 Research
 Open Access
 Published:
Using three statistical methods to analyze the association between exposure to 9 compounds and obesity in children and adolescents: NHANES 20052010
Environmental Health volumeÂ 19, ArticleÂ number:Â 94 (2020)
Abstract
Background
Various risk factors influence obesity differently, and environmental endocrine disruption may increase the occurrence of obesity. However, most of the previous studies have considered only a unitary exposure or a set of similar exposures instead of mixed exposures, which entail complicated interactions. We utilized three statistical models to evaluate the correlations between mixed chemicals to analyze the association between 9 different chemical exposures and obesity in children and adolescents.
Methods
We fitted the generalized linear regression, weighted quantile sum (WQS) regression, and Bayesian kernel machine regression (BKMR) to analyze the association between the mixed exposures and obesity in the participants aged 6â€“19 in the National Health and Nutrition Examination Survey (NHANES) 2005â€“2010.
Results
In the multivariable logistic regression model, 2,5dichlorophenol (2,5DCP) (OR (95% CI): 1.25 (1.11, 1.40)), monoethyl phthalate (MEP) (OR (95% CI): 1.28 (1.04, 1.58)), and monoisobutyl phthalate (MiBP) (OR (95% CI): 1.42 (1.07, 1.89)) were found to be positively associated with obesity, while methylparaben (MeP) (OR (95% CI): 0.80 (0.68, 0.94)) was negatively associated with obesity. In the multivariable linear regression, MEP was found to be positively associated with the body mass index (BMI) zscore (Î² (95% CI): 0.12 (0.02, 0.21)). In the WQS regression model, the WQS index had a significant association (OR (95% CI): 1.48 (1.16, 1.89)) with the outcome in the obesity model, in which 2,5DCP (weighted 0.41), bisphenol A (BPA) (weighted 0.17) and MEP (weighted 0.14) all had relatively high weights. In the BKMR model, despite no statistically significant difference in the overall association between the chemical mixtures and the outcome (obesity or BMI zscore), there was nonetheless an increasing trend. 2,5DCP and MEP were found to be positively associated with the outcome (obesity or BMI zscore), while fixing other chemicals at their median concentrations.
Conclusion
Comparing the three statistical models, we found that 2,5DCP and MEP may play an important role in obesity. Considering the advantages and disadvantages of the three statistical models, our study confirms the necessity to combine different statistical models on obesity when dealing with mixed exposures.
Introduction
The continuous increase in obesity has become an important worldwide health problem in the past 30â€‰years [1]. In 2016, about 18% of children and adolescents aged 5â€“19 were overweight or obese [2]. Obesity in children increases the risk of health conditions, such as coronary heart disease, diabetes mellitus, hypertension, and heart failure, and those obese children or adolescents can become obese adults [3,4,5]. Therefore, it is vital to identify potential risk factors contributing to obesity to reduce the prevalence and mortality rates in obesityrelated diseases. Although genetic predisposition, physical activity, and diet play an essential role in the occurrence of obesity, there is still a need for further explanation. More evidence indicates that environmental endocrinedisrupting chemicals might increase the occurrence of obesity [6,7,8,9]. Twum et al. demonstrated an underlying relation between exposure to 2,5dichlorophenol (2,5DCP) and obesity in children [4]. A significant association was found between bisphenol A (BPA) and general and abdominal obesity [10]. Deierlein showed that phthalatesâ€”specifically lowmolecular weight phthalates (monoethyl phthalate [MEP], a metabolite of diethyl phthalate (DEP); mononbutyl phthalate [MBP], a metabolite of dinbutyl phthalate (DBP), and monoisobutyl phthalate [MiBP], a metabolite of diisobutyl phthalate (DiBP))â€”had slight associations with girlsâ€™ anthropometric outcomes [11]. These substances are readily present in our daily lives, since consumer products usually use parabens as preservatives, building and food packaging materials use phthalates as plasticizers, and the production of pharmaceutical and agricultural products uses 2,5DCP as a chemical intermediate [12,13,14]. We can easily contact these environmental endocrinedisrupting chemicals via gastrointestinal intake, dermal contact, and applying products that contain these chemicals [15, 16]. However, most of the previous research studied only a unitary exposure or a set of similar exposures [17,18,19]. We are exposed to all kinds of chemical exposures simultaneously, which can result in complicated interactions. Therefore, it is necessary to use a suitable statistical model for risk assessment of exposure and obesity [20,21,22].
