Impact of ambient air pollution exposure during pregnancy on adverse birth outcomes: generalized structural equation modeling approach

Studies in the city of Durban have reported that increased levels of ambient air pollution in the city were found to be a major health concern [11, 32]. Recent studies also indicated that the participants in South Durban are exposed to high levels of NO_x [33, 34]. We analysed data from the Mother and Child in the Environment (MACE) birth cohort, a study with ongoing recruitment in Durban, South Africa. This is described in detail elsewhere [35]. Here we report on the enrolled cohort of 996 pregnant women from March 2013 to May 2017 from eight public sector antenatal clinics in Five communities in the south of Durban (Merebank, Bluff, Wentworth and Austerville), located in close proximity to major industries, as well as communities in less heavily industrial areas in the north of Durban (KwaMashu, and Newlands East) were selected (Fig. 1). The selected communities were similar in socio-demographic profiles.

All pregnant women that were at a gestational age of less than 20 weeks and resident for the full duration of the pregnancy in the geographical area within which the clinic was located as well as for the follow-up period of 5–6 years, were recruited into the study. Women with multiple pregnancies (n = 2); miscarriages (n = 55); stillbirths (n = 25); and termination of pregnancy (n = 2) were therefore excluded from the cohort. A further 225 participants who relocated outside the areas of interest and decide to use clinics closer to their new homes were excluded. This reduced the number of enrolled subjects followed up through to labour and delivery during their pregnancy to 687 participants. Finally, excluding 31 participants with postdate birth (gestational age above 42 weeks), the effective sample size is 656 mother–child pairs (see SupFigure 1 in Supplementary Material).

Air pollution exposure to PM_2.5, SO₂ and NOx measurements for participants in the MACE birth cohort was derived from a land use regression model. This is described in greater detail elsewhere [34, 37, 38]. Using the methodology used in the European Birth Cohorts (ESCAPE), samples of NOx were taken at 40 randomly selected sites in the north and south Durban areas (Fig. 1) using Ogawa samplers over two two-week periods during mid-summer and mid-winter, to account for seasonal variability. The air pollution monitoring campaign was undertaken over two-week periods with a one-week break in between to allow for sample preparation, for a duration of nine consecutive months. An annual average concentration was estimated from the results of the two measurements for each sampling site by adjusting it with data from the Air Quality Monitoring Station of the eThekwini Municipality. At one additional site (reference site), the pollutants were measured using the same sample media for the full year to allow for the site-specific measurements to be temporally adjusted to the long-term annual average for the observation period [38]. An annual adjusted average prenatal PM_2.5, SO₂ and NOx exposure (\(\mu g/{m}^{3}\)) was predicted, using a combined land use regression model based on pre-selected geographic predictors, such as land use types (area of industrial land use, open space land use, and the harbour), road length, topography, population, and housing density. The model developed accounted for 73% of the variance in ambient PM_2.5, SO₂ and NOx measurements. No temporal adjustments were made. The parameter estimates were used to predict PM_2.5, SO₂ and NOx exposure at the residential addresses of the study participants.

The three adverse birth outcomes examined in this study were the following: PB, LBW and SGA. All of the data were extracted from MACE data. Gestational age was assessed by obstetricians based on last menstruation or early ultrasound estimates. PB was coded as a dichotomous variable as infants were born before 37 completed weeks of gestation or not. Birthweight data were coded as a dichotomous variable, indicating LBW as a birthweight \(\le\) 2,500 g (g). SGA was defined as \(\le\) the 10.^th percentile for birth weight by gestational age [39, 40] across our sample and was categorised as a dichotomous variable. Birthweight (BW) measurement was obtained by trained nurses. Exploration of the data was performed using parallel coordinate plots, in order to examine trends of adverse birth outcomes across exposure to air pollution, and clinical factors. Details about parallel coordinate plots data visualization have been previously described [41].

