Low birthweight is a strong predictor of perinatal morbidity and mortality [1
]. Intrauterine growth has implications on cardiac health, bone development, and other health consequences later in life [2
]. Studies suggest that up to 80% of the variation in birthweight is determined by genetic factors [3
]. The residual variance is determined partially by maternal health and exposures during pregnancy [4
]. Environmental exposures are estimated to contribute about 25% of the variation in birthweight [3
]. The contribution of environmental exposures to birthweight reduction might not be large enough to result in a shift from normal to low birthweight (i.e., birthweight < 2500 g. However, given the multiple exposures, and the extent of the population exposed on a daily basis throughout the pregnancy, birthweight differences within the normal range may have meaningful implications on public health on the population level.
There is extensive literature documenting the association between multiple environmental exposures (e.g., air pollution, temperature, and noise) [11
] or neighborhood built environment characteristics (e.g., greenness) [16
] and birthweight. However, most studies assessed the effect of a single environmental exposure, either without considering other exposures or with the consideration of one additional exposure (usually temperature or fine particulate matter (PM2.5
]. Given the wide range of environmental exposures that can potentially affect birthweight, it has been challenging to isolate important environmental predictors of birthweight and to identify the exposures that are more highly weighted in their contribution to differences in birthweight. There is a gap in studies that consider the impact of multiple environmental exposures on birthweight while accounting for the correlations between individual and environmental risk factors.
For decades, the research focus of the effects of external environmental exposures has been on the effects of single pollutants. In recent years, the exposome concept has emphasized the need for epidemiological studies that consider many exposures [19
]. Only a few recent studies have taken such an exposomic approach when investigating the environmental predictors of birthweight. Hystad et al., for example, have assessed the effect of greenness exposure on birthweight while accounting for multiple environmental exposures (e.g., PM2.5
, black carbon, NO2
, noise, walkability, and park proximity) [20
]. Another recent study of six European cohorts have estimated associations between low birthweight and multiple exposures defined as the urban exposome including the built environment, air pollution, road traffic noise, meteorology, natural space, and road traffic [21
]. Finally, Agier et al. have examined over 100 predictors of birthweight and found lead exposure as a significant risk factor for lower birthweight. After correcting for multiple testing, no other exposure was significantly associated with birthweight [22
]. Still, the ability to account for multiple exposures is often limited by sample size and statistical power or by the availability of exposure data. In addition, with multiple spatially and temporally correlated exposures, a priori selection of predictors can be challenging.
In the current study, we apply a combination of variable selection, traditional, and novel regression methods. Using a large cohort of births in Massachusetts, and a wide range of prenatal exposures, we identified important predictors of birthweight out of a set of environmental exposures, built environment characteristics and economic indices, and assessed the relative contribution of each of the selected exposures to differences in birthweight.
shows the maternal and neonatal characteristics of the 640,659 births. The mean birthweight was 3379.3 g, 48.8% of the newborns were female, and the mean gestational age at birth was 39.0 weeks. The mothers were 30.1 years old on average, 69.9% were white, and 33.1% received governmental support.
show the maps with the distributions of the exposures across urban census block-groups. The mean (standard deviation) of the exposures during pregnancy were 10.4 (1.9) µg/m3
, 11.1 (4.7 °C) for temperature, and 0.5 (0.2) for NDVI. For the walkability index, where higher values indicate neighborhoods that encourage walking, the values ranged between −2.8 and 16.9 and the mean index was 1.1 (1.9). The mean nighttime noise level was 43.5 (2.9 dB). Thirty-nine percent of the mothers resided in block-groups with similar distribution of the high (≥$
100 K/year) and low (<$
25 K/year) income groups. The mean index of economic dissimilarity was 16.9%, indicating that in general, the distribution of the “high income” and “low income” groups were similar at the block-group and the census tract level (Table 2
We observed large negative correlations between the average temperature in the first and third trimesters (r = −0.9) and correlations around 0.5 between PM2.5
exposure in the first and third trimester and between temperature and NDVI exposures (Figure 1
shows the multivariable adjusted difference in birthweight for an IQR rise in each exposure, obtained from single exposure models. We observed lower birthweight associated with higher PM2.5
exposure in the three trimesters of pregnancy; temperature exposure in the 2nd and 3rd trimester; and with higher walkability, noise, and IED. Conversely, we found that higher greenspace in the three trimesters of pregnancy and residing in block groups with more residents of the ‘high income’ group were associated with higher birthweight.
