Study participants
China Health and Retirement Longitudinal Study (CHARLS) was a representative survey research nationwide conducted by the National School of Development of Peking University. Between June 2011 and March 2012 (baseline survey), a multistage process with random cluster sampling was conducted to select a representative sample of individuals older than 45 years in 10,287 households in 450 villages/cities. 17,708 individuals were enrolled in the baseline survey through face-to-face household interviews. After the initial recruitment, all 17,708 individuals have been re-surveyed every two years using the same questionnaire as at the baseline. Blood samples were gathered in the year of 2011 and 2015. The detailed information regarding the CHARLS has been described on the website of CHARLS: http://charls.pku.edu.cn/en. The study protocol has been approved by ethical committees of Peking University (IRB 00001052–11,014), and all individuals have provided written informed consent.
This study was a post hoc analysis of CHARLS from 2011 to 2015. Of the 17,708 individuals at baseline, 1099 subjects were younger than 45 years. Participants with detailed follow-up information, major study variables and confounding variables at baseline and follow-up period were included in the study. Participants who had cancer at baseline were excluded. Finally, 5727 subjects were included in our study. The flowchart in Fig. 1 was used to show the detailed information about the sample size of the subjects and the exclusion criteria in the current research.
Exposures and covariates
Information on participants demographics (gender, age, ethnicity, and education), lifestyle factors (smoking status and drinking status), and history of diseases (diabetes, chronic kidney disease, and hypertension) were collected during the structured household survey. Data on healthy behaviors were obtained from the subjects’ self-reported questionnaire, including smoking status (former, current smoker or never), frequency of alcohol consumption (more than once a month, once a month or never). Data on diabetes, chronic kidney disease, and hypertension were collected by trained health staff members.
Data on collection and measurement of cholesterol indexes, fasting plasma glucose, biochemical blood indexes, and other blood pressure indexes are detailed on the website of CHARLS: http://charls.pku.edu.cn/en. The mean of 3 blood pressure values was calculated and used for our analysis. The BMI was calculated using the weight and height indicators; the formula as follows: BMI = weight (Kg) / height2 (m2).
Dependent variable (Y)
The TyG index was calculated according to the following formula: TyG = ln [fasting TG (mmol/L) × FPG (mmol/L) × 0.5 × 159.37]. TG and FPG were measured by enzymatic colorimetric test method.
Independent variable (X)
A soft tape was inserted at the navel level to measure WC in a standing pose with a cloth measuring tape. At the same time, all participants needed to do a regular breathing exercise, holding the breath at the end of exhaling and letting the tape out slightly.
Mediators (M)
UA plus method was used to measure UA.
Statistical analysis
All analyses were conducted by IBM SPSS version 22.0 and SPSS Amos 22.0 with P ≤ 0.05 (two-sided) considered statistically significant differences.
Percentiles were calculated used to describe the categorical variables, while arithmetic mean with standard deviation (SD) was calculated for the description of continuous variables that met the normal distribution. The median with interquartile range was used to describe the continuous variables that didn’t meet the normal distribution. At the same time, the Student t test, the Mann-Whitney U test or the Pearson’s χ2 test was used for the comparison between male and female.
The linear regression model was constructed to explore whether WC (UA) could predict the future variation of TyG. Future TyG variation and baseline WC (UA) were the dependent and independent variables of the model, respectively. Firstly, the multicollinearity problem among independent variables was examined by the variance expansion factor (VIF). VIF greater than 10 was deemed significant multicollinearity. Secondly, three models were constructed to assess the relationship between baseline (examination at 2011) WC (UA) and future (follow-up at 2015) TyG as follows: model 1: adjusted for baseline WC (UA); model 2: model 1 plus sex, ethnicity and age; model 3: model 2 plus drinking, current smoking, education, BMI, LDL-C, HDL-C, HR, DBP, SBP, and creatinine.
Longitudinal changes of WC and UA indexes were measured at two follow-up periods. Previous study introduced the theory and application of cross-lagged panel design [26]. Overall, the cross-lagged panel model was performed for the longitudinal relationships among different interrelated variables [27]. In Fig. 2, the path with β1 showed the impact of baseline UA on follow-up WC, and β2 showed the effect of baseline WC on follow-up UA. WC and UA values were adjusted in linear regression analysis by baseline and follow-up variables: sex, age, ethnicity, alcohol consumption, current smoking, education, BMI, LDL-C, HDL-C, HR, DBP, SBP, and creatinine. Residuals were saved and after that, Z-transformation was used to standardize the saved residuals (mean = 0; SD = 1). Root-mean-square residual (RMR) and comparative Fit Index (CFI) and were enrolled for the model fits. RMR < 0.05 and CFI > 90 meant a relatively good model fit in the cross-lagged path model. In addition, this analysis model was constructed in groups of men and women, separately.
When the temporal relationship was establised between WC and UA, the mediation model would be fitted to explore the potential influence of the association between UA and WC on TyG. The values of TyG were analyzed by linear regression residual model and standardized by Z-transformation (with mean of 0 and SD of 1). According to the results of the cross-lagged path analysis model, X and M were determined as predictor and mediator, respectively. The detailed mediation model was showed in Fig. 3, which included three models, given by:
$$ \mathrm{Model}\ \mathrm{Y}={\beta}_{\mathrm{Tol}}\ \mathrm{X}\kern0.5em \left({\beta}_{\mathrm{Tol}}=\mathrm{total}\ \mathrm{effect}\right), $$
$$ \mathrm{Model}\ \mathrm{M}={\beta}_1\ \mathrm{X}\ \left({\beta}_1=\mathrm{indirect}\ \mathrm{effect}\ 1\right), $$
$$ \mathrm{Model}\ \mathrm{Y}={\beta}_2\ \mathrm{M}+{\beta}_{\mathrm{Dir}}\ \mathrm{X}\ \left({\beta}_2=\mathrm{indirect}\ \mathrm{effect}\ 2,{\beta}_{\mathrm{Dir}}=\mathrm{direct}\ \mathrm{effect}\right). $$
The following formula was used to calculate the proportion of the mediation effect, as follows: Mediation effect (%) = \( \frac{\beta_1\times {\beta}_2}{\beta_{\mathrm{Tol}}}\times 100\% \).