This study considered observed covariates as prenatal risk factors. These are exposure to ambient PM_2.5, SO₂ and NOx pollution and clinical (gestational weight gain, BMI in the first trimester, and HIV status). Maternal socio-demographic status included demographic variables (maternal age, infant gender and low socio-economic status (unemployment, multiparous, low income), primary or less education), and low socio-economic housing). The perinatal health behavioural characteristics included alcohol use, smoking, and passive exposure to tobacco smoke during pregnancy. We also included walking and physical exercise as indicators of physical activity. Gestational weight gain was obtained as the difference in kilograms between the weight at the third and first trimesters and maternal BMI (in kg/m²) was calculated using first-trimester weight and height.

A structural equation model [42] is a multivariate statistical model that involves relationships among endogenous and exogenous latent variables, accounting for measurement error. It provides a general framework for modelling stochastic dependence that arises through cause-effect relationships between random variables. SEM minimises the effect of residual confounding in associations, especially in observational studies [43]. It allowed including variables with a mediating effect on the exposure and outcome variables. A SEM model is composed of two sub-models: a measurement model and a structural or causal model. In path diagrams of SEM, the ovals signify latent variables and observed variables are shown in rectangles. A structural model constitutes a directional chain system that describes the hypothetical causal relationship between the constructs of theoretical interest (latent variables) using path diagrams [44, 45]. The structural component of the model has the following mathematical form:

where \({{\varvec{\eta}}}_{{\varvec{i}}}\) is a vector of latent endogenous variables for unit i, \({\boldsymbol{\alpha }}_{{\varvec{\eta}}}\) is a vector of intercept terms, \(\mathbf{\rm B}\) is the matrix of coefficients giving the expected effects of the latent endogenous variables (\({\varvec{\eta}})\) one each other, \({{\varvec{\xi}}}_{{\varvec{i}}}\) is a vector of latent exogenous variables, \({\varvec{\Gamma}}\) is a coefficient matrix giving the expected effect of latent exogenous variables (\({\varvec{\xi}}\)) on latent endogenous variables (\({\varvec{\eta}})\), and \({{\varvec{\zeta}}}_{{\varvec{i}}}\) is the vector of disturbances. I = 1, …, n, E(\({{\varvec{\zeta}}}_{{\varvec{i}}}\)) = 0, COV(\({{{\varvec{\xi}}}_{{\varvec{i}}}}^{\boldsymbol{^{\prime}}}\),\({{\varvec{\zeta}}}_{{\varvec{i}}}\)) = 0, and (I-\(\mathbf{\rm B}\)) is invertible.

A measurement model describes the relationships between latent variables and their manifest variables. The measurement model is represented as

where \({{\varvec{y}}}_{{\varvec{i}}}\) and \({{\varvec{x}}}_{{\varvec{i}}}\) are vectors of the observed indicators of \({{\varvec{\eta}}}_{{\varvec{i}}}\) and \({{\varvec{\xi}}}_{{\varvec{i}}}\), respectively, \({\boldsymbol{\alpha }}_{{\varvec{y}}}\) and \({\boldsymbol{\alpha }}_{{\varvec{x}}}\) are intercept vectors, \({{\varvec{\Lambda}}}_{{\varvec{y}}}\) and \({{\varvec{\Lambda}}}_{{\varvec{x}}}\) are matrices of factor loadings or regression coefficients giving the effect of the latent \({{\varvec{\eta}}}_{{\varvec{i}}}\) and \({{\varvec{\xi}}}_{{\varvec{i}}}\) on \({{\varvec{y}}}_{{\varvec{i}}}\) and \({{\varvec{x}}}_{{\varvec{i}}}\), respectively, and \(\bf {{\varvec{\varepsilon}}}_{{\varvec{i}}}\) and \({{\varvec{\delta}}}_{{\varvec{i}}}\) are the unique factors of \({{\varvec{y}}}_{{\varvec{i}}}\) and \({{\varvec{x}}}_{{\varvec{i}}}\). We assume that the unique factors (\(\bf {{\varvec{\varepsilon}}}_{{\varvec{i}}}\) and \({{\varvec{\delta}}}_{{\varvec{i}}}\)) have expected values of zero, have covariance matrices of \({{\varvec{\Sigma}}}_{{\varvec{\varepsilon}}{\varvec{\varepsilon}}}\) and \({{\varvec{\Sigma}}}_{{\varvec{\delta}}{\varvec{\delta}}}\), respectively, and are uncorrelated with each other and with \({{\varvec{\zeta}}}_{{\varvec{i}}}\) and \({{\varvec{\xi}}}_{{\varvec{i}}}\).

