Study population
This work takes place in a 214-participant personal PM2.5 exposure monitoring sub-study nested within the Maternal and Developmental Risks from Environmental and Social Stressors (MADRES) study, an ongoing prospective cohort study of just over 1000 pregnant, primarily Hispanic, low-income pregnant women in Los Angeles County41. MADRES aims to investigate the cumulative impact of environmental pollutants and psychosocial, behavioral, and built environmental risk factors on maternal and infant health outcomes as described in more detail elsewhere41. Briefly, pregnant women were enrolled in the cohort through partnerships with four prenatal care clinics in Los Angeles, CA from November 2015. Eligibility for this study included: (1) at least 18 years old, (2) a singleton birth, (3) less than 30 weeks gestation at recruitment, (4) HIV negative, (5) having no physical, mental, or cognitive disability that would prevent the participant from providing informed consent, and (6) not currently incarcerated.
Of the 214 participants in the personal exposure monitoring study, nine were removed due to incomplete or erroneous personal PM2.5 mass exposure data or birth outcome data. Four participants did not have PM2.5 source data and were removed from the analysis. A multivariate k-nearest neighbor outlier detection analysis revealed three outliers in terms of personal exposure to the six sources. These were excluded from further analysis. However, given these points were very influential in the models, results including and excluding them are presented side-by-side for completeness in this analysis. This resulted in a sample of 198 mother-infant dyads used in the final models (201 in the outlier-included models).
Participants were recruited by trained, bilingual MADRES staff members during a 3rd trimester visit to the University of Southern California (USC) clinic, where consenting women were asked to participate in the in-utero personal exposure monitoring sub-study for a 48-h monitoring period. This sample was comparable to the larger MADRES cohort on key demographics, birth outcomes, and ambient air pollution metrics.
Personal PM2.5 exposure monitoring
Total personal PM2.5 exposure was measured over an integrated 48-h monitoring period in the 3rd trimester using a custom-designed sampling protocol between October 2016 and February 2020. The 3rd trimester was chosen because most fetal weight gain occurs in this trimester25. Participants were shown and provided with instructions by trained staff members on the correct use of the personal exposure monitoring device, which was housed in a crossbody purse. Instructions included a demonstration of how to wear the purse, making sure to keep the sampling inlet located on the purse shoulder strap free from obstructions and in the breathing zone. Additionally, participants were requested to wear the device as much as possible during normal daily activities, with a limited number of exceptions, including driving, showering, sleeping, etc. Participants were asked to keep the sampling device safe and away from water, high humidity (such as showering), heat, children, and pets, and when unable to wear the device, place it as near as possible, such as on the passenger seat if driving, and a side-table while sleeping.
The purse contained a Gilian Plus Datalogging Pump (Sensidyne Inc.) connected to a Harvard PM2.5 Personal Environmental Monitor (PEM) with a pre-weighed 37 mm Pall Teflo filter. The device was programmed to start at midnight the day after enrollment into the sub-study, and actively sampled at a 50% cycle and a 1.8 L per minute (LPM) flow rate. The sampling device was programmed to shut off after the 48-h sampling period and collected by staff members the following day when a brief exit survey was conducted. The devices were then transferred to the USC Exposure Analytics lab for analysis. Pump data was downloaded, evaluated for errors, and stored securely. Filters were equilibrated within a dedicated chamber and gravimetrically weighed in temperature and relative humidity-controlled glove box using an MT-5 microbalance (Mettler Toledo, Inc.) to obtain PM2.5 mass concentration reported in μg/m3. The methodology of this personal monitoring study has been described in greater detail elsewhere42.
Elemental speciation analysis using X-ray fluorescence
Elemental data was obtained via an X-ray fluorescence analysis43 that determined the elemental composition of PM2.5 collected on personal sampling filters. Concentrations of elements (reported in ng/m3) identified in the source apportionment analysis44 as markers or high-loading species in the source profiles were used in this current analysis. These included: aluminum (Al), barium (Ba), bromine (Br), calcium (Ca), chlorine (Cl), copper (Cu), magnesium (Mg), nickel (Ni), silicon (Si), sodium (Na), sulfur (S), titanium (Ti), vanadium (V), and zinc (Zn).
