Study design and participants
This investigation was based on a large school-based cross-sectional study that was part of a population survey of autism spectrum disorders led by the government. Relevant sampling methods have been described in a previous study [7] and are briefly stated as follows. This study was conducted using multistage, stratified cluster random sampling among children ages 3 to 12 years old in Shanghai, China, in June 2014. We randomly selected three urban districts and four suburban districts from a total of 17 districts across Shanghai. In total, 134 of 949 (14.12%) kindergarten schools, as well as 70 of 436 (16.06%) primary schools, were randomly sampled from a set of schools located in the selected districts (Figure 1). In total, 84,075 of 576,621 (14.58%) children were recruited from these selected schools according to the proportion of students in each district to all of the sampled districts.
The child’s family, social environment, and growth questionnaires were administered to teachers, who accepted uniform training on completing, distributing, and collecting the questionnaires. Teachers informed their students to take the questionnaire home, and then the students’ parents were asked to complete the questionnaire to collect multilevel information on the child’s characteristics (e.g., age, sex, weight, height, number of siblings, birth order, birthweight, feeding pattern, parental ages at childbirth, workday TV time, Internet use time, and snacking frequency) and family structure (e.g., parental weight status, parental education level, family income and residential site). Then, the teachers collected the completed questionnaires and returned them to the investigators. Questionnaires with key information missing, including height, weight, number of siblings, or birth order, were excluded in the final analysis.
Measurement
Body mass index (BMI, kg/m2) was calculated as weight (kg) divided by height (m) squared. Thinness, overweight, and obesity were defined according to the International Obesity Task Force–recommended age- and sex-specific cutoff points for children ages 2 to 18 years old. The BMI cutoffs for grades 1, 2, and 3 were <18.5, <17.0, and <16.0 kg/m2, respectively, and the cutoff for overweight was ≥25.0 kg/m2. The cutoff for obesity was ≥30.0 kg/m2, and the cutoff for severe obesity was ≥35.0 kg/m2 [27, 28]. For adults, the weight status was categorized by BMI into underweight (<18.5 kg/m2), normal weight (18.5–25.0 kg/m2), and overweight (≥25.0 kg/m2) classes, which included obesity and severe obesity as defined based on the World Health Organization cutoffs.
According to a previous study [17], we divided the number of siblings into three groups as follows: none (only child), one, and two or more siblings. We categorized birth order into four groups as follows: only child, oldest child, youngest child, and middle child. We included the number of younger or older siblings in three groups: none (only child), one, and two or more siblings. For birth order, the middle child represented children who had younger sibling(s) and older sibling(s). For the number of younger siblings, the one-sibling group or the two-or-more-siblings group represented the children who were the oldest child and who had either one or two or more younger siblings. For the number of older siblings, one or two or more siblings represented children who were the youngest child and who had either one or two or more older siblings.
Childhood characteristics included age (in years), sex (boy, girl), birthweight (<2,500, 2,500–4,000, or ≥4,000 g), feeding pattern (breast-feeding, formulary-feeding, mixed-feeding), parental ages at childbirth (<25, 25–34, or ≥35 years old), workday TV time (<1, 1–3, or >3 hours/day), Internet use time (<2, 2–4, or >4 hours/day), and snacking frequency (0, 1–3, or >3 times/day) were considered as potential prenatal confounding factors [24, 29, 30]. Family characteristics included parental education level, which was divided into low (illiterate, primary school, or junior high school), middle (senior high school, technical school, or college), and high (undergraduate or above). Family income was categorized into three groups as follows: low (<10,000, 10,000–30,000, or 30,000–50,000 Chinese yuan), middle (50,000–100,000, 100,000–150,000, or 150,000–200,000 Chinese yuan), and high (200,000–300,000, 300,000–1,000,000, and >1,000,000 Chinese yuan), according to a social science definition [31]. Residential site was defined as urban or suburban residents according to the participants’ living district.
Statistical analysis
We used EpiData 3.1 (EpiData Association, Odense, Denmark) for data entry and applied a logic error check. To ensure the reliability, consistency, and correctness of inputted data, we randomly sampled 15% of questionnaires for repeat data entry. We obtained verbal consent from all participants and their parents before investigation. This study was approved by the Institutional Review Boards of the Shanghai Municipal Commission of Health and Family Planning.
We computed sampling weights using inverse probability weighting, which represented the inverse of the combined selection probability in each sampling stage. Sample weight (Wt_Sample) was the product of the sampling weights and the nonresponse weight, which was calculated by the following equation:
Wt_Sample = Wt_Strat1 × Wt_Strat2 × Wt_NR, (1)
where Wt_Strat1 is the inverse probability of an “urban district” or “suburban district” being selected in the central urban or suburban districts stratum in Shanghai, Wt_Strat2 represents the inverse probability of a “kindergarten” or “primary school” being selected in the kindergarten or primary school stratum in each selected district, and Wt_NR is the inverse probability of the nonresponse rate for questionnaires in each of the selected districts.
We used the Chi-square tests to compare the distribution of childhood and family characteristics, as well as prevalence of thinness among the groups for different numbers of siblings, birth order, number of younger siblings, and number of older siblings. On the basis of complex sampling, we used multinomial logistic regression models considering sample weight to calculate the OR and 95% CI of the number of siblings, birth order, number of younger siblings, and number of older siblings for grades 1, 2, and 3 thinness among boys and girls. We made additional adjustments for the multinomial regression models, including model I: adjusted for age, which was related to the BMI category; model II: adjusted for age and childhood characteristics, including birthweight, feeding pattern, parental age at childbirth, workday TV time, Internet use time, and snacking frequency, which were reported to be associated with sibship composition and BMI category; and model III: adjusted for age, childhood characteristics, and family characteristics, including parental weight status, parental education level, family income, and residential site, which could reflect the family resources for children to some degree. We conducted statistical analysis using the software package IBM SPSS Statistics (version 24.0) The statistical significance was defined as a P-value <0.05 by a two-tailed test.