We collected data on urinary chemicals or metabolites that had been reported to have an effect on obesity in the National Health and Nutrition Examination Survey (NHANES) from 2005 to 2010. We studied 9 chemical exposures including phenols (BPA, benzophenone3 (BP3)), parabens (methylparaben (MeP), propyl paraben (PrP)), pesticides (2,5DCP, 2,4DCP) and phthalate metabolites (Monobenzyl phthalate (MBzP), MEP, MiBP). We selected three statistical methods, including generalized linear regression, weighted quantile sum (WQS) regression, and Bayesian kernel machine regression (BKMR) models, to better analyze multiexposuresâ€™ cofunction on adolescent obesity. Among them, BKMR model can resolve the nonlinear and complicated interactions between chemical exposures and get more accurate results comparing with the generalized linear regression [23]. All of these three methods have their own advantages and disadvantages, and we expected that this comprehensive analysis would yield insightful and fruitful conclusions.
Methods
Study sample
The NHANES is a crosssectional nationally representative program, aiming to collect information on adultsâ€™ and childrenâ€™s health and nutritional condition in the United States, which is reviewed and approved by the National Center for Health Statistics, as one of the departments of Centers for Disease Control and Prevention (CDC). The NHANES program was conducted in the early 1960s and released the data in biennial datasets. In order to get the unbiased national health information on the noninstitutionalized population of the United States, the NHANES used a considerate, multistage stratification probability sampling design [24]. We collected publicly accessible data from 2005 and 2010. We selected participants between 6 and 19â€‰years old, with attainable measurements of urinary phenols, parabens, pesticides, and phthalate metabolites Body mass index (BMI) and waistcircumference simultaneously (nâ€‰=â€‰2629) and excluded the participants whose data on covariates, including age, gender, race, education level, family incometopoverty ratio, caloric intake, serum cotinine, and urinary creatinine, were missing (nâ€‰=â€‰257). Finally, 2372 participants were included in our study.
Measurement of chemical exposures
Urinary samples were collected and stored at âˆ’â€‰20â€‰Â°C. They were sent to the National Center for Environmental Health, the Organic Analytical Toxicology Branch, for analysis. BPA, BP3, MeP, PrP, 2,4DCP, and 2,5DCP,were extracted by online solidphase extraction (SPE). They were measured by highperformance liquid chromatography as well as tandem mass spectrometry (MS/MS). MBzP, MEP, and MiBP were measured by highperformance liquid chromatographyelectrospray ionizationtandem mass spectrometry (HPLCESIMS/MS). The limit of detection (LOD) for the compounds to be analyzed, including BPA, BP3, MeP, PrP, 2,4DCP, and 2,5DCP, were 0.4â€‰ng/mL, 0.4â€‰ng/mL, 1.0â€‰ng/mL, 0.2â€‰ng/mL, 0.2â€‰ng/mL, 0.2â€‰ng/mL in the data from 2005 to 2010, respectively. And the LOD for MBzP, MEP, and MiBP were 0.3â€‰ng/mL, 0.8â€‰ng/mL, and 0.3â€‰ng/mL in the data from 2005 to 2008 and 0.2â€‰ng/mL, 0.4â€‰ng/mL, and 0.2â€‰ng/mL in the data from 2009 to 2010. These values below the limit of detection were divided by the square root of 2 to replace the original values. As one study recommended [25], we treated urinary creatinine as a covariate to explain the urinary dilution. Urinary creatinine was measured by a Beckman Synchron CX3 Clinical Analyzer. The NHANES provides detailed information on the measurement method in the section on laboratory methods on its website [26, 27].
Anthropometric variables
Trained health technicians measured the body weight and height according to the standardized protocol. The BMI was calculated using each personâ€™s weight in kilograms to divide the square of their height in meters. However, because the standard BMI shows differences for the different ages and gender among children, measuring BMI percentiles and the BMI zscore was more appropriate. The BMI zscore was calculated in regards to the childrenâ€™s age, gender, and BMI. An appropriate standard was used, which reflected the number of SDs differing from the mean of the BMI with reference to the same age and gender. The methodology to calculate the BMI zscore specifically for different ages and gender was provided by the CDC [28]. We defined a child to be obese when their BMI was above or equal to the 95th percentile for their age and gender in accordance with the CDC recommendations [29].
Covariates
Covariates, including age, gender, race, education level, family incometopoverty ratio, caloric intake, serum cotinine, and urinary creatinine, were collected by interview or laboratory detection by NHANES. Race was grouped into Mexican American, Other Hispanic, NonHispanic White, NonHispanic Black, and Other Race. Education level was categorically grouped into â‰¤ 5 grade, 6 âˆ’ 8 grade, 9 âˆ’ 12 grade, or High School Graduate with No Diploma, High School Graduate and GED or Equivalent, or More than high school. The family incometopoverty ratio was divided into three groups: â‰¤ 1.30, 1.31 âˆ’ 3.50, and > 3.50. The caloric intake was dichotomously divided into normal intake and excessive intake, according to the Dietary Guidelines for Americans 2010 [30]. Serum cotinine indirectly reflected the exposure to environmental tobacco. Serum cotinine, age, and urinary creatinine were considered to be continuous variables.