GSEM is a more flexible modelling approach than SEM, similar to a generalised linear model (GLM), as a more flexible alternative to ordinary least squares regression. The GSEM allows responses of continuous or binary, ordinal, count, or multinomial variables. GSEMs represents a generalisation of SEMs by allowing the use of discrete variables and non-Gaussian distributions. They combine observed (or manifest) and latent variables representing unmeasured constructs. A GSEM can be defined as

where x and y are vectors of manifest variables and \({\varvec{\eta}},{\varvec{\xi}},{\varvec{\zeta}}\) represent the latent variables, while δ and ε denote the error terms. The functions (\({\varvec{f}}_{\varvec{\eta}}\) \({\varvec{f}}_{\varvec{x}}\), \({\varvec{f}}_{\varvec{y}}\)) provide a general way to represent the connections between the variables within the parentheses to those on the left-hand side of each equation in Eq. 4.

The models were estimated by using the robust maximum likelihood approach with a method of mode-curvature adaptive Gauss–Hermite quadrature (MCAGH), which is superior in terms of accuracy to the nonadaptive methods.

The goodness of fit for each model was assessed with the Akaike information criterion (AIC) and the Bayesian information criterion (BIC). Lower AIC and BIC values indicate better model fit. AIC and BIC both balance model fit with parsimony, and each penalises based on the number of parameters. BIC imposes a larger penalty for complex models. As a result, AIC may overfit the model while BIC may underfit the model, but generally, they correspond closely with one another [46]. In the current data, we have tested the GSEM model using the Bernoulli distribution with a probit, logit and complementary log–log link functions. The fitted model with a complementary log–log link was found with lower AIC and BIC values (Table 1). Then, our final model was fitted with a Bernoulli distribution and complementary log–log link.

Indirect effects were calculated by multiplying the slope coefficients on each path. They were then summed to obtain the overall indirect effect of the variable. Total effects were calculated as a sum of the direct and indirect effects and reported in effect coefficients. These values were obtained using the nlcom command. We have used multiple imputation to address these missing points. All analyses were performed at a 5% significance level using STATA 15.0.

The parallel coordinate plots revealed that mothers of infants with adverse birth outcomes LBW, SGA and PB tend to have average to higher prenatal exposure to PM_2.5 (Fig. 2). For SO₂ and NOx, similar trends of high range of variation from low to high were observed across different adverse birth outcomes. The parallel coordinate plots further displayed that infants with adverse birth outcomes were found to be born from mothers with lower BMI at first trimester (Fig. 2). Furthermore, both the scatter plot matrix and the colour map on correlations showed that exposure to NOx had a modest positive correlation with exposure to PM_2.5 and SO₂ pollution. On the other hand, PM_2.5 is weakly and negatively correlated with exposure to SO₂ (Fig. 3). This indicates the absence of multicollinearity among air pollution exposure measures. This all provokes the use of a comprehensive statistical model that considers the three adverse birth outcomes simultaneously to examine the adverse effect of air pollution and other adjusted factors.

The median annual air pollutant levels of PM_2.5, SO₂ and NOx for individual women and the percentage of adverse birth outcomes by prenatal exposure factors for individual women and their newborns in the MACE birth cohort are shown in Table 2. The overall median level of exposure to PM_2.5, SO₂ and NOx was 13.0 μg/m³ (range 8.9– 14.1 μg/m³), 2.8 μg/m³ (range 2.1 – 5.9 μg/m³), and 34.4 μg/m³ (range 2.5 – 45.4 μg/m³) respectively. The mean maternal age was 26 years (SD: 5.7 years). Of 656 infants in the birth cohort, 66.5% were from south Durban. The median level of exposure to PM_2.5 was similar across all adverse birth outcomes while a higher median exposure level to NOx (34.6 μg/m³ (range 2.5 – 45.4 μg/m³)) was observed among mothers with preterm birth (Table 2).

