilian adolescents (n = 16.336). National Adolescents School-based Survey—PeNSE, Brazil, 2015
Full size table
Fig. 1
figure 1

Cluster solutions for physical activity, diet, and sedentary behavior according to age group in Brazilian adolescents, National Adolescents School-based Survey—PeNSE, 2015 (n = 16,336). Note: PA physical activity; SB sedentary behavior; F&V fruit and vegetables; SSB snacks, salt and beverages

Full size image

Information on the characteristics and comparisons between the clusters according to the age group can be found in Table 2. Leisure-time PA, SB, F&V, and SSB were differentiated among clusters using multivariate analysis (p < 0.001). Leisure-time PA and SB contributed most to the distinction between clusters (Partial eta-squared largest effect size). Four different cluster were observed: The Actives; the All-day sitters; the Inactive 1 (with moderate consumption of F&V); and the Inactive 2 (with small consumption of F&V and SSB). The Actives cluster were present in all age groups and was characterized by more than 5 days/week of leisure-time PA and more than 4 days/week of F&V consumption. However, this group accumulated moderate levels of SB in daily life. The All-day sitters cluster too, were present in all age groups; and included high levels of SB (more than 7 h/day, except for the 18–19 age group); small participation in leisure-time PA; and consumed SSB on more than 3 days/week. The Inactive clusters were most prevalent in all age groups (except among the 18–19 age group) and were mainly characterized by lack of leisure-time PA (Fig. 1). The Inactive clusters were different from the All-day sitters, in that their SB was less than 3 h a day. More than 55% of 11–12 years old adolescents had inactive clusters. Despite similarities, the 18–19 age group seemed to differ in some aspects. It had a smaller proportion of the Actives cluster and the most prevalent cluster was the All-day sitters. However, SB was lower than in the other age groups (less than 5 h/day); and this cluster did not present low to moderate F&V consumption as observed in other age strata (Fig. 1).

Differences in sociodemographic characteristics according to cluster

Association between clusters and sociodemographic characteristics are shown in Table 3. In all age groups there were associations between sex and clusters. The All-day sitters, Inactive 1 and Inactive 2 clusters had a majority of girls; while the Actives cluster across all age groups had a majority of boys. Maternal education had an association with all age groups. In all clusters, except in the Actives, the 11–12 age group were in a majority in the category where maternal education was unknown. There were no significant differences between the medium, low, and unknown maternal education categories in the 13–15 and 16–17 age groups. In addition, in all clusters, the 18–19 age group were in a majority in the low maternal education category.

Table 3 Cross-sectional differences between sociodemographic characteristics and clusters according to adolescents’ age groups (n = 16,336). National Adolescents School-based Survey—PeNSE, Brazil, 2015
Full size table

Discussion

As far as we know, this is the first study to use cluster analysis to identify how healthy and unhealthy behaviors (including PA, diet and SB) are displayed in Brazilian adolescents according to age groups (11–12, 13–15, 16–17, and 18–19 years old) and its association with sex and maternal education. While similar studies that had used the cluster analysis approach in different age ranges were found in other populations and countries [12, 23, 30, 31], the synergic obesogenic behaviors in different stages of adolescence had been little explored. Four clusters were identified: Actives, All-day sitters, Inactive 1 (with moderate consumption of F&V); and the Inactive 2 (with small consumption of F&V and SSB). A large proportion of adolescents had an inactive cluster in all age groups. In addition, in all age groups, boys were more likely to be in the Actives cluster than girls and associations were found between clusters and educational levels. These results can contribute to create strategies to reduce SB, increase PA and stimulate a healthy diet concurrently in adolescents.

The Actives cluster was observed to be the healthiest and appeared across all age groups. They were characterized by high levels of PA and F&V consumption with a moderate time spent in SB. It seems that adolescents with higher levels of PA trigger the adoption of other healthy behaviors. This behavioral synergy that was observed was consistent with previous researches [6, 12, 30, 31]. Results similar to ours were found in studies conducted in Australia, France, and Spain that investigated clusters of obesogenic behavior in samples of younger (including children) and older adolescents. They also found that the healthiest cluster had spent the highest time in moderate-to-vigorous PA and had adopted other healthy behaviors (e.g., low consumption of energy-dense foods, spent least time watching TV, and had a high consumption of healthy foods) [12, 23, 30].

In all age groups, a pattern characterized by high time spent in SB with high consumption of SSB was observed in the All-day sitters cluster; and similar results were observed in others studies [14, 15]. Since this clusters was inductive of a moderate consumption of F&V; in this sense, compensatory health beliefs may have appeared for this group because of the adolescents rationalizing and believing that consuming healthy foods could compensate for SB and an unhealthy diet [32]. A combination of healthy and unhealthy behaviors also appears in the Inactive clusters and a good diet pattern had a synergic interaction with low levels of PA. A study indicated that calorie restriction is more effective than exercise programs for weight loss [33]. In addition, since adhering to healthy diet habits and PA involve discomfort and efforts [34]; perhaps eating healthy foods demands less effort than practicing PA. Antagonistic behaviors (positive diet and other unhealthy obesogenic behavior) are consistent with others studies [12, 14, 15].

