Study population
This is an observational study nested within the GUSTO birth cohort study. To provide context, GUSTO is an on-going multi-ethnic birth cohort which commenced in 2009 to investigate the associations between early life factors and the health and development outcomes of children; details of the cohort have been published previously (20). In brief, pregnant women were recruited from two major public maternity units in Singapore: KK Women’s and Children’s Hospital, and National University Hospital. Women who were aged at least 18 years, of Chinese, Malay, or Indian ethnicity with a same-ethnicity partner, Singapore citizens or permanent residents, and were intending to deliver in one of the two above-mentioned maternity units and remain in Singapore for the following 5 years were eligible. Of 1450 women recruited, 1219 babies (including 10 twin-births) were born and followed up regularly. For the current study, data from child participants at 10 years of age were analysed. This study was approved by the National Healthcare Group Domain Specific Review Board and the SingHealth Centralised Institutional Review Board, and all participants provided written informed consent at enrolment.
Data collection
Lifestyle information of children
My E-Diary for Activities and Lifestyle (MEDAL), a web-based lifestyle assessment application developed to collect time-use information was administered to collect self-reported lifestyle behaviour information. Details of the development, usability, and validity of the MEDAL application have been published (21,22,23).
During the clinic visit at 10 years, a trained researcher explained the MEDAL interface to the participant and allowed him or her to record the activities he or she did and everything he or she ate the day before the day of the clinic visit. The participant could clarify any difficulty they faced when using MEDAL during the clinic visit. Participants were instructed to complete the entry at home, and to record their diet and activities on MEDAL from the time they woke up to the time they went to bed in chronological order over two specified weekdays and two weekend days independently, without assistance from their parents or caregivers. These entries capture information related to their diet and movement behaviours, such as portion size of food and drinks consumed, duration and intensity of activities they engaged in, and whether any other activity occurred concurrently (e.g. screen-viewing during a meal).
The data collected via MEDAL were processed to generate information on averaged daily moderate-to-vigorous physical activity (MVPA), screen-viewing, and sleep, and averaged daily intakes of fruits, vegetables, wholegrains, dairy, sugar-sweetened beverages (SSB), sweet and savoury snacks, fast food, and processed food (i.e. ham or luncheon meat, instant noodles, nuggets, sausages, and meatballs). This was done by dividing total reported time spent in each of these behaviours or intakes by the number of days recorded on MEDAL by each participant (range: 2 to 4 days). These variables were selected in accordance with the Singapore Integrated 24-hour Activity Guidelines for Children and Adolescents, which have been identified as important for optimising the health and well-being of children (24). Engagement in MVPA, screen-viewing, and sleep were expressed in minutes or hours. Intakes of fruits, vegetables, dairy, and SSB were expressed in servings in accordance with the local Health Promotion Board’s quantification of one serving for items in these food groups (25). As what constitutes one serving of wholegrains, sweet and savoury snacks, fast food, and processed food were not available, we presented these variables based on the average daily frequency in which they were consumed.
Outliers were replaced with the maximum values of the acceptable distribution, i.e. value at 99th percentile of the distribution (10): movement behaviour-related behaviours with values above the 99th percentile were replaced for MVPA (207.5 min, n = 3), screen-viewing (13.0 h, n = 3), and sleep (13.0 h, n = 4). Diet-related variables with values above the 99th percentile were replaced for fruits (2.5 servings, n = 3), vegetables (3.3 servings, n = 3), wholegrains (2.3 times consumed per day, n = 4), dairy (1.4 servings, n = 3), SSB (3.8 servings, n = 3), sweet and savoury snacks (3.0 times consumed per day, n = 2), fast food (2.0 times consumed per day, n = 1), and processed food (2.5 times consumed per day, n = 2).
Family sociodemographic, parent- and child related variables
To examine the correlates of children’s lifestyle patterns, we selected variables with guidance from the Family Ecological Model (26) that were available in the GUSTO study and hypothesised to be associated with children’s lifestyles. These variables were grouped as distal, intermediate, and proximal in terms of level of influence on lifestyle patterns respectively (Fig. 1). Details of each variable are available in Supplementary Table 1.
Fig. 1
Conceptual framework of distal, intermediate, and proximal factors examined in association with children’s lifestyle patterns in the present study
In brief, distal factors (i.e. family sociodemographic) included variables such as child’s sex, child’s ethnicity, maternal marital status when the child was 8 years, whether the child had siblings when the child was 8 years, maternal and paternal education level when the child was 5 years, and pet ownership when the child was 7 years, were collected through questionnaires administered by interviewers to mothers of children in the GUSTO cohort between 5 and 8 years.
