Systematic Review and Meta-analysis of Birthweight and Perfluorohexane Sulfonate Exposures: Examination of Sample Timing and Study Confidence

Abstract

Objective:

We examined the association between mean birth weight (BW) differences and perfluorohexane sulfonate (PFHxS) exposure biomarkers.

Methods:

We fit a random effects model to estimate the overall pooled effect and for different strata based on sample timing and overall study confidence. We also conducted an analysis to examine the impact of gestational age at biomarker sample timing.

Results:

We detected a −7.9 g (95%CI:−15.0, −0.7;p_Q=0.85;I2=0%) BW decrease per ln ng/mL PFHxS increase based on 27 studies. The 11 medium confidence studies (β=−10.0g;95%CI:−21.1, 1.1) showed larger deficits than 12 high (β=−6.8g;95%CI:−16.3, 2.8) and 4 low (β=−1.5g; 95%CI:−51.6, 48.7). Ten studies with mid- to late-pregnancy sampling periods showed smaller deficits (β=−3.9g;95%CI:−17.7, 9.9) than 5 post-partum studies (β=−28.3g;95%CI:−69.3, 12.7) and 12 early sampling studies (β=−7.6g;95%CI:−16.2,1.1). Six of 12 studies with the earliest sampling timing showed results closer to the null.

Conclusions:

Overall, we detected a small but statistically significant BW deficit across 27 studies. We saw comparable BW deficit magnitudes in both the medium and high confidence studies as well as the early pregnancy group. Despite no definitive pattern by sample timing, larger deficits were seen in post-partum studies. We also saw results closer to the null for a subset of studies restricted to the earliest biomarker collection times. Serial pregnancy sampling, improved precision in gestational age estimates, and more standardized reporting of sample variation and exposure units in future epidemiologic research may offer greater understanding of the relationship between PFHxS on BW and any potential impact of pregnancy hemodynamics.

Introduction

Perfluorohexane sulfonate (PFHxS) is a legacy per- and poly-fluoroalkyl substances (PFAS) that is common in human biomarkers across different lifestages, since it can be found in firefighting foam, drinking water, dietary sources as meat and fish, and various other consumer and industrial products.[1, 2] Different PFAS are associated with a variety of adverse health endpoints in epidemiological and toxicological studies, including immunological, hormonal, reproductive, and developmental effects.[3, 4] Developmental effects are of particular concern since in utero exposures can lead to impacts in offspring at birth but also can have long-term and transgenerational health effects.[5, 6] Despite declining levels of legacy PFAS in biomarkers there is an increasing need to determine safe levels to address cleanup of environmental media such as contaminated soil and water. An ongoing challenge in characterizing risk from PFHxS exposure is the long half-life of the chemical, estimated at 8.30 years.[7] This results in accumulation of PFHxS, so that biomarker measurements reflect a combination of recent and previous exposure; this is of special concern for a developing fetus who is exposed to a PFHxS concentration in utero that has accumulated over the mother’s life. Developmental exposure is also more challenging to assess because the physiological changes that occur during pregnancy can affect the pharmacokinetics of PFAS (e.g., increased blood plasma volume due to decreased mean arterial pressure, increased cardiac output, and systemic vasodilation). In general, while most PFAS have been shown to decrease in serum or plasma concentration as pregnancy progresses, PFHxS has most often been shown to increase or remain constant.[8–11] Notable is Taibl et al.[11], which showed increases in many PFAS, and a more than doubling of PFHxS between the 1^st and 3^rd trimesters. While decreases in PFAS serum concentrations during pregnancy have been attributed to increased clearance, decreased plasma binding protein concentration or dilution due to increased plasma volume, such a sharp increase is difficult to explain though the influence of these physiological changes alone. More research, like more standardized serial sampling, is needed to determine whether temporal exposure differences during pregnancy noted across cohorts are due to actual changes in exposure, physiology, or pharmacokinetics or a combination of these factors. These differences have been difficult to disentangle in existing systematic review efforts PFAS meta-analyses especially since existing meta-analyses on PFHxS are based on a limited by disparate inclusion criteria resulting in few epidemiological studies available for pooled estimates.

To examine potential effects of PFHxS on mean birth weight (BW), we conducted a systematic review of the published literature and assessed the risk of bias and sensitivity related to eligible studies that met our inclusion criteria. We also conducted various sub-group and sensitivity analyses to highlight study characteristics that may explain between-study heterogeneity, including those identified in previous reviews/meta-analyses of other PFAS. For example, this included the challenge of evaluating the potential impact from hemodynamic changes during pregnancy based on differences in biomarker sample timing.

