Descriptive Profiles of Adolescent Overweight/Obesity in China
The overall trend of overweight prevalence for Chinese adolescents declined from 23% to 8% over time for the studied full sample. However, there exists substantial differences both between gender and place of residence in terms of the prevalence of childhood overweight/obesity. Fig. 2 displays the time trend of the prevalence of overweight by gender and by place of residence (urban/rural) respectively. The prevalence of overweight for boys (from 28% to 10%) is consistently higher than that found for girls (from 18% to 5%) in an almost parallel way over time. The prevalence of overweight for urban children aged 10-12 (22%) is similar to that for rural children (23%) in 2010. Nonetheless, the prevalence of overweight and obesity for rural adolescents declines steadily faster than that for urban adolescents during the follow-up period. At the age of 16 to 18, the prevalence of overweight for urban adolescents is 10% while it is 7% for their rural counterparts.
[Insert Figure 2 here: Prevalence of adolescent overweight by gender and place of residence across time]
Developmental trajectories of BMI among Chinese adolescents
We first estimated latent class growth analysis (LCGA) models that allowed for different numbers of latent classes of overweight trajectories. The corresponding model fit indices are shown in Table 2. Among all the potential models, the three-class solution has the lowest ABIC value and the most reasonable size and shape for each class. The retrieved three classes also make theoretical and substantive sense. Therefore, we will focus on the three-class model for the subsequent analysis[1].
Table 2 Model fit indices for difference number of latent classes
# of classes
|
1
|
2
|
3
|
4
|
# of parameters
|
2
|
5
|
8
|
11
|
Log likelihood
|
-1201.880
|
-1010.641
|
-994.395
|
-994.015
|
BIC
|
2417.240
|
2054.979
|
2042.705
|
2062.162
|
AIC
|
2407.761
|
2031.283
|
2004.790
|
2010.029
|
ABIC
|
2410.888
|
2039.101
|
2017.300
|
2027.230
|
Entropy
|
——
|
0.919
|
0.636
|
0.652
|
LMR
|
——
|
p = 0.0000
|
p = 0.0012
|
p = 0.0022
|
Class size
|
100%
|
91.3% / 8.7%
|
57.6% / 36.4% / 6.1%
|
50.5% / 1.2% / 6.3% / 42%
|
As is shown in Fig. 3, we defined Class 1 (57.6%) as a stable low-risk or normal group, Class 2 (36.4%) as a decreased-risk group, and Class 3 (6.1%) as a sustained high-risk group. For Class 1, the probability of getting overweight is consistently close to zero over time. Class 2 initially has a medium risk of being overweight, which however goes down over time and eventually converges with the first class. In comparison, Class 3 had a consistently high probability (over 90%) of being overweight over the study period.
[Insert Fig. 3 here: Heterogeneous developmental trajectories of overweight and obesity for the full sample]
We further estimated the three-class LCGA model separately by gender and by urban/rural division (Fig. 4). Among the sampled girls, only 2.6% were classified into the stable high-risk group (Class 3), while for the boys, it is 9.4%. There are 8.4% of urban adolescents who were classified into the stable high-risk group (Class 3), while it is only 4.8% for rural adolescents[2]. This verifies the gender and rural-urban differences in adolescent overweight/obesity that we observed in the above section.
[Insert Figure 4 here: BMI developmental trajectories by place of residence and gender]
SES predictors of class membership
We used the 3-step approach developed by Vermunt (2010) to estimate the association between the predictor variables and the latent class variable. R3STEP option in Mplus was used to specify the auxiliary variables in the model as predictors rather than distal outcomes of the latent class variable. Along with the regular latent class growth model, each observation was assigned to a most likely class variable N by taking measurement error or classification uncertainty rate into account [3]. In the last step, a multinomial logistic regression model was estimated, in which SES indicators and other controlling covariates predicted the most likely class variable N.