Statistical analysis
We used the Ï‡^{2} test and the ttest to analyze categorical variables and continuous variables, respectively. And for the serum cotinine and urinary creatinine, we used wilcoxon ranksum test. We calculated the descriptive statistics on BPA, BP3, MeP, PrP, 2,4DCP, 2,5DCP, MBzP, MEP, and MiBP. Because the distributions of the chemical exposures were skewed, we logtransformed the concentrations of all chemical exposures. We used the Pearson correlation to calculate the correlation coefficients among all chemical exposures. pâ€‰<â€‰0.05 was considered to be statistically significant.
Generalized linear regression
We conducted multivariable logistic regression to analyze each chemical exposure and the odds ratios (ORs) of obesity in different quartiles. We also fitted a multivariable linear regression model to assess the association between each chemical exposure and the continuous variable of the BMI zscore in different quantiles. In addition, we fitted the models, adjusting for all the chemical exposures. All the regression models were adjusted by age, gender, race, education level, family incometopoverty ratio, caloric intake, serum cotinine, and urinary creatinine. We used logtransformed urinary creatinine as an independent covariate instead of the creatinineadjusted concentration [25].
Weighted quantile sum (WQS) regression
The WQS model scored all the chemical exposures into quantiles and estimated the weight index:
where â„Š() represents any monotonic link function, Î¼ is the predictable variable, Ï‰ is the weight of the ith components to be estimated, q_{i} refers to different quantiles, and \( \left(\sum \limits_{i=1}^c{\omega}_i{q}_i\right) \) represents the weight quantile sum of the set of c components of interest. Furthermore, Î²_{1} denotes the regression coefficient for the weight quantile sum, Î²_{0} is the intercept, z^{â€²} refers to the covariates, including risk factors and confounders, and Ï† is the coefficients for the covariates. The weights were estimated between 0 and 1, and added up to 1. In this study, we divided the data into the training set (40%) and the validation set (60%), we also set Î²_{1} to be positive and the seed was set to be 2019. Besides, we also constrained Î²_{1} to be negative to find if there was a significant relationship in this way. We bootstrapped the training set 10,000 times and got the estimated weights, which maximized the likelihood of the nonlinear model. A significant level (pâ€‰<â€‰0.05) was set to test the significance of the weights in each bootstrap. We calculated the \( {\overline{\omega}}_i \) to estimate the weight quantile sum:
where n_{B} represents the number of bootstraps in which Î²_{1} was significant. The estimated WQS was then determined using the validation set. All the chemical exposures were included in the model, and a specific weight was calculated for each component, representing their contribution to the WQS index. The chemical exposures included were constrained to have the same effect with the outcome (all positive or all negative) [31].
Bayesian kernel machine regression (BKMR)
The BKMR model utilizes a nonparametric approach to flexibly model the association between chemical exposures and healthy outcomes, including the nonlinear and/or interactions in the exposureoutcome association. A highdimension exposureresponse relationship induced by multiple variables incorporated in the model would make it difficult to ascertain the basis function. Thus, we used a kernel machine regression:
where Y_{i} is the health outcome, i refers to the individual (i = 1, 2, 3â€¦n), z_{i} is the chemical exposures, x_{i} is the potential confounders, and Î² represents the effect of the covariates. Ïµ_{i} is the residual that obeys the normal distribution N (0, Ïƒ^{2}). h() is the function that fits the exposure and the outcome considering nonlinear and interactions between the exposures. We grouped the chemical exposures into three groups (group1: BPA, BP3, MeP, and PrP; group2: 2,5DCP and 2,4DCP; group3: MBzP, MEP, and MiBP), according to their source and correlation (chemical exposures with high correlation were grouped) with each other. A hierarchical variable selection approach was used to estimate the posterior inclusion probability of highly correlated variables, which was based on our prior knowledge. The model was fit with 10,000 iterations using a Markov chain Monte Carlo (MCMC) method. The parameter r.jump2 was separately set to 0.2 (in the BMI zscore model) and 0.001 (in the obesity model) to get suitable acceptance rates.
We also analyzed the association between the quantiles of the chemical exposures and binary healthy outcome (obesity and nonobesity) using a probit BKMR model:
where Î¦^{âˆ’1} is the link function and Î¼_{i} is the probability of the binary outcome [22, 23].
Trace plots of parameter in both BMI zscore and obesity model were visualized to investigate the convergence.
All of the statistical analysis were conducted using R software (version 3.6.0).
