Latent class analysis of obesity‐related characteristics and associations with body mass index among young children

Summary Objective Identifying how obesity‐related characteristics cluster in populations is important to understand disease risk. Objectives of this study were to identify classes of children based on obesity‐related variables and to evaluate the associations between the identified classes and overweight and obesity. Methods A cross‐sectional study was conducted among children 3–11 years of age (n = 5185) from the TARGet Kids! network (2008–2018). Latent class analysis was used to identify distinct classes of children based on 15 family, metabolic, health behaviours and school‐related variables. Associations between the identified latent classes and overweight and obesity were estimated using multinomial logistic regression. Results Six classes were identified: Class 1: ‘Family and health risk behaviours’ (20%), Class 2: ‘Metabolic risk’ (7%), Class 3: ‘High risk’ (6%), Class 4: ‘High triglycerides’ (21%), Class 5: ‘Health risk behaviours and developmental concern’ (22%), and Class 6: ‘Healthy’ (24%). Children in Classes 1–5 had increased odds of both overweight and obesity compared with ‘Healthy’ class. Class 3 'High risk' was most strongly associated with child overweight (odds ratio [OR] 1.9, 95% confidence interval [CI] 1.2, 3.2) and obesity (OR 3.3, 95% CI 1.7, 6.7). Conclusions Distinct classes of children identified based on obesity‐related characteristics were all associated with increased obesity; however, the magnitude of risk varied depending on number of at‐risk characteristics. Understanding the clustering of obesity characteristics in children may inform precision public health and population prevention interventions.

Obesity-related variables defined broadly to include both risk factors (e.g., poor diet, low physical activity and family history of obesity) and complications or consequences of obesity (e.g., increased cholesterol, blood pressure and behavioural problems), are generally considered in isolation. A 2014 review paper identified 18 studies on the clustering of diet and physical activity in children, and only three studies used latent class analysis (LCA). 7 LCA is a data-driven analytical method that groups heterogeneous populations into homogeneous groups based on categorical indicator variables and has some advantages over other cluster analysis methods. 8,9 In addition to the three studies [10][11][12] identified in the previous review, three other studies [13][14][15] that used LCA in children were identified-all of these studies were conducted in children >9 years of age and focused primarily on diet and physical activity. These six previous LCA studies [10][11][12][13][14][15] each identified between three to six specific classes (e.g., healthy, sedentary, physically active and health risk behaviour), and each of these classes were differentially associated with overweight and obesity.
Thus, the overall aim of this study was to identify distinct classes of children, as young as 3 years of age, based on a comprehensive range of obesity-related variables, including family, metabolic, health behaviours and developmental concerns. The primary objective of this study was to identify distinct groups of healthy children 3 to 11 years of age based on obesity-related variables, overall and stratified by sex.
The secondary objective was to evaluate the association between the identified latent classes and BMI categories (normal weight, overweight and obesity).

| Study design
A cross-sectional study was conducted using data from The Applied Research Group for Kids (TARGet Kids). TARGet Kids is a primary care research network in Toronto, Canada. Children <6 years of age were recruited between 2008 and 2018 during well-child visits at primary care paediatric and family physician practices in the Greater Toronto Area and are followed annually. 16 Exclusion criteria at enrolment were <32 weeks gestational age, growth-restricting health conditions such as cystic fibrosis or failure to thrive, severe developmental delay, or other chronic conditions (not including asthma and high functioning autism) and non-English speaking parents. 16 For this cross-sectional study, children with a visit between 3 to 11 years of age (36 to <144 months) were included. Children <3 years of age were excluded because blood pressure was not measured in this age group. 16 Children ≥12 years of age were excluded because there were relatively few children of this age in TARGet Kids, and this study was specifically interested in younger children as obesity interventions may be more successful at younger ages. 17 For children with multiple visits in the 3-to 11-year age range, the first visit with a blood sample was selected; for children who never provided a blood sample, their first visit with questionnaire data within this age range was used.
Trained research assistants recruited study participants and physical measurements, including anthropometrics and blood pressure, blood samples (nonfasted) and parent-completed nutrition and health questionnaires were collected at each visit. 16 16 The TARGet Kids! cohort study is registered at www.clinicaltrials.gov (NCT01869530).

