Women were enrolled in the Human Placenta and Phthalates Study prior to 14 weeks’ gestation from prenatal clinics at Eastern Virginia Medical School (EVMS) and the University of Texas Medical Branch (UTMB) from 2017–2018. Women were eligible if they were between the ages of 18–50, carrying a singleton pregnancy, and did not have any detected abnormalities in the fetus, umbilical cord, or in placental location within the uterus. Additional detail on this study has been previously described [20]. Between enrollment and 16 weeks’ gestation, women attended visits every 2 weeks (median 13, 15, 17 weeks). From 16 weeks’ gestation through delivery, women attended clinic visits every 4 weeks (median 21, 25, 29, 33, 37 weeks). At these 8 study visits, 303 women contributed at least one urine sample and completed a product use questionnaire. On average, women contributed 6.5 samples. Procedures were approved by the Institutional Review Boards at EVMS and UTMB and all participants signed informed consent forms prior to participating. Analysis of de-identified samples by the Centers for Disease Control and Prevention (CDC) laboratory was determined not to constitute engagement in human subjects research.

At each visit, women completed a questionnaire that asked them to select products from a list that they had used in the prior 24 hours. The list included vitamins and supplements, personal care products, cosmetics, insect repellants, air sprays, and cleaners. This analysis is restricted to personal care products and cosmetics, which have been linked to phthalate exposure biomarkers most consistently in the literature [19]. The full list of personal care products queried is included in Supplemental Table 1. Where relevant, women also reported if the product used was fragranced, fragrance-free, or the fragrance status was unknown. The questionnaire did not ask about specific brands or formulations.

Due to variability in sample size across study visits, we condensed responses from the 8 visits down to 3, corresponding to early, mid-, and late pregnancy, hereafter referred to as pregnancy periods. Visits 1 and 2 (gestational weeks 13–15) were considered early pregnancy, visits 3–5 (gestational weeks 17–25) were mid-pregnancy, and visits 6–8 (gestational weeks 29–37) were grouped as late pregnancy. Product use responses for the pregnancy periods were populated using the first completed response from the relevant visits (Supplemental Fig. 1). Women were only dropped from the LCA if they had completely blank questionnaire responses and there were no notable demographic differences between women who completed a questionnaire and those who did not.

Urine samples were collected from each woman in sterile polypropylene specimen cups at each visit. Women were instructed not to use any wipes prior to sampling. Specific gravity (SG) was measured using a PAL -10S refractometer. Samples were stored at –80 °C until being shipped overnight and on dry ice to the National Center for Environmental Health laboratory at CDC for analysis. At CDC, urine was stored at or below –40 °C until analysis. Additional detail on collection processes has been previously described [20]. Quantification of phthalate and replacement metabolites involved enzymatic hydrolysis of the metabolites from their conjugated form, automated online solid phase extraction, separation with high-performance liquid chromatography, and detection using isotope-dilution tandem mass spectrometry [21]. The following metabolites were measured: monoethyl phthalate (MEP), mono-n-butyl phthalate (MBP), mono-hydroxybutyl phthalate (MHBP), mono-isobutyl phthalate (MiBP), mono-hydroxy-isobutyl phthalate (MHiBP), monobenzyl phthalate (MBzP), mono-3-carboxypropyl phthalate (MCPP), mono-2-ethylhexyl phthalate (MEHP), mono-2-ethyl-5-hydroxyhexyl phthalate (MEHHP), mono-2-ethyl-5-oxohexyl phthalate (MEOHP), mono-2-ethyl-5-carboxypentyl phthalate (MECPP), mono oxononyl phthalate (MONP), mono carboxyisooctyl phthalate (MCOP), and mono carboxyisononyl phthalate (MCNP). Additionally, four metabolites of replacements were measured: mono-2-ethyl-5-hydrohexyl terephthalate (MEHHTP), mono-2-ethyl-5-carboxypentyl terephthalate (MECPTP), both metabolites of di(2-ethylhexyl) terephthalate (DEHTP), and cyclohexane-1,2-dicarboxylic acid, monohydroxy isononyl ester (MHiNCH), and cyclohexane-1,2-dicarboxylic acid, monocarboxy isooctyl ester (MCOCH). Instrument-reading concentrations below the limit of detection (LOD) were retained if they were >0 and concentrations reported as 0 (i.e., absence of analytical signal) were imputed using LOD/√2, a method shown to be valid in situations with levels of values < LOD similar to what was present in our study [22]. We did not impute instrument-read values < LOD [23, 24].