Results
There were 2372 children and adolescents included in our study. The general characteristics of the participants are presented in Table 1. The prevalence of obesity was 20.53%. It showed that the mean age of obesity and nonobesity is close: approximately 12andahalf years old. About half (44.98%) of the participants were â‰¤ 5 grade, and 53.03% had a normal caloric intake. The mean (SD) BMI and waist circumferences were 30.41 (6.99) and 96.17 (18.05) cm in the obesity group and 19.68 (3.66) and 69.79 (11.36) cm in the nonobesity group, respectively. The mean (SD) BMI zscores were 2.12 (0.32) in the obesity group and 0.18 (0.94) in the nonobesity group. There were significant differences between the obesity and nonobesity participants in terms of race, family income, caloric intake, urinary creatinine, BMI, BMI zscore, and waist circumference.
The LOD and the detection frequency of the chemicals above the LOD are shown in Table 2. The detection frequency of MEP (99.9%) had the highest detection frequency of chemical exposures and the detection frequency of all chemical exposures was above 90%. Table 2 also shows the geometric mean, the mean, and the distribution of the chemical exposures. The highest and the lowest geometric means of the chemical exposures were related to the MEP (87.12) ng/mL and 2,4,5TCP (0.09) Î¼gâ€‰âˆ•â€‰L.
We found significant correlations (Pâ€‰<â€‰0.05) among 9 chemicals (Fig. 1), except for the correlation between BP3 and 2,4DCP (Pâ€‰=â€‰0.69). There was a positive correlation between other compounds, except for a nearly no correlation of BP3 with 2,5DCP (r =â€‰âˆ’â€‰0.06). 2,5DCP was found to have a strongly correlation with 2,4DCP (r = 0.87). Additionally, a high correlation between MeP and PrP (r = 0.81) was found.
The results from the multivariable logistic and linear regression models adjusted for the covariates are shown in Tables 3 and 4, respectively. The adjusted multivariable logistic regression analysis revealed a statistically significant association between obesity and MeP (OR (95% CI): 0.80 (0.68, 0.94)), 2,5DCP (OR (95% CI): 1.25 (1.11, 1.40)), MEP (OR (95% CI): 1.28 (1.04, 1.58)), and MiBP (OR (95% CI): 1.42 (1.07, 1.89)), with MeP showing a negative association with dichotomous variable obesity. PrP was found to have a negative association with obesity only when comparing the 4th quartile with the reference quartile (OR (95% CI): 0.69 (0.49, 0.98)). When comparing the 2nd, 3rd, and 4th 2,5DCP quartiles with the reference quartile, 2,5DCP had a higher odds ratio (OR (95% CI): 1.49 (1.07, 2.07); 1.80 (1.30, 2.51), and 2.06 (1.47, 2.89), respectively) (Table 3). When comparing the second, third, and fourth quartiles of MEP with the reference quartile, MEP had a higher odds ratio (OR (95% CI): 1.04 (0.75, 1.43); 1.28 (0.92, 1.79), and 1.39 (0.98, 1.98), respectively; Table 3). We used adjusted multivariable linear regression to evaluate the relation between 9 chemical exposures and the BMI zscore (Table 4). We found MeP (second vs. first quartile) to be negatively associated with the BMI zscore (Î² (95% CI): âˆ’â€‰0.14 (âˆ’â€‰0.27, âˆ’â€‰0.01)), and 2,5DCP (third vs. first quartile) as well as MEP to be positively associated with the BMI zscore (Î² (95% CI): 0.16 (0.02, 0.30); 0.12 (0.02, 0.21), respectively). The second, third, and fourth MEP quartiles had a higher BMI zscore (Î² (95% CI): 0.02 (âˆ’â€‰0.12, 0.16); 0.12 (âˆ’â€‰0.03, 0.27), and 0.14 (âˆ’â€‰0.02, 0.30), respectively) compared with the lowest reference quartile (Table 4).
In the multivariable logistic and linear regression models, including all the chemical exposures, adjusting for the confounding effects of other chemicals, 2,5DCP, 2,4DCP, and MEP were found to have a significant association with both the dichotomous variable obesity (OR (95% CI): 1.73 (1.35, 2.24), 0.57 (0.40, 0.82), and 1.35 (1.08, 1.69), respectively) and continuous variate BMI zscore (Î² (95% CI): 0.14 (0.04, 0.24), âˆ’â€‰0.20 (âˆ’â€‰0.36, âˆ’â€‰0.05), and 0.15 (0.05, 0.25), respectively) (see Additional File 1, Tables S1 and S2). We calculated the variance inflation factors (VIFs) (see Additional File 1, Tables S3), and none of them was higher than 10.
We fitted the WQS regression model to the data to evaluate the relationship between the chemical exposures and the outcome in three models, adjusting for different covariates respectively (Table 5). The WQS index had a significant association with obesity in Model 1 (OR (95% CI): 1.50 (1.19, 1.90)). In Model2, the WQS index had a significant association with obesity (OR (95% CI): 1.51 (1.19, 1.91)). In Model 3, the WQS index also had a significantly positive association with obesity after being adjusted for all covariates (OR (95% CI): 1.48 (1.16, 1.89)). We also calculated the estimated chemical weights of the dichotomous variable obesity in obesity model, which are presented in Fig. 2a. The highest weighted chemical in the fully adjusted obesity model was 2,5DCP (weighted 0.41), followed by BPA and MEP (weighted 0.17 and 0.16, respectively). We also treated the BMI zscore as a continuous variable and fitted the BMI zscore model (Table 5). However, we did not find any significant association between the exposures and the BMI zscore in all three models. The estimated chemical weights of BMI zscore are presented in Fig. 2b. The highest weighted chemical in the BMI zscore model was 2,5DCP (weighted 0.30). Next to this were BP3 and MEP, weighted 0.28 and 0.18, respectively. In addition, we also fitted WQS model including all covariates with Î²_{1} constrained to be negative. However, no statistical difference was found in this way. (see Additional File 1, Tables S4).