Methods

Systematic review methods

This systematic review was conducted in accordance with US EPA’s PFAS systematic review protocol[12] based on methods detailed in the ORD Staff Handbook for Developing IRIS Assessments[13]. The literature search and study evaluation results are detailed in the Supplemental file. The following databases were searched: PubMed (National Library of Medicine), Web of Science (Thomson Reuters), Toxline (National Library of Medicine), and TSCATS (Toxic Substances Control Act Test Submissions). Publications were identified by conducting a broad literature search of PFHxS exposure using the search terms and databases listed in Supplemental Table 2. The initial search took place in 2017, with annual updates until the final search was performed in April 2023. Results were screened by two independent reviewers with a process for conflict resolution at two stages. First, at the title and abstract level, and subsequently at the full-text level, using structured forms in DistillerSR (Evidence Partners; https://distillercer.com/products/distillersr-systematic-review-software/). Literature inventories for Populations, Exposures, Comparators, Outcomes (PECO) relevant studies and additional studies tagged as “potentially relevant supplemental material” during the screening phases were created to facilitate subsequent review of various sets of studies by different topic specific experts; studies of birth weight were tagged at this stage.

Critical appraisal of studies was performed using the US EPA’s IRIS approach for study evaluation as detailed in US EPA PFAS protocol[12]. At least two independent reviewers rated each study based on potential risk of bias and study sensitivity using the following domains: Participant Selection, Exposure Measurement, Outcome Ascertainment, Confounding, Analysis, Selective Reporting, and Study Sensitivity[12]. They determined a consensus rating specific to each of these domains as well as the overall confidence level. Ratings for each domain were either Good, Adequate, Deficient, or Critically Deficient, while overall confidence ratings included high, medium, low, or uninformative. The overall confidence ratings were based on the reviewer’s expert judgment of the likely impact that the identified domain limitations had on the results; no pre-defined weighting of domains or numerical scores were used. Details on the ratings, which reflect a consensus judgment between reviewers are found in the ORD Staff Handbook for Developing IRIS Assessments[13]. For example, study sensitivity addresses the extent that factors in the design and conduct of the study might reduce its ability to observe an effect if present.

Inclusion criteria for meta-analysis

This meta-analysis included epidemiological studies which reported regression coefficients (i.e., “betas” or βs) for the association between mean BW changes and PFHxS exposure concentrations. Included studies were also required to report 95% confidence intervals (CIs) or other measures of variance such as a standard error or a p-value. PFHxS exposure must have been measured as concentrations in umbilical cord or maternal blood collected before, during or after pregnancy. The study population could either be the overall population (i.e., combined male and female offspring) or stratified by sex. Studies which only reported effect estimates for categorical exposure groupings were included only in a sensitivity analysis, and only after converting and combining the effect estimates per each non-referent exposure quantile to a continuous measure as was done in Steenland et al.[14].

Data pre-processing

Following study screening and evaluation, two data pre-processing steps were carried out prior to the meta-analysis. First, the βs and 95% CIs of the mean BW changes were mathematically rescaled within either the logarithmic or the natural scale, or re-expressed between scales from the natural scale to grams (g) per natural logarithmic (ln) PFHxS exposure unit (ng/ml) according to Dzierlenga et al.[15]. Additional details were provided in the Supplemental file. Second, for those studies only reporting sex-specific estimates for boys and girls, effect estimates for the overall study population were estimated by combining sex-specific results using inverse-variance weighting.

Statistical analysis

Data pre-processing, including rescaling and re-expression, was conducted using SAS Version 9.4 (SAS Institute Inc., Cary, NC, USA). All meta-analytic techniques were carried out using the meta-analysis package, metafor, in R Version 4.0.3 (R Core Team 2020).

Overall meta-analysis

The meta-analysis was performed using random effect models, estimated through restricted maximum likelihood method, with inverse-variance weighting. This allowed for differences in study designs, populations, and other unknown factors to be included as random effects, minimizing the influence of sampling variance and between-study variance (i.e., heterogeneity) on the pooled effect estimates. The a priori choice of a random effects model was intended to best reflect the assumption that each study produced an estimate of a study-specific true effect that varies across studies.[16] Two types of heterogeneity measures, Cochran’s Q test statistics and p-value, and Higgin’s and Thompson’s I², were reported. Publication bias was evaluated visually using enhanced funnel plot asymmetry[17] and the Egger’s regression test[18] with different alpha levels (i.e., p<0.01; 0.01–0.05; 0.05–0.10; >0.10).

Stratified analyses

To evaluate the effect of overall study confidence and biomarker sample timing on the relationship between BW deficits and PFHxS, studies were binned into groups for stratified meta-analyses. All stratified meta-analyses were carried out using the same approach as the overall meta-analysis. Statistical tests for differences among strata were performed using fixed effect models according to the method of Borenstein et al.[16]. However, strata-specific statistical tests should be interpreted with caution due to small sample sizes in certain subgroups. Studies were assigned into three confidence levels (e.g., high, medium, and low) based on study evaluation for subgroup analyses. In addition to the three-strata classification for study confidence, stratified meta-analyses were also carried out for high and medium confidence studies combined, contrasted with low confidence studies (i.e., a two-strata approach). Studies were binned into three sample timing strata (early-pregnancy, mid-to late-pregnancy, and post-pregnancy) based on the timing of biomarker collection (see Supplemental Table 4 for details on sample timing distributions and strata assignments). Stratified meta-analyses were also carried out for mid-to late-pregnancy and post-pregnancy studies combined, contrasted with early pregnancy studies as a two-strata classification approach for sample timing. Early pregnancy studies are defined as those with any trimester 1 sampling or earlier (e.g., pre-conception, Trimester 1-only, Trimester 1 and 2, and Trimester 1 to 3). Mid- to late-pregnancy sampling is defined as the remaining maternal sampling during pregnancy (e.g., Trimester 2-only; Trimester 2 and 3; Trimester 3-only; Trimester 2, 3 and at delivery).