| Measurement of body mass index z-scores
Height and weight anthropometric data were collected by trained research assistants using a precision digital scale (seca, Germany) for weight and a stadiometer (seca) for measuring standard height. BMI was calculated by dividing weight (in kilogrammes) and height (in metres squared) 16 and BMI z-scores (zBMI) were calculated using the World Health Organization (WHO) growth standards for children under five and growth reference for children over five as recommended for the Canadian population. 18,19 For the purpose of this study and to conduct analysis combining all children from 3 to 11 years of age, the labels recommended for children over 5 years of age were consistently applied to all the children in this study: zBMI ≤ 1 was defined as 'normal weight'; 1 < zBMI ≤ 2 was defined as 'overweight'; zBMI > 2 was defined as 'obesity'. 18  Family characteristics included family history of cardiometabolic disorder defined as 'yes' if a mother, father or siblings reported diagnosis of heart disease, hypertension, high cholesterol or diabetes versus 'no' if none reported. Parental BMI was calculated based on measured height and weight of either the mother or father collected by the research assistant at the child's primary care visit. 16 For this study, parental BMI was 83% from mothers and 17% from fathers.

| Measurement of obesity-related characteristics
Parental BMI ≥ 25 was defined as overweight or obesity versus BMI < 25. 20 Low income was categorized as median neighbourhood income <$50,000 per year versus ≥$50,000 per year. Low income was defined as <$50,000 based on the low-income measure threshold for before-tax income for a family of four in Toronto, which was $51 031 in 2015. 21 Cardiometabolic characteristics for children included high systolic blood pressure (SBP) or diastolic blood pressure (DBP), measured by trained research assistants according to recommendations, 22 and dichotomized as high versus low based on the age-specific guidelines (SBP or DBP ≥ 90th percentile for sex, age and height). 23 Abnormal cholesterol was defined from nonfasting blood samples as described previously. 24 The following cut-offs were used to define abnormal lipid levels: HDL ≤ 1.17 mmol/L, LDL ≥ 2.85 mmol/L), non-HDL ≥ 3.11 mmol/L and triglycerides (0-9 years ≥0.85 mmol/L, TG (10-19 years) ≥ 1.01 mmol/L). 25 Health risk behaviours included parent-reported physical activity, screen time, sleep and sugar-containing beverages (SCBs). The measurement of each of these variables is described in more detail elsewhere. 16 Physical activity was measured as weekday freeplay, and children were classified as not meeting physical activity recommendations if they had less than 180 min/day of unstructured free play (under 5 years), or less than 60 min per day (5 years and over). 26,27 Sedentary behaviour was estimated by screen time in reference to how much time the child spent awake using a TV, DVD, video game or mobile device on a typical weekday. Children under five that reported greater than 60 min per day of screen time, and children 5 years and older that reported greater than 120 min per day were identified as having high screen time. 27,28 Parent-report of child sleep duration was dichotomized as not meeting recommendations if sleep duration was <10 h for children 3-5 years or <9 h for children 6-13 years, versus meeting or exceeding recommendations. 29 SCBs were calculated as the number of cups of 100% juice, sweetened drinks and soda or pop the child consumed on a typical day. Based on recommendations outlined by the American Academy of Pediatrics, children were categorized as SCB intake not meeting recommendations if they consumed >0.5 cup/day for age less than 3, >0.75 cup/day for age 4-6, and >1 cup/day for ages greater than 7. 30 Developmental concerns included children that were reported to have been diagnosed with attention deficit hyperactivity disorder, autism, learning problem or developmental delay on the Nutrition and Health Questionnaire (NHQ) that were identified as having developmental concern. Children were identified as having school concerns if parents reported that their child's school expressed concerns about the child relating to the following: speech and language, learning, attention, behaviour, social relationships, physical coordination, fine motor coordination or self-help skills and independence. Children were identified as requiring extra resources if parents reported the provision of one or more of the following extra resources from their child's school: speech and language therapy, occupational therapy, educational assistance or other extra resources provided at school.

| Statistical analysis
Descriptive analyses of covariates including sex, age and weight category were examined for participants, as well as the frequency of  distributions of the item response probabilities were evaluated, and the identified classes were named based on which characteristics were more likely to be exhibited by members of the class. Participants were assigned to the class in which they had the highest probability of membership; that is, they exhibited the traits that are representative of that class. The associations between the identified latent classes and zBMI categories were then evaluated using multinomial logistic regression adjusted for child age and sex. Age and sexadjusted odds ratios (ORs) and 95% confidence intervals (CIs) were reported. All analyses were conducted overall (among males and females combined) and sex stratified.
Under the assumption that missing data was missing at random (MAR), the model was fit using an Expectation-Maximization (EM) approach, where all available data was used to estimate classes and assign the posterior probabilities of belonging to each class for each individual. 9 All individuals were assigned to a class using this approach, and this class was then used in the regression analysis resulting in no missing data for the final regression analysis.