We grouped 9 chemical exposures into three groups, according to their source and correlation with each other, and fitted the BKMR model to analyze the simultaneous exposure with obesity and BMI zscore. In the obesity model, the group posterior inclusion probabilities (PIP) of the pesticides group was 0.966, while the group PIP of phenol and phthalates metabolites was higher than 0.5 (Table 6). In the pesticides group, 2,5DCP seemed to drive the effect of the whole group (CondPIPâ€‰=â€‰0.978; Table 6). In the phthalate metabolites group, MEP drove the main effect of the whole group (CondPIP: 0.656), while MeP drove the main effect in the phenols group (CondPIP = 0.903) (Table 6). The overall association between the chemical mixtures and the binomial outcome is shown in Fig. 3a. We found a positive tendency between chemical exposures and the outcome, in spite of no statistically significant difference. Figure 4 a illustrates the positive associations of 2,5DCP, MEP, and MiBP with obesity in the BKMR models, while controlling all other chemical exposures at their median level. MeP demonstrated an inverse association with obesity, while no other chemical exposures showed a noteworthy change in slope. We also investigated the relationship between the outcome and a unitary predictor in exposures while fixing another predictor in exposures at the 10th, 50th, and 90th quantiles (and holding the remnant predictors to their median level), and the results are shown in Fig. 5 a. Since the slopes were different between 2,5DCP and obesity, MEP and obesity while fixing MeP at the 10th, 50th, and 90th quantiles, potential interactions might exist between 2,5DCP and MeP as well as MEP and MeP. In the BMI zscore model, the values of the group PIP in three groups were 0.329, 0.256, and 0.707, respectively. (Table 6). MEP drove the main effect in its group (CondPIP: 0.831). The overall risk of the chemical mixtures on the outcome are presented in Fig. 3b. Although no statistically significant difference was found, they revealed a positive association of the mixed exposures with the BMI zscore, when we compared all the predictors fixed at different levels with their 50th percentiles. 2,5DCP and MEP had a trend of a positive association with the BMI zscore, while 2,4DCP had an inverse association (Fig. 4 b). No obvious interaction was found in the BMI zscore model (Fig. 5 b).
To ensure the convergence, we plotted the trace plots, which showed a more or less homogeneously covered space and indicated our model had a good convergence. (see Additional File 1, Fig. 1 and Fig. 2).
For 2,5DCP and MEP seemed to drive the whole effect in pesticides group (in obesity model) and in phthalate group (in BMI zscore model), we further modeled 2,5DCP and other groups (phenols group, parabens group, and phthalate group) in obesity model and MEP and other groups (phenols group, parabens group, and pesticides group) in BMI zscore model. The credibility intervals tighten a little (see Additional File 1, Fig. 3 a and b), which meant 2,4DCP, MiBP and MBzP showed little relevance for the outcome.
Discussion
Due to the interactions between chemicals, it would be inaccurate to fit only the generalized linear regression model. Therefore, we further used the WQS and BKMR models, which can deal with the interaction between chemicals.
The generalized linear regression showed a positive association between 2,5DCP, MEP, and MiBP and obesity; however, MeP was negative with the outcome. 2,5DCP and MEP were significantly associated with the BMI zscore. In the WQS model, 2,5DCP, BPA, and MEP were found to have relatively high weights in the obesity model, while 2,5DCP and MEP were found to weight relatively high in the BMI zscore model. In the BKMR model, although no significant association was found between the overall risk of the mixed chemicals and obesity (either obesity or the BMI zscore), there was an upward trend. 2,5DCP, MEP, and MiBP were found to have a positive association in the obesity model, when fixing others at their median concentration, while in the BMI zscore model, 2,5DCP, and MEP were positively correlated with the BMI zscore. These results point out the necessity for combining three different models, considering their various advantages and disadvantages.