Meta-regression

Biomarker sample timing has previously been identified as potentially influencing the size of BW deficits while evaluating the association between PFAS exposure and fetal growth restriction[14, 15]. To further test this hypothesis, a meta-regression, with a continuous measure of sample timing as an explanatory variable, was carried out using a mixed effect model as detailed in the Supplemental file. The continuous measure of sample timing was calculated for each study based on available central tendency measures of gestational age (i.e., mean, median, range midpoints, weighted mean of means/medians/midpoint and exact week when samples were collected) (see Supplemental Table 4 for details on calculations).

Sensitivity and additional analyses

In addition to the overall and stratified meta-analyses, subgroup analyses were also conducted within the set of studies with early sample timing studies by reclassifying them into very early pregnancy (i.e., predominately first trimester and/or smaller estimates of mean, median or mode of gestational age at sample timing) versus the remaining early sample timing studies. To further examine the impact of both sample timing and study quality on magnitude and direction of meta-effect estimate, early sample timing studies were also compared to mid- or late- sample timing studies within the high confidence group. Uncertainty may be introduced during the process of re-expressing effect estimates for some studies into ln units[15, 19]; therefore, subgroup analyses were performed for studies that did not involve re-expression and for studies that reported log-based results. Finally, a sensitivity analysis was conducted to weigh the benefit of increasing the meta-analysis sample size by including studies that only reported categorical results. The results were converted to continuous measures of BW deficits according to the method of Dzierlenga et al.[15] and added to the group of studies considered in the main meta-analysis. Lastly, we conducted an exploratory analysis of regional groupings of different study populations to further examine between-study heterogeneity.

Results

Literature Search and Study Inclusion

The literature search, completed in April 2023 from the databases listed in Supplemental Table 2, yielded 415 human health effect studies that met the PECO eligibility criteria (See Supplemental Table 1 for details) after full text screening. Among the studies identified through the literature search, as well as additional reference screening and author knowledge, 36 non-overlapping developmental epidemiological publications of PFHxS and mean birth weight that met our inclusion criteria. Five studies were classified as uninformative; seven were low, 11 were medium and 13 were high confidence, respectively (Figure 1 and Supplemental Figure 1). Additional details on study evaluation results were provided in the Supplemental file. Of the 31 informative epidemiological publications with mean BW data, three[20–22] reported categorical results only. Our primary analysis was restricted to BW studies that were most comparable, i.e., those based on continuous PFHxS exposures. Among the 28 publications with results based on continuous data, 24 provided effect estimates in the overall population (i.e., male and female combined). Data were pooled for the four publications that only reported sex-specific findings[23–26] within each manuscript using inverse-variance weighting to provide an effect estimate in each study’s overall population. This resulted in the inclusion of 27 nonoverlapping informative PFHxS studies (from 28 publications; detailed in Supplemental file including Supplemental Figure 2 and Supplemental Table 3) with continuous exposure expressions for the meta-analysis.

Figure 1.

Summary of study evaluations for studies of PFHxS and birth weight.

Of the 27 studies included here, six studies reported the effect estimates on the natural scale or other transformation of exposure (i.e., Buck Louis et al.[27] reported results per SD increase in ln-transformed of 1 plus exposure units). The results from these studies were first rescaled, if necessary, and then re-expressed into ln units (i.e., per ln (ng/mL)). The original study results, exposure distributions, and the rescaled or re-expressed results for the 27 studies included in the meta-analysis are shown in Table 1.

Table 1:

Original study and re-expressed results based on reported PFHxS exposure distributions in 27^a studies (from 28 publications) used in primary analysis.