In the multivariate model including all exposures, temperature exposure in all three trimesters was significantly associated with birthweight. Although attenuated due to adjustment to the other covariates, the associations with noise, ERS, walkability, and NDVI in the 1st and 3rd trimesters remained significantly associated with birthweight. IED, NDVI in the 2nd trimester, and PM2.5
were no longer associated with birthweight (Table 3
). The associations with the full set of model covariates are presented in Table S3
All the effect estimates for the associations with birthweight were obtained from linear regressions. The single exposure models tested the association with each exposure separately, with adjustment for maternal and neonatal covariates. The multivariate model includes all the exposures selected by the elastic net model, with adjustment for maternal and neonatal covariates. Each model was adjusted for the following covariates: season and year of birth, government support for prenatal care, the maternal highest level of education attained, race, maternal age, parity, maternal smoking before or during pregnancy, diabetes mellitus, chronic hypertension, and gestational age at birth. We present the results as the difference in birthweight (in grams) for an interquartile (IQR increase in the exposures).
We conducted several sensitivity analyses to make sure the associations found in our study were robust. The addition of random effects to the multivariate regression yielded larger standard errors, as expected, but the significance and the magnitude of the effects remained similar for all the exposures except for temperature. With the inclusion of a random intercept, temperature in the first and second trimesters was no longer significantly associated with birthweight (Table S4a
). The inclusion of neighborhood SES adjustment (Table S4b
) and the restriction of the study population to term births (Table S4c
) did not change the results.
shows results from the WQS showing the weighted contribution associated with rises in each of the exposures and birthweight.
shows the weighted contribution of each exposure to a negative difference in birth weight. Results were obtained from a WQS regression. The WQS regression summarizes all the exposures into one index while taking the association with the outcome into account, and the contribution of each exposure is weighted based on its relevance to the overall association with the outcome. The weights were assigned to quartile-scored exposures within the composite index and were constrained to be between zero and one and summing up to one.
The combined index of exposures was significantly associated with lower birthweight. We found nighttime noise to be the largest contributor to lower birthweight, accounting for 18% of the weights. This result was consistent across the sensitivity analyses using different quantiles. Temperature in the third trimester was equally weighted to noise, but results differed in the sensitivity analyses. IED, temperature in the first trimester, and PM2.5
exposures across pregnancy had the smallest contribution to lower birthweight. The relative contribution of NDVI exposures largely differed between the analyses. However, when summing all three trimesters of exposure, NDVI accounted for 34% of the wights in the main analysis and 40% of the weights in the sensitivity analyses (Figures S9 and S10
We found environmental exposures, built environment exposures, and economic indices during pregnancy to be important predictors of birthweight, even after accounting for known individual maternal and neonatal risk factors. We observed lower birthweights in infants born to women with higher temperature exposures during pregnancy, living in areas with less greenness and higher noise, living in more walkable areas, and in areas with more of the “low income” population. Treating the exposures individually, nighttime noise had the highest weight in its contribution to lower birthweight (graphical abstract). Summing trimestral exposures together, NDVI was the most highly weighed exposure.