The generalized linear model, which is used frequently to deal with the exposureresponse model, has a fast modeling speed and allowed us to obtain an understandable interpretation of the coefficients. Usually, in the analysis to evaluate the association between exposures and outcome, a unitary exposure or a set of similar exposures is included [12, 32, 33]. Our study included 9 chemical exposures of different sorts. It should be noted that the generalized linear model could not analyze the interactions between exposures. The results may be confusing due to the colinear or interactions between the exposures.
The WQS mode can include mixed chemicals exposures, with possible high correlations and interactions that are common in real life. In our analysis, 2,5DCP and MEP were weighted highly in the WQS model. Among these, it is worth noting that BPA and BP3 were found to weigh highly in the WQS model, yet was found to have a negligible relationship with obesity in the other two models, which may be due to the limitation of the WQS model. The WQS model may lose the full exposure information of the chemical exposures using the quantiles to score the exposures. MeP weighed slightly in the WQS model, which differed from the results in the the other two models. This may result from its negative correlation with the outcome. Since one limitation of WQS is that all chemical exposures included in the model must have the same effective trend with the outcome, otherwise they will be distributed to a negligible weight in the WQS model [34]. In addition, the WQS model may result in a slight weight if a large number of exposures were included, or if there were complex interactions within mixed exposures. Two likely important exposures would have smaller weights if one of them was highly correlated with another one that was assigned a slight weight [31]. However, as for the interactions between chemical exposures, the WQS model still has a high specificity and sensitivity when dealing with mixed predictors, considering the correlated highdimensional mixtures.
The BKMR model is a new approach to deal with the complexity of mixed exposures, which can analyze not only the exposureresponse function of the overall risk of mixed chemical exposures but also the interaction between two chemical exposures. In our study, 2,5DCP and MEP have a positive association with the continuous variable BMI zscore, which was consistent with the results of our findings in the other two models. However, with the nonlinear exposureresponse function, other exposures were slightly or negatively associated with the outcomes, which showed consistency with its slight weight in the WQS model. Among the three groups, the MeP was found to have an inverse association with obesity, which is consistent with a previous study [12]. Previous studies could not reach consensus concerning phthalate and BPA, [35,36,37], and further studies are needed. It is worth noting that MiBP had a positive relationship with the dichotomous variable of obesity but had no relationship with the continuous variable. This may be due to the misleading information when we artificially classified the continuous variable into a dichotomous variable. Besides, we also found potential interactions between 2,5DCP and MeP as well as MEP and MeP in obesity model, while in the BMI zscore model there was no oblivious interactions. And further investigation is needed on these interactions. The BKMR model also has some limitations. An inconspicuous overall risk association may be observed when exposures which were positive with the outcome or were negative with the outcome both exist [22].
There were several limitations to our study. First, because of the design of the crosssectional survey project, which collected all of the data at a single time point, there was a limit to the inference of the causation between the chemical exposures and obesity. Second, we used the education level of the individuals themselves instead of their parentsâ€™ education level, which can be a factor, since parental education can change their intention to alter the obesity risk factor [38]. Third, chemical concentrations below the limit of detection were simply replaced by the value of the limit of detection divided by the square root of 2, which may cause inaccurate results. Thus, we selected chemical exposures with a high detection frequency. Fourth, obesity is the result of a combination of the longterm effects of various factors. We determined that the concentration of various exposures in urine does not justify a full inference about the mixed chemical exposures on individuals. Further prospective studies are required to investigate the longterm exposure.
Conclusion
Our study uses three statistical models to analyze the mixed chemical exposures with obesity. 2,5DCP and MEP were found to have a significant association with the outcome in all models, these results may lead to a false conclusion if only one model is considered. Since all of the models have their own advantages and disadvantages, our study confirms the necessity of combining different statistical models when dealing with the effects of mixed exposures on obesity.
Availability of data and materials
The dataset supporting the conclusions of this article is included within the article (AdditionalÂ FileÂ 2).
Abbreviations
 2,4DCP:

2,4Dichlorophenol
 2,5DCP:

2,5Dichlorophenol
 BP3:

Benzophenone3
 BKMR:

Bayesian kernel machine regression
 BMI:

Body Mass Index
 BPA:

Bisphenol A
 CDC:

Centers for Disease Control and Prevention
 CI:

Confidence interval
 DBP:

Dinbutyl phthalate
 DF:

Detection frequency
 DiBP:

Diisobutyl phthalate
 GM:

Geometric mean
 HPLCESIMS/MS:

Highperformance liquid chromatographyelectrospray ionizationtandem mass spectrometry
 LOD:

Limit of detection
 MBP:

Mononbutyl phthalate
 MCMC:

Markov chain Monte Carlo
 MeP:

Methyl paraben
 MEP:

Monoethyl phthalate
 MiBP:

Monoisobutyl phthalate
 MS:

Mass spectrometry
 NHANES:

National Health and Nutrition Examination Survey
 ORs:

Odds ratios
 PIP:

Posterior inclusion probabilities
 PrP:

Propyl paraben
 SD:

Standard Deviation
 SPE:

Solid phase extraction
 VIFs:

Variance inflation factors
 WQS:

Weighted quantile sum
References
Engin A. The definition and prevalence of obesity and metabolic syndrome. Adv Exp Med Biol. 2017;960:1â€“17.
Prevalence of obesity among children and adolescents [https://www.who.int/gho/ncd/risk_factors/overweight_obesity/obesity_adolescents/en/],(accessed 31 October 2019).
Simmonds M, Llewellyn A, Owen CG, Woolacott N. Predicting adult obesity from childhood obesity: a systematic review and metaanalysis. Obes Rev. 2016;17(2):95â€“107.
Twum C, Wei Y. The association between urinary concentrations of dichlorophenol pesticides and obesity in children. Rev Environ Health. 2011;26(3):215â€“9.
Lavie CJ, De Schutter A, Parto P, Jahangir E, Kokkinos P, Ortega FB, Arena R, Milani RV. Obesity and prevalence of cardiovascular diseases and prognosisthe obesity paradox updated. Prog Cardiovasc Dis. 2016;58(5):537â€“47.
Heindel JJ, Blumberg B, Cave M, Machtinger R, Mantovani A, Mendez MA, Nadal A, Palanza P, Panzica G, Sargis R, et al. Metabolism disrupting chemicals and metabolic disorders. Reprod Toxicol. 2017;68:3â€“33.
Kim JT, Lee HK. Childhood obesity and endocrine disrupting chemicals. Ann Pediatr Endocrinol Metab. 2017;22(4):219â€“25.
Karoutsou E, Polymeris A. Environmental endocrine disruptors and obesity. Endocr Regul. 2012;46(1):37â€“46.
Nadal A, Quesada I, Tuduri E, Nogueiras R, AlonsoMagdalena P. Endocrinedisrupting chemicals and the regulation of energy balance. Nat Rev Endocrinol. 2017;13(9):536â€“46.
Liu B, Lehmler HJ, Sun Y, Xu G, Liu Y, Zong G, Sun Q, Hu FB, Wallace RB, Bao W. Bisphenol a substitutes and obesity in US adults: analysis of a populationbased, crosssectional study. Lancet Planetary Health. 2017;1(3):e114â€“22.
Deierlein AL, Wolff MS, Pajak A, Pinney SM, Windham GC, Galvez MP, Silva MJ, Calafat AM, Kushi LH, Biro FM, et al. Longitudinal associations of phthalate exposures during childhood and body size measurements in young girls. Epidemiology. 2016;27(4):492â€“9.
QuirosAlcala L, Buckley JP, Boyle M. Parabens and measures of adiposity among adults and children from the U.S. general population: NHANES 20072014. Int J Hyg Environ Health. 2018;221(4):652â€“60.
Xia B, Zhu Q, Zhao Y, Ge W, Zhao Y, Song Q, Zhou Y, Shi H, Zhang Y. Phthalate exposure and childhood overweight and obesity: urinary metabolomic evidence. Environ Int. 2018;121(Pt 1):159â€“68.
Park H, Kim K. Concentrations of 2,4Dichlorophenol and 2,5Dichlorophenol in Urine of Korean Adults. Int J Environ Res Public Health. 2018;15(4):589.
Bui TT, Giovanoulis G, Cousins AP, Magner J, Cousins IT, de Wit CA. Human exposure, hazard and risk of alternative plasticizers to phthalate esters. Sci Total Environ. 2016;541:451â€“67.
Dodge LE, Kelley KE, Williams PL, Williams MA, HernandezDiaz S, Missmer SA, Hauser R. Medications as a source of paraben exposure. Reprod Toxicol. 2015;52:93â€“100.
Shoaff J, Papandonatos GD, Calafat AM, Ye X, Chen A, Lanphear BP, Yolton K, Braun JM. Earlylife phthalate exposure and adiposity at 8 years of age. Environ Health Perspect. 2017;125(9):097008.
Buckley JP, Engel SM, Mendez MA, Richardson DB, Daniels JL, Calafat AM, Wolff MS, Herring AH. Prenatal phthalate exposures and childhood fat mass in a New York City cohort. Environ Health Perspect. 2016;124(4):507â€“13.
Wang H, Zhou Y, Tang C, He Y, Wu J, Chen Y, Jiang Q. Urinary phthalate metabolites are associated with body mass index and waist circumference in Chinese school children. PLoS One. 2013;8(2):e56800.
Valeri L, Mazumdar MM, Bobb JF, Claus Henn B, Rodrigues E, Sharif OIA, Kile ML, Quamruzzaman Q, Afroz S, Golam M, et al. The joint effect of prenatal exposure to metal mixtures on neurodevelopmental outcomes at 2040 months of age: evidence from rural Bangladesh. Environ Health Perspect. 2017;125(6):067015.
Warner M, Rauch S, Coker ES, Harley K, Kogut K, Sjodin A, Eskenazi B. Obesity in relation to serum persistent organic pollutant concentrations in CHAMACOS women. Environ Epidemiol. 2018;2(4):e032.
Bobb JF, Claus Henn B, Valeri L, Coull BA. Statistical software for analyzing the health effects of multiple concurrent exposures via Bayesian kernel machine regression. Environ Health. 2018;17(1):67.
Bobb JF, Valeri L, Claus Henn B, Christiani DC, Wright RO, Mazumdar M, Godleski JJ, Coull BA. Bayesian kernel machine regression for estimating the health effects of multipollutant mixtures. Biostatistics. 2015;16(3):493â€“508.