Optical carbon fractions analysis
A multiwavelength optical absorption approach was used to measure concentrations of several carbon fractions (reported in μg/m3) in the personal PM2.5 samples, including: (1) Black Carbon (BC), (2) Brown Carbon (BrC), and (3) Environmental Tobacco Smoke (ETS). This method is described in more detail elsewhere45, and its performance is consistent with other carbon apportionment approaches45. Briefly, this method uses an integrating sphere radiometer which measures the difference in absorption of transmitted light passed through a mass-loaded Teflo filter. Each of the three carbon components measured with this approach has a different optical density at varying wavelengths, which allows for the identification and quantification of their respective concentration from their optical properties. For the purposes of this study, ETS refers to the carbon fraction concentration obtained via this multiwavelength optical analysis, while the secondhand smoke (SHS) source (explained below) refers to one of the six major contributing sources of personal PM2.5 identified in the PMF analysis. This source had high loadings of several different but highly correlated components, namely ETS and BrC.
Personal PM2.5 sources
Six major contributing sources of personal PM2.5 were used in this analysis, obtained from an earlier source apportionment analysis of these personal exposure filter samples using the EPA Positive Matrix Factorization model (EPA PMF v5.0)44. The PMF analysis used PM2.5 mass and the concentrations of 36 components (33 elements and 3 optical carbon fractions) as inputs to derive the six sources and their predicted mass contributions. The elements and carbon species were obtained from X-ray fluorescence (XRF) and multiwavelength optical absorption carbon speciation analyses, respectively, at the Research Triangle Institute International, Inc (described in more detail below). For this current study, only 17 (14 elements and 3 optical carbon fractions) high-loading components or signature tracers of the six sources (noted in parentheses) were investigated with birthweight including:
(1) Traffic (BC, Zn, Ba), (2) Secondhand Smoke (BrC, ETS, Br), (3) Aged Sea Salt (S, Na, Mg), (4) Fresh Sea Salt (Cl, Na, Mg), Fuel oil (Cu, Ni, V), and Crustal (Si, Ca, Ti, Al).
Birthweight outcome
Infant birthweight (grams) was abstracted from participants’ electronic medical records (EMR). Given that birthweight and gestational age are closely linked, birthweight-for-gestational age z-scores that were either sex or parity specific were also assessed, as described in46. However, the results were not materially different from continuous birth; therefore, only the continuous birthweight models are presented.
Covariate data
Possible covariates were determined a priori from the air pollution and birth outcomes literature. Factors assessed included maternal demographics, pregnancy and birth outcomes, study design characteristics (such as hospital of birth), and meteorology. Self-report data were collected during the MADRES cohort follow-up through a sequence of staff administered in-person and telephone-based questionnaires. All questionnaires were available in either English or Spanish. Anthropometric assessments were conducted via regular clinic visits. Due to the timing of this personal monitoring study in the 3rd trimester of pregnancy, data primarily came from the 3rd trimester visit, the personal monitoring study exit survey, and birth-related datasets and variables, with additional variables assessed at the baseline visit including race/ethnicity and pre-pregnancy body mass index (BMI; kg/m2).
Additional pregnancy and birth-related covariates were also evaluated for confounding. Gestational age at birth (GA; weeks) was estimated with a hierarchical approach of differing methods from the preferred ultrasound measurement of crown-rump length at < 14 weeks gestation (60%), ultrasound measurement of fetal biparietal diameter at < 28 weeks’ gestation (30%), and from physicians’ clinical estimate from EMR (10%). Parity was defined as 1 or more previous births and included a missing category with 6 participants so as not to lose sample size. Infant sex was obtained through EMR.
Maternal demographics included the following: Age at baseline (continuous; years), education (completed < 12th grade, completed high school, at least some college), household income (less than $15,000, $15,000–29,999, $30,000+, don’t know), and diabetes status [no diabetes, glucose intolerant, diabetes (chronic and gestational)]. Race/ethnicity was categorized as Hispanic, non-Hispanic Black, and non-Hispanic Other. Pre-pregnancy BMI (continuous; kg/m2) was calculated from self-reported pre-pregnancy weight and standing height measured by MADRES staff at participants’ first visit by either a stadiometer (Perspectives model PE-AIM-101) or EMR. Self-report weight was chosen because participants entered the study at differing weeks of gestation.
Meteorological factors evaluated in this study included ambient air temperature (Celsius) (calculated as the average of minimum and maximum air temperature) and relative humidity (%), averaged over the 3rd trimester and estimated at the residential location based on a high-resolution (4 km × 4 km) gridded surface meteorological dataset 47, and seasonality using the following levels, “Cool (Winter), Warm (Summer), and Transition (Spring and Autumn)”.