			Original β		Rescaled or Re-expressed β (g/ln(ng/ml))
	Exposure Summary	Exposure Distribution in
Study	Mean/q50/GM (Dispersion)	ng/mL (µ, σ)	β (LCL, UCL)	Unit	Beta (LCL, UCL)
Ashley-Martin, 2016[23] Female	1.00 (0.70–1.60 IQR ng/ml)	(0, 0.61)	−24.7 (−140, 90.6)	g/log10(ng/ml)	−10.7 (−60.8, 39.3)
Ashley-Martin, 2016[23] Male	1.00 (0.70–1.60 IQR ng/ml)	(0, 0.61)	53.7 (−53.7, 161.2)	g/log10(ng/ml)	23.3 (−23.3, 70.0)
Bach, 2016[44]c	0.47 (0.36–0.63 IQR ng/ml)	(−0.76, 0.41)	−11.0 (−32.0, 9.0)	g/IQR (ng/ml)	−19.4 (−56.3, 15.8)
Buck Louis, 2018[27]d	0.71 (0.44–1.23 IQR ng/ml)	(−0.35, 0.75)	−17.1 (−40.7, 6.5)	g/SD ln(1+ng/ml)	−22.1 (−52.5, 8.4)
Callan, 2016[45]	0.47 (0.44 SD ng/ml)	(−1.07, 0.79)	−72.0 (−194.0, 50.0)	g/ln(ng/ml)	−72.0 (−194.0, 50.0)
Chang, 2022[46]	1.02 (1.90 GSD ng/ml)	(0.02, 0.64)	−14.0 (−58.0, 31.0)	g/log2(ng/ml)	−20.2 (−83.7, 44.7)
Chen, 2021[10]	0.64 (0.47–0.99 IQR ng/ml)	(−0.45, 0.55)	27.6 (−64.7, 119.9)	g/ln(ng/ml)	27.6 (−64.7, 119.9)
Gyllenhammar, 2018[41]	2.40 (1.40–4.00 IQR ng/ml)	(0.88, 0.78)	−53.3 (−104.5, −2.1)	g/ln(ng/ml)	−53.3 (−104.5, −2.1)
Hamm, 2010[47]	1.10 (3.00 GSD ng/ml)	(0.10, 1.10)	21.9 (−23.4, 67.2)	g/ln(ng/ml)	21.9 (−23.4, 67.2)
Hjermitslev, 2020[48]	0.51 (0.34–0.80 IQR ng/ml)	(−0.67, 0.63)	−93.0 (−230.0, 44.1)	g/ng/ml	−48.6 (−120.2, 23.1)
Kashino, 2020[49]	0.30 (0.20–0.40 IQR ng/ml)	(−1.20, 0.51)	−3.0 (−60.5, 54.4)	g/log10(ng/ml)	−1.3 (−26.3, 23.6)
Kwon, 2016[50]	0.34 (1.81 GSD ng/ml)	(−1.08, 0.59)	−60.1 (−136.4, 16.3)	g/ln(ng/ml)	−60.1 (−136.4, 16.3)
Lenters, 2016[51]	1.84 (1.24 2SD ln(ng/ml))	(0.61, 0.62)	−6.3 (−55.2, 42.6)	g/2SD ln(ng/ml)	−5.1 (−44.5, 34.3)
Li, 2017[52]	4.39 (3.01 SD ng/ml)	(1.29, 0.62)	−30.0 (−83.4, 23.5)	g/ln(ng/ml)	−30.0 (−83.4, 23.5)
Lind, 2017[24] Male	0.30 (0.20–0.40 IQR ng/ml)	(−1.20, 0.51)	9.0 (−71.0, 89.0)	g/ln(ng/ml)	9.0 (−71.0, 89.0)
Lind, 2017[24] Female	0.30 (0.20–0.40 IQR ng/ml)	(−1.20, 0.51)	0 (−65.0, 64.0)	g/ln(ng/ml)	0 (−65.0, 64.0)
Luo, 2021[53]	10.36 (7.67–13.44 IQR ng/ml)	(2.34, 0.42)	−12.5 (−106.8, 81.8)	g/ln(ng/ml)	−12.5 (−106.8, 81.8)
Manzano-Salgado, 2017[42]	0.58 (0.37 SD ng/ml)	(−0.72, 0.58)	−8.6 (−32.0, 14.8)	g/log2(ng/ml)	−12.4 (−46.2, 21.4)
Maisonet, 2012[25] Femaleb	1.60 (1.30–2.00 33–67 PCTLR ng/ml)	(0.47, 0.49)	−35.7 (−105.4, 34.1)	g/ng/ml	−57.9 (−171.1, 55.3)
Marks, 2019[26] Maleb	1.90 (1.40–2.50 IQR ng/ml)	(0.64, 0.43)	−5.2 (−14.4, 3.9)	g/ng/ml	−10.1 (−27.6, 7.5)
Meng, 2018[28]	1.00 (0.70–1.30 IQR ng/ml)	(0, 0.46)	3.1 (−24.9, 31.2)	g/log2(ng/ml)	4.5 (−35.9, 45)
Sagiv, 2018[29]c	2.40 (1.60–3.70 IQR ng/ml)	(0.88, 0.62)	−2.8 (−16.0, 10.5)	g/IQR (ng/ml)	−3.3 (−18.7, 12.3)
Shi, 2017[54]	0.23 (0.28 SD ng/ml)	(−1.93, 0.96)	108.8 (−53.8, 271.5)	g/log10(ng/ml)	47.3 (−23.4, 117.9)
Shoaff, 2018[55]b	1.50 (1.00–2.00 33–67 PCTLR ng/ml)	(0.41, 0.79)	−13.4 (−35.9, 9.1)	g/ng/ml	−20.9 (−55.9, 14.1)
Starling, 2017[56]	0.80 (0.50–1.20 IQR ng/ml)	(−0.22, 0.65)	−13.5 (−50.7, 23.7)	g/ln(ng/ml)	−13.5 (−50.7, 23.7)
Valvi, 2017[57]	4.54 (2.24–8.52 IQR ng/ml)	(1.51, 0.99)	15.0 (−18.0, 47.0)	g/log2(ng/ml)	21.6 (−26, 67.8)
Wikström, 2020[30]	1.23 (0.86–1.99 IQR ng/ml)	(0.21, 0.62)	−0.1 (−38.0, 38.0)	g/ln(ng/ml)	−0.1 (−38.0, 38.0)
Workman, 2019[58]	0.73 (1.40 SD ng/ml)	(−1.09, 1.24)	−6.6 (−66.9, 53.7)	g/ln(ng/ml)	−6.6 (−66.9, 53.7)
Xu, 2019[59]	5.69 (4.19 SD ng/ml)	(1.52, 0.66)	−173.8 (−627.9, 280.1)	g/log10(ng/ml)	−75.5 (−272.7, 121.6)
Yao, 2021[60]	0.32 (0.26–0.39 IQR ng/ml)	(−1.14, 0.30)	−10.2 (−130.1, 109.7)	g/ln(ng/ml)	−10.2 (−130.1, 109.7)