Sample Design [https://wwwn.cdc.gov/nchs/nhanes/tutorials/module2.aspx],(accessed 1 April 2020).
Barr DB, Wilder LC, Caudill SP, Gonzalez AJ, Needham LL, Pirkle JL. Urinary creatinine concentrations in the U.S. population: implications for urinary biologic monitoring measurements. Environ Health Perspect. 2005;113(2):192â€“200.
Laboratory Procedure Manual (Method No: 6301.01 ) [https://wwwn.cdc.gov/nchs/data/nhanes/20092010/labmethods/PP_F_met_phenols.pdf],(accessed 31 October 2019).
Laboratory Procedure Manual (Method No: 6306.03) [https://wwwn.cdc.gov/nchs/data/nhanes/20092010/labmethods/PHTHTE_F_met.pdf],(accessed 31 October 2019).
A SAS Program for the 2000 CDC Growth Charts (ages 0 to <20 years) [https://www.cdc.gov/nccdphp/dnpao/growthcharts/resources/sas.htm],(accessed 31 October 2019).
Defining Childhood Obesity [https://www.cdc.gov/obesity/childhood/defining.html],(accessed 31 March 2020).
2010 Dietary Guidelines [https://health.gov/dietaryguidelines/2010/],(accessed 31 October 2019).
Carrico C, Gennings C, Wheeler DC, FactorLitvak P. Characterization of weighted Quantile sum regression for highly correlated data in a risk analysis setting. J Agric Biol Environ Stat. 2014;20(1):100â€“20.
Warner M, Ye M, Harley K, Kogut K, Bradman A, Eskenazi B. Prenatal DDT exposure and child adiposity at age 12: the CHAMACOS study. Environ Res. 2017;159:606â€“12.
Bhandari R, Xiao J, Shankar A. Urinary bisphenol a and obesity in U.S. children. Am J Epidemiol. 2013;177(11):1263â€“70.
Czarnota J, Gennings C, Colt JS, De Roos AJ, Cerhan JR, Severson RK, Hartge P, Ward MH, Wheeler DC. Analysis of environmental chemical mixtures and nonHodgkin lymphoma risk in the NCISEER NHL study. Environ Health Perspect. 2015;123(10):965â€“70.
Goodman M, Lakind JS, Mattison DR. Do phthalates act as obesogens in humans? A systematic review of the epidemiological literature. Crit Rev Toxicol. 2014;44(2):151â€“75.
Jacobson MH, Woodward M, Bao W, Liu B, Trasande L. Urinary Bisphenols and obesity prevalence among U.S. children and adolescents. J Endocr Soc. 2019;3(9):1715â€“26.
Liu B, Lehmler HJ, Sun Y, Xu G, Sun Q, Snetselaar LG, Wallace RB, Bao W. Association of Bisphenol a and its Substitutes, Bisphenol F and Bisphenol S, with obesity in United States children and adolescents. Diabetes Metab J. 2019;43(1):59â€“75.
BaileyDavis L, Peyer KL, Fang Y, Kim JK, Welk GJ. Effects of enhancing schoolbased body mass index screening reports with parent education on report utility and parental intent to modify obesity risk factors. Childhood Obes (Print). 2017;13(2):164â€“71.
Acknowledgements
We thank LetPub (www.letpub.com) for its linguistic assistance during the preparation of this manuscript.
Funding
The authors have no sources of funding to report.
Author information
Authors and Affiliations
Contributions
B.S. Wu participated in the study design, collected and organized data, carried out the statistical analysis, and prepared the first draft of the manuscript. Y. Jiang participated in the study design, in the coordination and the execution of data collection, statistical analysis and in writing the manuscript. X.Q. Jin participated in the study design, and gave critical appraisal of the manuscript. L. He coordinated the study design, and gave critical appraisal of the manuscript. All authors read and approved the final manuscript.
Corresponding authors
Ethics declarations
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisherâ€™s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Additional file 1 Table S1.
Association between chemical exposures and obesity with all the chemicals included in NHANES 2005â€“2010 (Nâ€‰=â€‰2372). Table S2. Association between chemical exposures and BMI zscore with all of the chemicals included in NHANES 2005â€“2010 (Nâ€‰=â€‰2372). Table S3. Variance inflation factors (VIFs) in the multivariable logistic and linear regression models, including all the chemical exposures, adjusting for the confounding effects of other chemicals in NHANES 2005â€“2010 (Nâ€‰=â€‰2372). Table S4. Association between the WQS index and obesity in negative direction. FigureÂ 1 The change of beta1 parameter values as the sampler runs in BMI zscore model. FigureÂ 2 The change of beta1 parameter values as the sampler runs in obesity model. FigureÂ 3 Overall risk (95% CI) of chemical exposures on obesity (A) and BMI zscore (B) when comparing all the chemicals at different percentiles with their median level.
Additional file 2.
Datasets generated and analyzed during the current study.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Wu, B., Jiang, Y., Jin, X. et al. Using three statistical methods to analyze the association between exposure to 9 compounds and obesity in children and adolescents: NHANES 20052010. Environ Health 19, 94 (2020). https://doi.org/10.1186/s12940020006426
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12940020006426
Keywords
 Obesity
 Adolescent
 Child
 Weighted quantile sum (WQS) regression
 Bayesian kernel machine regression (BKMR)