Intermediate factors (i.e. parenting practices and health behaviours) included variables such as self-reported maternal smoking habits, physical activity, and diet quality (scored using a diet quality index based on completed food frequency questionnaires) when the child was 10, 8, and 6 years respectively, and the identity of caregivers of the child when the child was 10.5 years, collected using and/or derived from questionnaires administered to mothers of the GUSTO child participants. For the present study, we defined caregivers as someone who spent at least two hours with the child in a week and was responsible for certain aspects of the child’s daily routine. We then categorised caregiver information as whether the child was cared for by their parents only, or by a combination of their parents and another caregiver (e.g. grandparents, domestic helper, and others). Mothers reported their own parenting style as well as their perception of their partner’s parenting style using the Parenting Styles and Dimensions Questionnaire Short Form (PSDQ) (27) when their child was 8.5 years. Parental bonding between parent and child, defined as how caring and overprotective the child perceived their mother and father to be, was collected using the Parental Bonding Instrument (PBI) (28) administered to the child when they were 8.5 years.
Proximal factors included variables that had most direct relationship with children’s diet and movement behaviours. This included their weight status based on height and weight measurements obtained during their follow up visit at age 10 years, and contextual activities they engaged in, obtained from the child’s MEDAL entries. For the present study, we classified children as being underweight, of healthy weight, and overweight based on population-derived sex- and age-specific percentiles (29). Contextual activities were defined as activities that happened in specific contexts (e.g. time of day, location, type, or situation), such as active transport (i.e. walking or cycling), leisure sports (i.e. engagement in sports that did not occur in school and were assumed to be for leisure), screen-viewing while travelling or during an eating or drinking occasion, outdoor time, educational activities (i.e. reading, studying, or enrichment lessons that did not occur in school), and breakfast consumption.
Statistical analysis
Participants who recorded at least two days on MEDAL with at least two meals (i.e. breakfast, lunch, or dinner) each day were included for analyses. Due to high homogeneity in variables relating to maternal marital status (95.5% married) and maternal smoking habit (94.5% non-smoker), these variables were excluded from subsequent analysis. Test for differences in characteristics between included and excluded participants were performed using Pearson’s chi-squared test and two-sample t-test for categorical and continuous variables respectively. To provide some description of the lifestyle behaviours of children, we present the mean and standard deviation (s.d.) for 11 lifestyle behaviours, i.e. durations of MVPA, screen-viewing and sleep, and intakes (servings or frequency) of fruits, vegetables, wholegrains, dairy, SSB, sweet and savoury snacks, fast food, and processed food. We further stratified these results by sex and compared girls and boys using two-sample t-tests.
We derived the lifestyle patterns of the participants using principal component analysis (PCA) with varimax rotation on the 11 lifestyle behaviours. Among two methods of rotation we are familiar with, the orthogonal varimax rotation produced slightly more interpretable PCs, thus guiding our current approach. Three orthogonal patterns were retained based on eigenvalues (> 1.0), an inspection of the scree plot, and pattern interpretability (10). Scores for each lifestyle pattern were generated for each participant by summing the 11 input variables and multiplying these by the corresponding loadings of each input variable. A higher pattern score indicates higher adherence to the lifestyle pattern. These patterns were characterised by variables that had loadings greater than 0.25, a cut-off that has been applied and accepted previously (10, 30).
To examine the associations between factors of distal, intermediate, and proximal levels of influence on lifestyle patterns, we first imputed missing distal and intermediate-level factors (no missing proximal-level factors) to maximise power. Linear regression, logistic, ordered logistic, and multinomial logistic methods were used to impute missing continuous, binary, ordinal, and nominal variables respectively, assuming these data were missing at random (MAR). This assumption was ascertained by taking each column with missingness and recoding as “1” for not missing and “0” otherwise. We then regressed each variable on each other using logistic regression where all associations returned significant (p < 0.001), suggesting that our data were MAR. Using the Markov chain Monte Carlo method, 45 datasets were imputed (31). Estimates and standard errors across the imputed datasets were pooled using Rubin’s Rules (32). The 95% confidence intervals were estimated, and p-values were obtained from these pooled estimates. To inform our model building, we performed simple linear regression analyses between each of the exposure variable and each lifestyle pattern and only included variables with p < 0.10 in the subsequent multivariable regression analysis (results are available in Supplementary Table 2); for categorical variables with more than 2 groups, overall p-values were evaluated.
In the multivariable regression analysis, each component of the conceptual framework that returned significant in the univariate analysis was entered into the regression model in a hierarchical manner (Fig. 1) starting with the factors of distal, intermediate, then proximal level of influence on the outcome of interest (i.e. children’s lifestyle pattern at 10 years) (10, 19). The first model (Model 1) regressed distal variables on the lifestyle patterns. The second model (Model 2) further included the intermediate variables in addition to variables that returned significant in Model 1 (at p < 0.05), and the third model (Model 3) further included the proximal variables in addition to variables that returned significant in Model 2 (at p < 0.05). In these multivariable analyses, each variable was interpreted within the first model in which it was included, independent of the significance of the association between the variable and lifestyle pattern in subsequent models. Sensitivity analyses were performed to evaluate if imputation of missing data may have altered our conclusion, once by excluding variables with large proportion of missing data (i.e. variables relating to caregiver information, parenting style, and parental bonding with 40–44% missingness), and a second time by performing the hierarchical regression analysis on participants with complete data.
All statistical tests were performed using Stata Special Edition version 14.2 (StataCorp LP, USA). All evaluations were made assuming a two-sided test at 5% level of significance unless otherwise specified.