| Participant characteristics
A total of 5185 children between 3 and 11 years of age were included in this study; 74% of children were in the 3-to <6-year age range. As described in Table 1, 2471 (48%) children were female, and 2714 (52%) were male. The mean age was 61 months (5 years) with a standard deviation of 21.1 months. Based on zBMI, 81% of children were considered 'normal' weight; 14% had overweight, and 5% had obesity; the proportion of children with obesity was slighter higher in males than females. The frequency of all obesity-related characteristics included as indicators in the LCA are also provided in Table 1. About half of the children (48%) did not provide a blood sample and thus had no lipid measures.

| Latent class analysis
For the primary analysis, a six-class model was found to be the optimal model as it had lower cAIC and BIC values compared with the other models (Supporting Information Table S1). Response probabilities for each of the 15 indicators by class are presented graphically in Figure 1 (and numerically in Table S2). The 6 identified classes were time, extra school resources and school concern. In comparison, Class 5 was defined as Health risk behaviours and developmental concern only with high screen time, extra school resources and school concern.
F I G U R E 1 Graphical display of item response probabilities for each class (see Supporting Information Table S2 for numerical display)

| Association of the latent classes with body mass index z-score categories
The distribution of zBMI categories across the six identified latent classes is presented in Figure 2. The proportion of children with overweight and obesity was highest in Class 3 High risk at 19% and 9%, respectively, and was lowest in Class 6 Healthy at 11% and 3%, respectively.

| Sex-stratified analysis
Results of the sex-stratified analyses revealed that a four-class model fit the data best for the LCA in both males and females separately school, whereas in females this class included 22% and was defined as by high parent BMI, low neighbourhood income, high triglycerides, screen time and SSB (not school resources or concern). Class 3 was the Healthy class observed in 39% of males and 35% of females; in males, this class exhibited no high risk characteristics (except low physical activity which was observed in across every class as discussed previously), but in females, this class also had relatively high probability of high triglycerides. Class 4 in males (23%) was defined as High triglycerides only, whereas Class 4 in females (20%) was defined as 'Developmental concern' with a high probability of both extra resources required at school and school concerns.
F I G U R E 2 Distribution of body mass index zscore (zBMI)-defined weight categories by latent class T A B L E 2 Adjusted odds ratio (OR) and 95% confidence intervals (CIs) of having overweight, and obesity by latent class for the total sample   were associated with increased odds of overweight and obesity. Class 4 Developmental concern was not associated with overweight or obesity in females.

| DISCUSSION
Results of this study suggest that there are distinct classes of young children defined based on presence of obesity-related risk factors, and several of these classes were differentially associated with odds of overweight and obesity. A total of six classes were identified, and only one class (Class 6 Healthy), which included 24% of children, was defined by the absence of any obesity-related risk factors (except for low physical activity that was highly prevalent across all classes). The other five latent classes, which included 76% of the children, all exhibited at least two obesity-related characteristics. Of concern were the 6% of children identified in Class 3 with multiple obesity-related characteristics, including family, metabolic, health risk behaviours and developmental concerns.
Each of the identified classes, compared with the healthy class as reference, was strongly associated with increased odds of overweight, and obesity in a consistent dose-response manner with much stronger associations for obesity than overweight. However, the magnitude of the strengths of the associations varied by class.
Classes 1-3 defined as Family and health risk behaviours, Metabolic risk and High risk were significantly associated with substantial increases in odds of both overweight and obesity, with ORs ranging from 1.6 to 3.3.
Sex-stratified analyses revealed that a LCA model with only four classes fit the data best for both boys and girls, compared with the six-class model identified overall. Some sex-differences were observed, but there were also many similarities between the four classes that were identified separately in males and females. Of the six previous LCA studies, [10][11][12][13][14][15] four studies evaluated sexdifferences. [11][12][13]15 Two of these studies 11,12 found that there were less girls in the high physical activity class compared with boys, while one study conversely found that girls were less likely to be in the high sedentary group compared with boys. 10 This is consistent with a recent review that found that females were more likely to be in classes defined by low physical activity. 7 Of the 15 indicator variables that were evaluated based on a priori hypothesis, the following five variables had little evidence of an association with any class membership: family history of cardiometabolic disorders, high systolic or diastolic blood pressure, abnormal HDL, sleep duration below age-specific recommendations and parent-reported developmental concern. In this study, low physical activity was also not a helpful indicator in determining the classes T A B L E 3 Adjusted odds ratio (OR) and 95% confidence intervals (CIs) of having overweight, and obesity by latent class for males and females separately Population-based interventions for the prevention 35

DATA AVAILABILITY STATEMENT
Data are available upon request by contacting www.targetkids.ca/ contact-us/. The full data are not freely available to respect the confidentiality of our participants, ensure data integrity and avoid scientific overlap between projects. Once initial contact has been made, we request a short research proposal that will be subject to review by the TARGet Kids! Scientific Committee and approval by institutional review boards.