As is mentioned above, the substantial economic and socio-cultural disparities between urban and rural China may lead to fundamental differences in lifestyles, nutrition status, and SES-obesity patterns between urban and rural children. Thus, it was critical to analyze the SES-BMI trajectories separately for both urban and rural adolescents. In fact, in the analysis with the full sample, neither family income nor parental education was found to be significantly associated with the latent class membership of BMI trajectories. However, separate analyses for the urban and rural sub-samples revealed more nuances about the potential underlying patterns (Table 3).
Specifically, for the urban sample, since less than one percent of the sampled girls was classified into the sustained high-risk group, we estimated the association between family SES and class memberships only for the urban boys, among whom 14% was classified into the sustained high-risk group. The results show that higher level of family income in urban area consistently predicts a lower probability of getting into stable high-risk group (ORs = 0.23, 0.14, and 0.23 for the three levels of family income). However, the opposite pattern was found for rural adolescents, in which higher level of family income seemed to increase the probability of stably staying in the high-risk group (ORs = 5.85, 4.34, and 2.66), with only the first income level being statistically significant. In addition, compared to agricultural work, rural adolescents whose parents engage in administrative/technical jobs (OR = 24.21) were also more likely to get into the high-risk class. Further complementary analyses for rural girls and rural boys show similar patterns as reported above.
Table 3 Multinomial Logit Models of Stable High-Risk BMI Class and Normal Class for the Urban and Rural Samples a
|
Urban (Male)
|
Rural
|
|
High vs. Normal
|
High vs. Normal
|
SES Determinants
|
O. R.
|
S. E.
|
O. R.
|
S. E.
|
Family Income
[Low income (below 25 %) as reference]
|
|
|
|
|
Lower medium (25% to 50 %)
|
0.23+
|
0.88
|
5.85*
|
0.92
|
Upper medium (50 % to 75 %)
|
0.14*
|
0.98
|
4.34
|
1.09
|
High (above 75 %)
|
0.23+
|
0.95
|
2.66
|
1.13
|
Parental Education
[6 years or less of education as reference]
|
|
|
|
|
9 years
|
1.45
|
1.24
|
0.54
|
0.75
|
12 or more years
|
0.99
|
1.45
|
0.54
|
0.92
|
Parental Occupation
[Agricultural related job as reference]
|
|
|
|
|
Administrative/technical job
|
3.89
|
1.45
|
24.21**
|
1.09
|
Non-agricultural service / factorial job
|
1.59
|
1.28
|
0.56
|
0.85
|
Other job
|
2.02
|
1.16
|
5.77
|
1.60
|
Note: a. The dependent variable was a three-category nominal variable, but only the comparison between stable high-risk class and normal class was reported in Table 2; b. Results for the control variables in the model was omitted in Table 2; c. +p 0.1; *p 0.05; **p 0.01; ***p 0.001.
Health consequences of Adolescent Overweight/Obesity
We first compared the mean differences among the three classes of overweight/obesity trajectories in adolescents’ self-rated health without controlling any covariates. This was done by using the automatic BCH procedure in Mplus. The results show that the sustained high-risk class (class 3) shows significantly lower self-rated health than the decreased risk class. We then further checked these results by including more controls in the model. That is, we predicted the distal outcomes (self-rated health) from the three latent classes while controlling for the effects of other possible confounding variables. This procedure was conducted by using the manual BCH method which estimated a multiple group model with each observation adjusted by a group-specific weight. The class-specific intercepts of self-rated health in 2016 are 3.72 (Class 1), 3.81 (Class 2), and 3.46 (Class 3), after controlling for gender, parental education, family income and baseline self-rated health in 2010. The Wald’s Chi-square test shows that the sustained high-risk group (Class 3) had significantly lower self-rated health than the normal group (Class 1) and the decreased-risk group (Class 2).
[1] We also replicated the 3-class solution by an alternative model specification allowing for growth factor variation within each class.
[2] In the complementary analysis, we found that almost all of the 2.6% of the sampled girls who were classified into Class 3 are from rural areas, whereas a much higher proportion of urban boys were classified into Class 3 than rural boys (14.1% vs. 6.7%).
[3] N is treated as an imperfect indicator of the latent class variable C with measurement error defined by the logits for the classification probabilities for most likely class membership [33].