Statistical analysis
Descriptive statistics
Descriptive statistics of key sample demographics and mean and standard deviations were calculated for concentrations of personal PM2.5 mass, six PMF-derived sources of personal PM2.5, and 17 high-loading components. The distribution of birthweight, personal PM2.5 mass concentration, and each source and component were investigated to identify any issues with normality and potential influential points. Bivariate analyses using Kruskal–Wallis one-way analysis of variance tests and Spearman’s correlation coefficients were conducted between personal PM2.5 mass, its major contributing sources, and birthweight by various sample characteristics to elicit any additional potential confounders for our regression analysis, in addition to those identified in previous literature27.
Linear regression models
Single- and multi-pollutant linear regression models were used to investigate the primary aim of this study, that is, the association between in-utero exposure to major personal PM2.5 sources and birthweight, adjusting for gestational age at birth, maternal age, race/ethnicity, infant sex, parity, diabetes status, temperature, maternal education, and personal smoking history. Even though this study assessed SHS as a source of PM2.5, it did not correlate strongly with our smoking covariate (never/ever smoker). However, this smoking covariate did seem to be a confounder and impact our main effects, therefore, it was kept within the model. Meteorological variables, such as season, were excluded from the models after statistical confounding checks showed no material differences in the main effects. This was done to conserve statistical power. The effect of total personal PM2.5 on birthweight, previously reported by this group32, was included in relevant tables for comparison purposes. PM2.5 sources that were not highly correlated with one another, as determined by a bivariate Spearman correlation analysis and/or a variance inflation factor (VIF) below 10 in the regression, were used to evaluate the effect of each source on birthweight, adjusting for one or more other PM2.5 sources. Multi-pollutant models were conducted with up to four personal PM2.5 sources included in each model; however, three- and four-pollutant model results did not materially differ. Therefore, only single- and two-pollutant models are reported. Additionally, the association between PM2.5 and birthweight has been shown to differ by the sex of the infant, therefore, this study evaluated whether the effect of each PM2.5 source exposure on birthweight was modified by sex. Non-linear effects were evaluated by modeling each PM2.5 source on birthweight using generalized additive models (GAMs) and assessing Akaike information criterion (AICs) vs. linear regression models. As a sensitivity analysis, the association of each PM2.5 source on birthweight was examined only among full-term births (37 weeks or older gestation) due to apprehensions about potential bias from factoring in gestational age as a confounder, given the possibility that it might act as a mediator48. This was conducted with and without the inclusion of gestational age as a covariate.
For the secondary aim of this study, this study attempted to determine whether it was the source itself (a particular mix of components) driving any adverse health effects or an individual element or component. To evaluate whether it is the PM2.5 source (the mixture) or any of its high-loading components that are driving the observed association between sources and birthweight, the effect of the 17 high-loading PM2.5 components on birthweight was investigated. Additionally, because PM2.5 mass concentration may be related to both birthweight and the concentration of the PM2.5 components (especially more abundant ones), further analyses adjusting component models for PM2.5 mass were performed via two approaches, which have different interpretations, while similarly attempting to account for PM2.5 mass itself. The first was by adjusting for PM2.5 mass in the individual component models by directly including it as a simple covariate or potential confounder. According to Mostofsky et al.49, this parameter “represents the impact of higher levels of the constituent (and its correlates), holding the other constituents constant”49. The second approach was to create component residuals by regressing the mass concentration of each PM2.5 component (dependent) on the total PM2.5 mass concentration (independent). The component coefficient “represents the increase in risk associated with higher levels of the constituent while holding PM2.5 constant”49.
Due to concerns with outliers being influential as determined by model diagnostics in 3 out of 6 main source models, a multivariate K-nearest neighbor outlier detection analysis was conducted in JMP Pro 16 (SAS Institute, Inc., Cary, NC, USA). This was used to identify outliers up to a distance of 8 nearest neighbors along the concentrations of all six personal PM2.5 sources. This analysis allowed us to objectively identify data points that were materially different from the overall sample across six dimensions. All effect estimates and 95% confidence intervals were scaled and reported per 1 SD increase in the main exposure of interest. An alpha of 0.05 was selected as a priori significance level for our main exposure/outcome analyses, while 0.10 was used for the infant sex interaction analyses to allow more leniency for statistical interactions due to the increased power requirement. Model diagnostics were conducted to ensure models were not affected by multicollinearity or influential points. The analysis was conducted using SAS v9.4 (SAS Institute, Inc., Cary, NC, USA).
Ethical approval and consent to participate
Study procedures were approved by the USC Institutional Review Board (IRB) and all participants completed written informed consent at the first study visit (IRB: HS-16-00530). This study was performed according to the ethical guidelines expressed in the Strengthening of the Reporting of Observational Studies in Epidemiology (STORB) guideline and the Declaration of Helsinki.