Abbreviations: Q25: 25% percentile; Q50: 50% percentile (median); Q75; 75% percentile; BLOD: below limit of detection; SD: standard deviation; GM: geometric mean; GSD geometric standard deviation; SE: standard error; IQR: Interquartile range; 33–67 PCTLR: 33% percentile to 67 percentile range; µ: mean of log-normal distribution; σ: standard deviation of log-normal distribution: β: beta coefficient; LCI: lower 95% confidence interval; UCI: upper 95% confidence interval; IQR: inter-quartile range; NR: not reported.

^aTwo publications (Maisonet et al.[25] and Marks et al.[26] are from the same ALSPAC cohort.

^bShoaff et al[55], Maisonet et al.[25] and Marks et al.[26] presented results on the natural scale (i.e., per ng/mL), which were re-expressed into ln units (i.e., per ln (ng/mL)).

^cThe results from Bach et al.[44] and Sagiv et al.[29] were first mathematically rescaled to effect estimates per (natural scale) of exposure units and then re-expressed into ln units (i.e., per ln (ng/mL)).

^dThe results from Buck Louis et al.[27] were first rescaled to per ln-transformed of 1 plus exposure units and then re-expressed into ln units (i.e., per ln (ng/mL)).

Overall Study Population Results

The meta-analysis of all 27 studies yielded a decrease of −7.9 g (95%CI: −15.0, −0.7) in birth weight per each ln ng/mL increase in PFHxS (Figure 2; Supplemental Table 5). The between-study heterogeneity was negligible (p_Q= 0.85; I²=0%). No significant relationship was found between the observed effect sizes and their standard error (p_E= 0.44) via the Egger’s regression test, and there was no evidence of publication bias based on funnel plot symmetry (Figure 3).

Figure 2.

Forest plot of the 27 studies included in the meta-analysis on PFHxS exposure and changes in birth weight.

Symbols and abbreviations: N: study sample size; n: number of studies; CI: Confidence Interval; 1st: Trimester 1; 2nd: Trimester 2; 3rd: Trimester 3; RE: random effect model; Q: Cochran’s Q test statistics; df: degree of freedom; p: p-value; I²: Higgin’s and Thompson’s I²; Q_M: test statistics for subgroup differences.

*There are three sample timing strata; Early-pregnancy group: studies with biomarker samples taken either in the 1st trimester, a combination of 1st and 2nd trimester, or a combination of 1st, 2nd and 3rd trimester; Mid- to late-pregnancy group: measurements from the 2nd and 3rd trimester, a combination thereof, or a combination the 2nd, 3rd trimester and at delivery; Post-pregnancy group: studies with biomarker samples from post-birth or at delivery/birth.

Figure 3.

Enhanced Funnel Plot of 27 studies examining PFHxS exposures and birth weight differences.

Stratified meta-analysis – study confidence level

The pooled effect for the 12 high confidence studies (β= −6.8g; 95%CI: −16.3, 2.8; p_Q= 0.44; I²= 0%) was smaller than the 11 medium confidence studies (β= −10.0g; 95%CI: −21.1, 1.1; p_Q= 0.94; I² = 0%) with negligible between-study heterogeneity. These results were larger than the 4 low confidence studies which showed low heterogeneity (β= −1.5g; 95%CI: −51.6, 48.7; p_Q= 0.30; I²=20.1%) (Figure 2; Supplemental Table 5). There was no statistically significant difference (p= 0.85) in effect size among these three strata.

Stratum-specific estimates – biomarker sample timing

The mean BW deficit among 12 studies with early sampling periods based on our classification (e.g., pre-conception, Trimester 1-only, Trimester 1 and 2, and Trimester 1 to 3) (β= −7.6g; 95%CI: −16.2, 1.1) showed slightly larger deficits per each PFHxS ln-unit change than the 10 studies with mid- to late-pregnancy sampling (e.g., Trimester 2-only; Trimester 2 and 3; Trimester 3-only; Trimester 2, 3 and at delivery) (β= −3.9g; 95%CI: −17.7, 9.9) (Figure 2; Supplemental Table 5). The mean BW deficit for five post-partum serum or plasma sample studies (e.g., Birth and Post-birth) studies was even larger (β= −28.3g; 95%CI: −69.3, 12.7). The results showed that while some mean BW deficits differed in magnitude by sampling window, especially in contrast to early and post-partum strata, statistical significance (p= 0.30) was not observed when testing the hypothesis for differences across any strata.