We calculated molar sums for metabolites of the same parent compounds by summing the molar concentrations (nmol/mL) [20, 25]: MEHP, MEHHP, MEOHP, and MECPP for the sum of di(2-ethylhexyl) phthalate metabolites (∑DEHP); MCNP and MCOP for the sum of di-isononyl phthalate metabolites (∑DiNP); MBP and MHBP for the sum of di-n-butyl phthalate metabolites (∑DnBP); MiBP and MhiBP for the sum of di-iso-butyl phthalate metabolites (∑DiBP); MHiNCH and MCOCH for the sum of 1,2-cyclohexane dicarboxylic acid, diisononyl ester metabolites (∑DiNCH); and MEHHTP and MECPTP for the sum of DEHTP metabolites (∑DEHTP). Molar concentrations were then converted to ng/mL by multiplying the molar sum by the molecular weights of MECPP, MCOP, MBP, MiBP, MHiNCH, and MECPTP, respectively.

All phthalate and replacement exposure biomarkers were corrected for urine dilution using covariate-adjusted standardization in a model that included maternal age (continuous), gestational age at sample collection (continuous), pre-pregnancy body mass index (continuous), maternal education (3 level categorical), and maternal race and ethnicity (3 level categorical) [26, 27]. Natural log-transformed SG was modeled as a function of the above covariates to generate predicted SG values. Exposure biomarkers were then divided by the ratio of observed SG to predicted SG. The urine dilution standardization was conducted after the calculation of the molar sums. Maternal age, education, height, and race and ethnicity were self-reported at enrollment. Maternal weight was extracted from medical records and gestational age at sample collection was recorded as part of the study protocol.

To create more stable estimates of exposure within pregnancy periods, we condensed phthalate and replacement biomarker concentrations from the 8 visits into 3 time points, using the same timing scheme described for the questionnaire. We calculated the geometric mean of all available measurements within each pregnancy period, and, if only one concentration was available, that was used in place of the geometric mean.

Maternal race was self-reported with options of Caucasian, Black or African-American, Asian, Native Hawaiian/Other Pacific Islander, or “other.” Women also indicated their ethnicity as Non-Hispanic/Latina, Hispanic/Latina, or “refuse to disclose.” We created a composite race/ethnicity variable with levels of “non-Hispanic White”, “non-Hispanic Black”, “Hispanic, any race”, and “Asian, Native Hawaiian/Other Pacific Islander, or Other.” Highest education achieved was captured as high school graduate or below; some college, technical school, or associates degree; 4-year college degree. Insurance status was categorized using enrollment values as private, or self-pay/uninsured/government-assisted.

All analyses were conducted in SAS 9.4 (Cary, NC, USA). We examined distributions of phthalate and replacement biomarker concentrations by pregnancy period, using SG-corrected and period-averaged values.

Latent class analysis (LCA) was used to identify groups of individuals with similar patterns of personal care product use among participants within each pregnancy period using PROC LCA in SAS 9.4. The number of classes was selected based on model fit statistics like Akaike and Bayesian Information Criterion, membership percentages, posterior probabilities, and interpretability of classes [28]. To reduce the number of items included in the LCA model, we used an a priori cut-off to drop items with less than <10 women reporting use at each visit (Supplemental Table 1). The remaining products were evaluated based on whether they had previously been associated with phthalates in the literature. If a product was lowly used, not previously linked to phthalates, or strongly racially colinear, it was excluded. Additionally, similar products were combined (e.g., conditioner and leave-in conditioner). In the cases of these composite variables, the participant was classified as a fragranced user if one instance of fragranced use was noted. Conversely, both products in a combined variable had to be reported as no use or fragrance-free use for a participant to be coded as those respective responses. For early pregnancy, responses were only available at one visit for 19-30% of women, varying by product. Of the remaining women, approximately 60% had the same responses for product use at each visit. In mid- and late pregnancy, approximately 40% of women had concordant values at all 3 visits, while an additional 30-40% had either a response only at one visit, or a response at two visits with those responses in agreement (data not shown).