The unhealthiest cluster (Inactive 1) was the most prevalent in the 11–12, 13–15, and 16–17 age groups. In addition, the present study showed that all clusters presented at least one unhealthy behavior in all age groups, even in the healthiest cluster, which corroborates with others studies [6, 12]. A fact worth highlighting is that adolescents in all cluster behaviors and age groups spent more than 2 h in SB. Perhaps, right from early adolescence, greater attention should be given to promoting synergistically healthy behaviors that also focus on reducing sedentary time [3] once, prolonged sitting have been related as a risk factor to chronic diseases such as mental health, cancer, obesity and cardiovascular disease [35]. Thus, considering cluster characterized with at least one unhealthy behavior it supports the need of targeting increase PA, reduce SB and improve the quality of diet in adolescents at the same time. Also, distinct strategies should be considered in accordance with the characteristics of adolescents in each cluster; for example, adolescents in the Actives cluster can be provided with actions to be taken to decrease SB and SSB consumption.

Considering the differences between age groups, previous studies reported a higher prevalence of younger adolescents in clusters with high levels of PA [6, 13, 14]. While the present study is consistent with literature, the Actives clusters was most prevalent in the 11–12 and 16–17 age groups. In addition, older adolescents had a higher presence in the Inactive 1 cluster that corroborated a study [15], which found that older adolescents had a higher presence in clusters considered unhealthy (e.g., low PA, diet quality, and time in SB). It is noteworthy that the proportion of adolescents in the Inactive 1 cluster increased with advancing age, and there was a slight decrease in the 18–19 age group. While we were not able to infer a causal relationship between health risk clusters and age groups; more than 70% of the sample in all age groups seemed to be at a risk for obesity.

As consistent with previous studies [6, 15,16,17,18,19]; boys in all age groups, were more likely to be in the Actives cluster than girls; whereas girls were more likely to be in the All-day sitters, Inactive 1, and Inactive 2 clusters than boys. There is, therefore, a need for intervention programs that would target the reduction in SB and increase of PA for Brazilian girls in all age groups. Investigate how these obesogenic behaviors are grouped according to different age groups, in a representative sample, was reported as a gap in a previous review [6]. It is worth mentioning that clustering of people who share similar characteristics is a concept that has been successfully applied to understanding the relationships between different lifestyle behaviors [6]. Since these obesogenic behaviors combine in complex ways have a cumulative effect on the development of overweight, obesity, and/or non-communicable diseases, this has important implications for public health because understanding which behaviors need to be targeted simultaneously and in whom obesogenic behaviors cluster together can be used to assist in the elaboration of targeted obesity prevention initiatives [6]. Tailoring interventions for appropriately targeted and effectively bring about multiple behavior changes, may be more cost-effective and maximize reach to those most in need [6].

There were associations between clusters and educational levels in all groups, except the 11–12 age group, which perhaps, occurred because of their difficulties in reporting about their mothers’ schooling. In our study, the Actives group had a higher proportion of adolescents with low maternal education; which was consistent with studies from Brazil and Portugal [36] and inconsistent with previous studies conducted in Brazil [37] and high-income countries [6, 13, 20, 23]. Indeed, health-risk clusters were associated with low maternal education as was previously observed in the literature on younger and older adolescents [6, 13, 20, 23]. It would expect that parents with a higher level of education are generally better informed and are more aware of the importance of PA on the health status of their children and adolescents [38]. Then, better-educated parents more likely to encourage their children to engage in some form of sports or PA. The educational level reflects parental knowledge on healthy lifestyle and influence in youth engagement in healthy behaviors [39]. However, even that parents recognized the importance of PA, barriers like lack of environmental barriers (e.g., safety issues, heavy traffic, lack local and unsafe neighborhood) and personal (e.g., prefer SB activities as screen-based technologies) are barriers to PA [39]. As these screen-based technologies are probably few accessible to families with low levels of education, adolescents may become more physically active, especially outside the home. However, caution is needed to extrapolate these results since we do not test this hypothesis. In addition, differences, particularly in the Actives, call attention to Brazil’s high social inequality [40], where adverse social conditions seem to impose a negative synergy in adolescents’ behaviors, even in healthy groups.

It needs to be acknowledged some limitations since the measures were self-reported, it could have resulted in participants’ answers being influenced by social desirability and memory bias. Moreover, since the design included the biases of cross-sectional cluster analysis where the causal relationship cannot be assumed; the assumed similarities and differences between the age groups cannot be expected only as a consequence of early age behavior antecedents. The transition and stability of clustering behavior over time should be distinguished through more robust research designs considering PA, diet, and SB using recalls and objective measures. Despite these limitations, this study was a nationally representative sample of Brazilian adolescents, utilizing a previously validated questionnaire, applying a sophisticated statistical analysis based on a data-driven procedure, and it facilitated the investigation of adolescents’ groups with similar behaviors based on age in an upper-middle-income country.

Conclusions

Four clusters of obesogenic behaviors were identified in Brazilian adolescents and its formation was similar across different adolescent age groups. All the age groups were highly exposed to an unhealthy cluster and even those with a healthy cluster were exposed to higher SB. Despite the fact that unhealthy behavior declines in older age groups; this proportion reaches a maximum drop of one third. We considered risk-groups of girls and we suggested intervention including more than one behavior to this group. Finally, our study suggested a reduction of SB in clusters at all ages and both sexes; an increase in PA in girls, and in adolescents aged 13–15 and 18–19 years.