Meta-Regression

The coefficient γ used to represent the change in the effect of PFHxS exposure on birth weight per each gestational week increase was estimated to evaluate how sampling time may affect BW deficits. The γ of −0.34g (95%CI: −1.08, 0.41) indicated that there is no statistically significant impact of biomarker sample timing on the pooled estimate of mean BW deficits.

Supplemental and sensitivity analyses

Random effect meta-analyses for different later sampled study groupings were based on two- or three-strata (Figure 2; Supplemental Table 5). The pooled effect estimates for PFHxS measured in mid- to late-pregnancy and postpartum (β= −8.5g; 95%CI: −21.0, 4.1) combined in the two-strata approach were comparable to the early pregnancy or pre-conception (β= −7.6g; 95%CI: −16.2, 1.1); these were vastly smaller than those seen for post-partum studies only. Among the 12 studies sampled during early pregnancy, 6[23, 24, 27–30] were re-categorized into very early pregnancy based on predominately first trimester or earlier sampling and/or low measures of centrality of sample timing (e.g. mean, median or mode of gestational age week at sampling ≤ 10) (Table 2). The mean BW deficit per ln exposure unit change in PFHxS for these six studies (β= −3.5g; 95 CI: −14.8, 7.9) was smaller compared to the remaining six early sampled studies (β= −13.4g; 95%CI: −26.9, 0.1) (Table 2). Among the 12 high confidence studies, 7 reported early biomarker sampling and 5 used mid- to late-pregnancy measures. The pooled effect for the mid-/late- pregnancy high confidence studies was slightly larger (β= −8.8g; 95%CI: −30.2, 12.7) than the early high confidence studies (β= −6.3g; 95%CI: −16.9, 4.4). Results from European studies were slightly larger (β= −9.8g; 95%CI: −20.7, 1.1) than both North American (β= −6.2g; 95%CI: −17.0, 4.5) and Western Pacific study populations (β= −7.0g; 95%CI: −26.2, 12.2) (Supplemental Table 5). Twenty-one studies reported their results that could be rescaled to per ln exposure units, and 22 reported their results based on any type of log transformation. The mean BW deficits for these 21 studies (β= −4.2g; 95%CI: −14.5, 6.2 per ln unit) and 22 studies (β= −6.0g; 95%CI: −15.8, 3.8) were slightly lower than our overall estimate (β=−7.9g;95%CI: −15.0, −0.7) (Figure 2).

Table 2.

Random effect estimates (β are in g/ln(ng/ml) of mean birth weight differences (and tests for heterogeneity) for PFHxS exposures for sensitivity analyses.

Set of studies	n	β (g per ln (ng/ml))	95% Confidence Interval	I2 (%)	pQ
All studies	27	−7.9	−15.0, −0.7	0	0.85
All studies plus 3 studies with categorical dataa	30	−13.4	−32.4, 5.7	89.3	<0.0001
Studiesb without re-expression of β	21	−4.2	−14.5, 6.2	0	0.79
Studiesc with log based β	22	−6.0	−15.8, 3.8	0	0.78
Early sampling with exclusive T1 or smaller measures of centrality	6	−3.5	−14.8, 7.9	0	0.84
Early sampling with larger measures of centrality	6	−13.4	−26.9, 0.1	0	0.86
Early sampling in High confidence	7	−6.3	−16.9, 4.4	0	0.85
Mid to Late sampling in High confidence	5	−8.8	−30.2, 12.7	0	0.71

Symbols and abbreviations: n = number of studies; β = combined estimate of change in birth weight (g) per ln (ng/mL) PFHxS exposure; I² = % variation in the pooled effect due to study heterogeneity; p_Q = p-value for Cochran’s Q test for heterogeneity; T1: trimester 1.

^aOriginal Categorical Study Results: Cao et al. [20]: Tertile 2: −13.7g; 95%CI: −136.9, 109.6). Tertile 3: −25.1 g (95%CI: −149.1, 98.9); Gao et al.[22]: Tertile 2: −154.1 g (95%CI: −332.2, 24.0); Tertile 3: −101.2 g (95%CI: −275.5, 73.1); Eick et al.[21]: Tertile 2: 82.2 g (95%CI: −24.8, 189.2); Tertile 3: 75.7 g (95%CI: −51.4, 202.8). [See Supplemental file for further discussion]

^bStudies reported results per ln(ng/mL), log₂(ng/mL) or log₁₀(ng/mL) that can be mathematically rescaled to per ln(ng/mL).

^cStudies reported results based on some types of log transformation, including Buck Louis et al.[27] which reported results per ln(1+ng/mL).

