Participants, Design and Setting
Data for this secondary analysis were obtained from the Child and Adolescent Behaviors in Long-term Evolution (CABLE) project. The study commenced in 2001 and followed participants annually until 2016. The CABLE project aimed to investigate the development of healthy behaviors from childhood to adolescence based on a socioecological model. Participants were cluster-sampled from all public elementary schools in Taipei City and Hsinchu County in Taiwan based on a list of names provided by the Ministry of Education in 2001. Nine schools from each area were selected. Two cohorts, first and fourth graders in each school, were followed. Further details regarding the sampling process, sample size calculations, and instrument development are described elsewhere [25]. Signed informed consent was provided by the parents or primary caregivers of all participating students.
In the CABLE project, data were collected from students and their parents. Regarding the students, first to ninth graders completed their questionnaires in the classroom under the supervision of trained instructors. From the 10th grade onwards, students were interviewed individually by trained interviewers. In the first four annual assessments, primary caregivers also completed parent-version questionnaires, which inquired about the parental education level, family income, marital status, and parental behaviors. The project was approved by the Human Research Medical Ethics Committee of the National Health Research Institutes in Taiwan (EC9009003).
In this study, we analyzed data from the second cohort (fourth graders), who were followed annually from 2001 to 2013 (aged 9 to 21 years). The completion rate ranged from 81.6% to 98.1% during the study period. The final analytical sample comprised 2,072 participants (1,075 males and 997 females) who were enrolled in 2001 and provided at least one wave of data on the measure of ECE. Overall, 46.53% of participants provided ECE data in all 13 waves, and 13.61%, 7.53%, 7.92%, 6.03%, 4.01%, 2.12%, 3.47%, 1.69%, 1.30%, 1.88%, 0.97%, and 2.94% of participants provided ECE data in twelve waves to one wave, respectively. Reasons for the attrition rate included moving, refusal to be interviewed, health issues, and loss of contact.
Measures
ECE. The measurement of ECE was assessed annually by asking participants “Not counting normal physical exercise courses in school, have you exercised/participated in sports in the past week?”. The possible responses were “1 = never,” “2 = rarely (one or two days),” “3 = often (three to six days),” and “4 = always (every day).”
Related factors. This study examined the sex-specific effects of several individuals and parental factors on the trajectories of ECE. Specific measures for each factor were as follows.
Individual factors. All individual factors were measured from 2001 (aged 9 years) to 2006 (aged 14 years), and scores at each wave were averaged to reflect the mean levels of each factor. BMI was measured by self-reported weight (kg) divided by the square of height (m2). Body dissatisfaction was measured using four items, namely self-perceived satisfaction with appearance, figure, height, and weight [26], all of which were rated on a 5-point scale ranging from 1 (very satisfied) to 5 (very unsatisfied). The body dissatisfaction score was obtained by summing the four items, with higher scores indicating a higher level of body dissatisfaction (Cronbach’s alpha of 0.68 to 0.72). Stress was assessed using questions adapted from a previous study [27] that asked participants to rate their perceived levels of stress from eight different sources (e.g., academic performance, relationships with friends, and relationships with parents). All items were rated on a 5-point scale ranging from 0 (absolutely no stress) to 4 (extremely high-level stress) and were summed to create an overall stress score (Cronbach’s alpha of 0.79 to 0.84). Screen behavior was assessed using two items: “Have you used a computer or played video games continuously for more than two hours in the past week?” and “Have you watched television continuously for more than two hours in the past week?” All responses were rated on a four-point scale ranging from 1 (never) to 4 (every day). The level of screen behavior was computed by summing the frequencies of these two items.
Parental factors. Parental exercise was measured by one item asking parents “Have you exercised in the past week?” and was dichotomized as “regular exercise” (defined as at least one parent had often or always exercised) and “irregular exercise or none exercise.” Parental screen behavior was measured at each wave from 2001 to 2004 by using the same two items that assessed students’ screen behavior. Response categories for these two items, which ranged from 1 (never) to 5 (always), were summed and averaged across 4 years to reflect the level of parental screen behavior.
Control variables. Parental education, household income, and marital status were included as control variables based on the previous evidence of associations between these social demographic variables and exercise [18]. All control variables were collected between 2001 and 2004 and reported by parents or primary caregivers. Parental education was measured as the highest level of education attained by either parent and was coded as low (less than or equal to 12 years) or high (more than or equal to 13 years). Monthly household income was averaged across 4 years and categorized as low (less than 59,999 new Taiwan dollars (NTD; 1 NTD ≈ 0·03 $US), medium (60,000–119,999 NTD), or high-income (more than 120,000 NTD) groups. Parental marital status was dichotomized as married or not married.
Analytic Procedure
Descriptive statistics included means, standard deviations, and study variables’ distribution. Student’s t-test and chi-square test were used to identify associations between the study variables and sex. Repeated-measures latent class analysis (RMLCA) was used to identify distinct ECE trajectories from childhood to young adulthood [28]. RMLCA is a statistical method that can be used to cluster individuals into a number of latent classes based on the pattern of responses to ECE questions at discrete time points [28]. All models were estimated using SAS version 9.4 (SAS Institute Inc., Cary, NC) via Proc LCA. The number of latent patterns was determined using 1) fit indices, namely the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and sample size-adjusted BIC, where a lower value indicates a better model; 2) average classification probability (ACP), where a higher value indicates better classification; and 3) interpretation of latent groups. According to a recent simulation study, BIC was preferred when comparing models [29]. We performed multiple group comparisons to determine potential sex differences [30] and found that the latent classes of exercise varied by sex. RMLCA and subsequent analyses were therefore stratified by sex. Finally, multinomial logit models were used to examine associations between related factors and different trajectories of ECE across sex.
Missing Data
In RMLCA, missing data were addressed using the maximum likelihood estimation [28]. For the multinomial logistic model, only complete data were used (n = 1701, 82.09%). We compared the analytic sample with those that had missing data and found no differences between the two samples in terms of sex, body dissatisfaction, stress, screen behavior, parental education, parental exercise, and parental screen behavior. However, those with missing data were significantly more likely to have a lower level of BMI (18.9 vs. 19.6), parents who were not married (18.6% vs. 12.8%), or parents with low education (32.7% vs. 29.16%). A sensitivity analysis was performed to assess the robustness of findings. This involved comparing results by using a multiple imputation strategy with results applying a list-wise deletion technique [31]. Twenty sets of missing values were imputed when conducting multiple imputations by using Markov Chain Monte Carlo methods, and the results of each data set were then combined to perform multinomial logistic regressions.