Figure 4 and Table 3 presents the final GSEM model containing both the structural and measurement components. The fitted model had a minimum AIC and BIC values compared to other competing models. It was found relatively parsimonious. The path diagram for the final model with all variables is given in Fig. 4. The single-headed arrows indicate causal effects and the associated parameter values show the coefficient estimates.

Table 3 and Fig. 4 presents the coefficient estimates of direct effects of pathways between prenatal air pollution exposure and adverse birth outcomes from the fitted generalised structural equation model. Results showed that increased prenatal exposure to particulate matter PM_2.5 increased the risk of LBW (AOR = 1.3, 95% CI:1.02–1.42). Prenatal exposure to SO₂ was directly associated with SGA (AOR = 1.1, 95% CI:1.01–1.13). i.e. as exposure level to SO₂ increases the probability of being born small for gestational age increases. The direct effects of prenatal exposure to NOx on adverse birth outcomes were significant, but not in the expected direction (Fig. 4 and Table 3). Our results suggest that infants born from smoker mother have a significantly increased risk of PB (AOR = 1.9, 95% CI: 1.27–2.89). Moreover, infants from HIV positive mothers had a higher tendency to be born preterm (AOR = 0.8, 95% CI:0.49–0.96).

The results also revealed that LBW (AOR = 0.9, 95% CI: 0.92–0.97), SGA (AOR = 0.9, 95% CI: 0.91–0.95) and PB (AOR = 0.94, 95% CI: 0.93–0.95) were negatively associated with increased BMI at first trimester. Increased gestational weight gain had associated with decreased odds of PB (AOR = 0.98, 95% CI:0.97 – 0.99). Among socio-economic variables, the results showed that the primary or less mother's education level had a negative significant association with PB (Fig. 4). Infants from mothers who had primary or less education have a significantly higher probability of being preterm (AOR = 2.7, 95% CI: 1.15–6.17). Our results also suggest that infants born from mother's living in lower socio-economic housing had associated with increased risk of LBW (AOR = 1.7, 95% CI: 1.61–1.85). Compared to male infants, females were less likely to be born with SGA (AOR = 0.9, 95% CI: 0.86–0.89) (Fig. 4 and Table 3).

Table 4 displays the indirect, and total effects of prenatal exposure factors with the adverse birth outcomes. We found that prenatal exposure to PM_2.5 (AOR = 0.03, 95% CI: 0.02 – 0.04), and SO₂ (AOR = 0.002, 95% CI: 0.01 – 0.02) and NOx (AOR = 0.001, 95% CI: 0.0003 – 0.002) exhibits indirect effect on LBW through PB. Furthermore, all of the three pollutants were associated with SGA indirectly through PB. However, these indirect effects were considerably of low effect sizes (Table 4). Infants from mothers with a higher level of exposure to PM_2.5, SO₂ and NOx are more likely to be LBW and SGA partly because of being preterm. Even if the direct effects were not in the expected direction, we observed a low-level indirect effect of prenatal exposure to NOx on LBW and SGA through PB (Table 4).

Both the estimated direct effect of PM_2.5 on LBW (AOR = 1.3, 95% CI:1.02–1.42) and indirect effect on through PB (AOR = 0.03, 95% CI: 0.02 – 0.04) are relatively higher, resulting a significant stronger positive total effect (total effect = 1.94, 95% CI:1.49, 2.34). The indirect effect points to the existence of a mediating effect of PB on the effects of prenatal exposure to air pollution on being born LBW. This suggests that preterm infants with increased prenatal exposure to air pollution were more likely to be born with LBW. Lastly, PB has a mediating effect on how BMI at first trimester affects LBW (indirect effect = 0.003, 95% CI:0.0.002, 0.005) and SGA (indirect effect = 0.003, 95% CI:0.0.002, 0.005) (Table 4).