Discussion

The primary meta-analysis of all 27 studies yielded an estimated −7.9 g (95%CI: −15.0, −0.7) decrease in BW per each ln ng/mL increase in PFHxS. The large sample size allowed us to evaluate whether meta-effect estimates varied by study confidence and different study characteristics, such as sample timing. Although some differences were noted, we did not observe any clear gradients in BW deficits across all study confidence levels. For example, the pooled results based on only four low confidence studies were null (β= −1.5g; 95%CI: −51.6, 48.7) with high variability in the individual study results. The low confidence results were divergent from the other stratum-specific effect estimates which were comparable in magnitude and precision including the high (β= −6.8g; 95%CI: −16.3, 2.8), medium (β= −10.0g; 95%CI: −21.1, 1.1) and high and medium confidence studies combined (β= −8.1g; 95%CI: −15.4, −0.9). This demonstrates that the overall pooled effect estimates of this meta-analysis was predominately influenced by these high and medium confidence study subsets of which we would expect a low risk of bias.

Although a gradient across each sampling window was also not evident, our data show a distinct pattern of larger BW deficits among studies measuring PFHxS in post-partum samples. For example, one of the largest stratum-specific differences detected was between the studies with post-partum (β= −28.3g; 95%CI: −69.3, 12.7) and early (β= −7.6g; 95%CI: −16.2, 1.1) sampling. The BW decrease among the 12 early sampled studies was comparable in magnitude to the overall estimate across 27 studies. However, our analysis examining six studies with the earliest sampling times (predominately trimester 1) showed null results (β= −3.5g; 95%CI: −14.8, 7.9). The contrast between this result and that for five post-partum sampled studies raises concern for potential bias due to changing hemodynamics during pregnancy. As noted earlier, hemodynamic changes are due to increased blood plasma volume as a result of decreased mean arterial pressure, increased cardiac output, and systemic vasodilation[29] during pregnancy and may lead to bias in epidemiological studies especially among those studies based on samples collected late in pregnancy. The larger deficits for the combined late-/post-sampling group compared to early pregnancy is coherent with largely null BW findings for early sampling based on meta-analyses of both PFOA and PFOS.[14, 15] Those meta-analyses of early sampling period studies were based on a binary classification comparison to later samples (i.e., exclusive second trimester sampling and later including post-partum). Null results in early sampling groups in published meta-analyses on PFOS and PFOS are in contrast to a recent meta-analysis of PFNA exposures[4] which showed larger and more robust overall BW deficits that were evident across various sample timing categorization approaches (e.g. 3-stratum that separated out later pregnancy from post-partum sampling).

Of the five PFHxS studies reporting post-partum sampling, four were based on measurements taken from umbilical cord blood. While umbilical cord blood is only capable of providing fetal exposure information around the time of birth, it has the advantage of being a more direct measure of fetal exposure, compared to maternal blood. Maternal and cord measurements of PFHxS are typically highly correlated, with coefficients between 0.59 and 0.88 for most studies[31–36], except one study[37] which observed a correlation coefficient of 0.18. This supports the use of maternal measurements as a surrogate for fetal exposure. However, cord blood concentrations are typically lower than maternal blood, with median cord to maternal concentration ratios of between 0.32 and 0.73.[31–38] This has typically been attributed to the function of placental transporters and could also be influenced by differences in binding protein affinity or concentration between maternal and fetal plasma. As cord blood PFHxS concentration is typically correlated with but lower than maternal blood concentration, a quantitative estimate of the same effect per ng/ml increase in cord blood PFHxS would be greater than per ng/ml increase in maternal blood PFHxS, if both measures are equally representative of the true exposure. This agrees with the stronger relationship between BW and PFHxS in post-partum samples compared to mid/late-pregnancy samples in our analysis. Cord blood measurements could also be more sensitive as a more direct measure of fetal exposure, though effects on birth weight may be mediated by effects in the placenta, which receives blood flow from both the mother and fetus.[5]

A strength of our meta-analysis was the ability to better differentiate sampling timing subsets than previous research. This is evident in the observation of a difference in results for the earliest sampled studies in our three-strata analysis (as well as sensitivity analyses restricting to high confidence studies only), as we would have detected little difference for PFHxS across sample timing based solely on the two-strata approach. Further investigation in epidemiological studies and systematic reviews should be undertaken to identify if the large differences between mid- to late-pregnancy and postpartum samples are unique to PFHxS. One of the challenges with the discrete sampling time analysis is that studies within the same stratum were based on heterogeneous sampling approaches that spanned different time periods during or near pregnancy. For example, the early pregnancy strata included studies with sampling which ranged from trimesters 1 to 3, whereas studies that sampled in trimester 2 only are included in the mid-to-late pregnancy group. This may decrease our sensitivity to detect differences and adds some uncertainty introduced by the dependencies (overlapping) of sample timing strata that cannot be readily estimated or accounted for in the statistical analysis used here. Our analysis of the impact of sampling timing on BW deficits was conducted using a meta-regression with the continuous gestational age as a measure of sample timing and potential effect modifier. Although the inverse beta showed some limited evidence of increasing deficit with each week of sample timing, this result (γ= −0.34g; 95%CI: −1.08, 0.41) was not statistically significant. These results seem to be influenced by the five post-partum studies, as removal of these studies showed an even more null result (γ= −0.07g; 95%CI: −0.97, 0.84). There is also uncertainty when evaluating a continuous estimate of gestational age, which depends on the level of detail provided by studies. First, measurement error in gestational age estimates can vary across dating techniques. Second, gestational age estimates at sample timing were derived from a combination of measures of central tendency (median, mean, midpoint, etc.), assuming that all measures of centrality are interchangeable. In addition, uncertainty related to measurement error associated with underlying gestational age estimates may be increased by the assumption of a linear relationship between sample timing and BW deficits. The uncertainty introduced by this type of measurement error is mitigated by our use of discrete sampling timing groupings. Given the sources of uncertainty related to measurement and reporting of gestational age distributions, further delineation of the impact of sampling timing differences may require more refined data or more deliberate sampling strategies employed in the underlying epidemiological studies.