Our findings are similar to previous studies where associations with birthweight were observed with maternal exposure to various environmental exposures, among them PM2.5
], temperature [18
], noise [6
], greenness [5
], and walkability [20
]. Although, most of these studies have assessed the effect of a single exposure at a time, some did try to distinguish between these spatially correlated exposures and studied the effects of multiple exposures simultaneously. For example, Smith et al. found an increased risk for low birthweight associated with PM2.5
exposure, independent of road traffic noise. Traffic noise, however, was no longer associated with the outcome once PM2.5
was controlled for [6
]. Hystad et al. explored the pathways related to the protective effect of greenness on birthweight and found that the beneficial effect of greenness was robust to adjustment for PM2.5
, black carbon, NO2
, noise, walkability, and park proximity [20
Although we observed significant associations of PM2.5
with birthweight in the single exposure models, the effects were attenuated in the multi-exposure model. Moreover, although PM2.5
exposures were selected by the elastic net, the effect sizes were very small. It is possible that our analysis was underpowered to detect a significant effect of PM2.5
exposure in a multivariate model, while adjusting for other environmental exposures. Similar to our findings, Hystad et al. found no association between PM2.5
and birthweight when including NDVI in the model [20
]. Nieuwenhuijsen et al., however, did find a significant association between PM2.5
and birthweight when including other urban exposome exposures in the model. Like our study, which identified neighborhood greenness as one of the most important predictors of differences in birthweight, this study also found the most consistent statistically significant associations between green space exposure and birthweight [21
Low socioeconomic status (SES) has been linked to poor obstetric outcomes in many studies [48
]. Most studies that assessed the effect of segregation on birth outcomes focused mainly on the effect of racial segregation in the neighborhood rather than economic segregation [51
]. One study, by Vinikoor et al., found that being economically privileged is protective against low birthweight only in predominantly black neighborhoods. The Vinikoor study, however, used a different economic segregation measure and assessed the effect of maternal residence in census tracts with a higher household income than would be expected based on the mother’s education and marital status [54
]. Here, we show that residing in a census block-groups with “low income” economic segregation (i.e., high proportion of people living in low income versus high income households) poses a risk for low birthweight deliveries, independent of the mother’s race. Economic dissimilarity was not associated with birthweight in the multivariate model.
Unlike the established association between birthweight and exposure to air pollution and SES, evidence on noise, greenness, walkability, and temperature exposures is scarce and results are inconsistent. We found birthweight to be inversely associated with temperature and nighttime noise and positively associated with greenness. Most previous studies found birthweight decreases associated with hot temperatures [18
], some with cold temperatures [55
], and others observed no association at all [8
]. A recent systematic review found evidence for an association between noise exposure during pregnancy and lower birthweight; however, it concluded that most current studies are of low quality and require better exposure assessment and adjustment for SES and environmental confounders [7
Several studies have demonstrated health benefits of residence in greener neighborhoods [57
]. Although a recent review of residential greenness and birth outcomes found consistent positive associations between neighborhood greenness and birthweight, the associations were weak [16
]. We found increases in birthweight associated with greenness exposure in the first and third trimesters of pregnancy, among women who reside in urban neighborhoods. We also found lower birthweight associated with higher walkability. Similar to our findings, a study in Connecticut found that urbanicity was negatively associated with birthweight [5
The exposures tested in our study can affect birthweight through different mechanisms. Exposure to high temperatures during early stages of pregnancy may cause a reduction in placental weight and umbilical cord flow [4
]. High-temperature exposures during the final stages of pregnancy may also affect birthweight, but the mechanism for this association is yet to be clear [4
]. Exposure to higher air pollution levels may cause oxidative stress, impairing the transport of oxygen and nutrient to the fetus and affect the intrauterine growth [58
]. Exposure to high noise levels can alter the mother’s levels of stress hormones, reduce ovarian and uterine blood flow and inhibit fetal growth [60
]. Residence in a greener neighborhood can have a beneficial effect on fetal growth through mitigation of harmful environmental exposures (i.e., filter air pollution and provide shade), or through the opportunity for social interactions and physical activity, contributing to better maternal health and lower stress levels [9
]. Higher neighborhood walkability can be associated with higher physical activity and therefore a healthier maternal lifestyle and higher birthweight. On the one hand, it can be associated with higher urbanicity, traffic, and pollution levels, and therefore associated with lower birthweight [20
]. To the best of our knowledge, there are no studies that assessed the effect of economic residential segregation on birthweight. However, there is a study that examined the effect of racial segregation on birthweight among black mothers, which found a higher probability of low birthweight in more segregated metropolitan areas [61
]. Residence in neighborhoods with lower SES can increase the risk for lower birthweight through limited access to housing, healthcare services, healthy nutrition, quality education and more [48
]. Similarly, we hypothesize that residence in block groups with higher economic segregation, similar to racial segregation, can have adverse effects on birthweight due to higher levels of poverty and crime, higher exposure to stress, fewer options for purchasing healthy food, and limited access to health care [61
The major strength of our study is the inclusion of a large sample of births, and a wide range of environmental predictors of birthweight. Another strength is the combination of two statistical approaches. The elastic net regression, which allowed us to select important predictors of birthweight while accounting for the complex correlations between these multiple exposures.