Product responses were coded using one 3-level variable with levels corresponding to: no use, fragranced use (reported yes to use in the last 24 hours, and reported that the product was fragranced), and fragrance-free use (reported yes to use in the last 24 hours and reported that the product was fragrance-free). If the use of a product was reported but the participant also reported “fragrance unknown,” the response for the item was set to missing. If one product in a composite variable was indicated as “fragrance unknown”, we used the available response to determine the fragrance status. This was only applicable for the hand soap variable, as the other composite variables either did not have a fragrance component (i.e., cosmetics) or were not included in the main analysis (i.e., conditioner). Cosmetics and perfume were coded dichotomously as any use vs. none.

Using bar charts, we reported the number of identified groups in each pregnancy period, the proportion of total participants in each group, and the proportion of participants within each group who reported no use, fragrance-free use, or fragranced use of each personal care product included in the model. “Any use” for cosmetics and perfume was coded as “fragranced use” for the visualization only since the questionnaire did not specify whether these products were fragranced and responses had to be represented visually. Furthermore, perfume is, by nature, fragranced and many cosmetics include fragrance either to include a scent or to mask the scent of other ingredients [29].

Latent transition analysis (LTA) was then conducted to assess changes in class membership across pregnancy using PROC LTA. LTA uses data from all time points to determine class profiles and identify item response probabilities. Output of LTA includes conditional item response probabilities, prevalence of latent classes at each time point, and transition probabilities, representing the likelihood of membership in each given class conditional on membership in a specific class at the previous time point. We compared the demographic distributions of women who remained in the same product use groups between consecutive pregnancy periods and those who moved between groups.

To assess the relationship between product use group membership and measured phthalate and replacement biomarker concentrations, we compared mean biomarker concentrations across groups. Mean concentrations were calculated from linear regression generalized estimating equations. Models were weighted using inverse probability of treatment weights so that each latent class group was standardized to the distribution of covariates in the overall population at each pregnancy period. Thus, this weighting allowed a more direct comparison of mean urinary biomarker concentrations within groups to the overall population. Without weights, differences could be attributable to demographic differences across groups. Inverse probability weights were based on maternal race and ethnicity, age, highest educational achievement, and health insurance status. Women missing information on any of the included covariates were dropped from the model (n = 8, 7, 6 women at each point in pregnancy, respectively). Because product use variables representing the pregnancy period were derived from different visits for different items, it was not possible to align concentrations and product responses to be used from the same visit. As such, we used biomarker concentrations representing the entire pregnancy period.

To visualize the associations between product use group and biomarker concentrations, we created heat maps displaying the relative difference in biomarker concentration for women in the product use group compared to the period-mean concentration. Tables are separated by biomarkers of low molecular weight (LMW) phthalates, consisting of MEP, ∑DnBP, and ∑DiBP, high molecular weight (HMW) phthalate biomarkers, consisting of MBzP, MCPP, ∑DEHP, ∑DiNP, and MCNP, and replacement biomarkers, consisting of ∑DiNCH and ∑DEHTP.

As a sensitivity analyses, we repeated analyses with the next best candidate models from the latent class analyses. Findings from LCA rely heavily on the user-selected model and rerunning analyses using an alternative model allows us to assess how much, if at all, conclusions change. We reran LCA using different subsets of products from the questionnaire to identify robustness of groups, as well as rerunning results retaining the “fragrance unknown” response instead of setting it to missing.