The analysis of health outcomes that are affected by a complex net of exposures requires tools that will help distinguish the contribution and weight of each exposure while accounting for other exposures. As the elastic net model has the ability to distinguish between the redundant exposures, and to identify exposures that are important predictors of birthweight, this method is an important tool in environmental research, where health outcomes are affected by multiple exposures, and where exposures are often correlated. Within the pool of important predictors, we also used WQS regression, and identified the exposures that had higher weight in their contribution to a negative difference in birthweight. These results can be of use to expecting families in their choice of place of residence. No place of residence is free of environmental exposures, and there is always a tradeoff of exposures. This tool, however, can provide insights on exposures that contribute more to differences in birthweight, and help focus the efforts to reduce exposure to nighttime noise and temperature, and increase exposure to greenness.
Our study had several limitations. First, we excluded 12% of the records due to missing data. However, this problem is common in studies that use routinely collected data. In addition, as the excluded population was mostly similar to the population included in the analysis, we do not expect it to cause a selection bias. Second, as we assigned exposure based on maternal residence reported at the time of birth, we might have had misclassified exposure for women who changed addresses as conception. It is unlikely that women based their decision regarding the new place of residence on air pollution or temperature exposure. Additionally, it is unlikely that a pregnancy will result in a move to a new residence with a very different socioeconomic environment. We therefore expect the measurement error for these exposures to be non-differential and therefore we do not expect it to bias our results. It is possible, however, that pregnant women will relocate to quieter, greener, and/or less walkable areas. Therefore, for these exposures, the measurement error could potentially bias the results toward the null. Third, we did not have information on other maternal factors that may confound the association between the tested exposures and birth weight. Maternal weight for example was reported to be associated with birthweight [63
], as well as with air pollution, neighborhood greenness, and noise exposure [64
]. Factors such as genetic susceptibility [3
] and substance-using mothers [66
], although related to birthweight, are not likely to be related to environmental exposures and therefore are not likely to confound the associations in our study. Finally, we could not evaluate trimester-specific exposures to walkability and noise, which were only calculated once during the study period. However, as these land use characteristics likely vary very slowly over time, we do not expect it to bias our results.
In addition to the aforementioned limitations, our study is subjected to the limitations of the methods used in our analysis. The elastic net model selects variables with non-zero coefficients, regardless of the effect size and clinical meaning of this effect. In addition, although the elastic net model is capable of identifying important predictors of birthweight, it may have a tendency to incorrectly select a large number of predictors in cases of strong associations with the outcome [67
]. However, as we estimated small environmental effect sizes, we do not expect this limitation to affect the variable selection process in our analysis. Lastly, a multi-phase approach in a statistical analysis may result in model overfitting. However, as the aim of the first stage in our statistical approach (i.e., single exposure models) was descriptive and was not used to select variables, this stage was not a part of the selection process and therefore unlikely to result in overfitting. Furthermore, as the elastic net model found all the covariates as predictors of the outcome, overfitting of the final multivariate regression, which included the exact same covariates, is unlikely in our analysis.
A major limitation of the WQS regression is that it can only include variables for which the effect on the outcome is in the same direction. We therefore could not include economic residential segregation as an exposure in this analysis. In addition, the WQS regression may lose information on the continuous range of exposure due to the use of quartiles. We therefore added sensitivity analyses, assigning different quantiles to exposures to create the weights. Although the results differed depending on the quantiles used, nighttime noise was consistently the most highly weighted exposure. Last, WQS requires strong assumptions of linear effects and lack of interactions between exposures [43