Our primary meta-analysis was performed estimating change in mean BW per ln (ng/mL) exposure units, since the majority of the studies reported results on the log scale of exposure units. The studies that reported mean BW per log₂ (ng/mL) or log₁₀ (ng/mL) exposure units were mathematically rescaled to mean BW per ln (ng/mL) exposure units. However, 6 of 27 studies that reported results in the overall population or across both sexes were re-expressed to results per ln (ng/mL) exposure units using a method of re-expression such as used by Steenland et al.[14] and Dzierlenga et al.[15]. According to Linakis et al.[19], this method of effect estimate re-expression can result in substantial bias depending on the underlying distribution of raw data. Sensitivity analysis using 21 studies without any re-expression yielded a −4.2 g (95%CI: −14.5, 6.2) decrease in mean BW per ln (ng/mL) exposure unit, which is slightly lower in magnitude and less precise compared to the overall estimate of −7.9 g (95%CI: −15.0, −0.7) from 27 studies. Although these results are slightly different, the advantage of more inclusive analysis allowed for a larger sample size and more extensive analyses of heterogeneity and may outweigh the additional uncertainty this re-expression may create.

A study strength was the inclusion of limited to 30 informative studies in our systematic review process. Our primary analysis of 27 studies with continuous PFHxS data showed a small but statistically significant BW decrease (β= −7.9g; 95%CI: −15.0, −0.7) per each ln-unit increase. This is larger in magnitude than the estimate from the 11 maternal PFHxS studies included in Lan et al.[39] especially when their log₁₀ results (β= −5.7g; 95%CI: −33.9, 22.6) are rescaled on a ln-unit scale (β= −2.5g; 95%CI: −14.7, 9.8). This in contrast to what Gui et al.[40] reported in their analysis of 15 BW studies (out of 23 identified publications). Although they reported a deficit based on categorical exposure studies alone (−31 g in highest quantile), their estimate per each ln-unit increase (β= −0.47g; 95%CI: −1.20, 0.25 per each ln-unit) was much lower. However, their analyses appear to be based on a questionable combination of both (unitless) standardized and absolute BW measures (measured in grams) and precludes direct comparison with our analysis based on studies of mean BW differences only. Their weighting approach is also curious as one study of 398 participants[55] accounts for nearly all of the weight (98.1%) of all the combined studies. Following conversion of 3 additional studies reporting only categorical PFHxS measurement data, our analysis of 30 studies showed a larger BW difference (β=−13.4g; 95%CI: −32.4, 5.7) compared to our primary analysis of 27 studies (β= −7.9g; 95%CI: −15.0, −0.7). However, we determined that the method for estimation of a continuous PFHxS measurement from categorical observations was not reliable in this case. The continuous estimates we calculated from those studies were large and precise compared to the original categorical estimates and the values from other studies. Furthermore, inclusion of those studies resulted in high heterogeneity statistics compared to the analysis without those studies as well as every other stratum analyzed here. These together led us to conclude that the continuous estimates were likely substantially larger and more precise than the true values one would obtain from a continuous analysis of the data. The unreliability of the conversion method is potentially due to the small range of PFHxS exposure in the study populations, which resulted in the conversion extrapolating beyond the exposure contrast originally observed. More work is needed to evaluate this method for conversion and the conditions under which it produces acceptable continuous estimates.

Our systematic review represents the most extensive literature search, re-expression effort, and analyses conducted to date on the effects of PFHxS on birth weight. While the number of studies for the primary analysis may be large enough (n=27) to not be a major concern, some imprecision and uncertainty exist for the stratified analyses with considerably fewer studies per strata. This included some stratified analyses such as the low confidence, post-partum stratum, early/late among high confidence studies and the portioning of the early studies into two groups. The smaller numbers of studies (range: 4 to 7) in some subgroups impact the ability to discern statistically significant differences in effect estimates between groups and the examination of heterogeneity. Thus, more research with larger sample sizes, more standardized reporting of biomarker distributions and even earlier gestational age sampling windows may further clarify some of the uncertainties identified across the PFAS literature, including the potential impact of pregnancy hemodynamics. Although previous studies have not demonstrated that hemodynamic measures such as GFR and albumin[29, 41–43] were empirical confounders, uncertainty remains as to whether these measures account for the complex and dynamic effect of physiological changes over pregnancy. Future efforts that include meta-regression analyses may further help clarify the extent to which sample timing and hemodynamics influence associations of PFAS and birthweight, and if that varies across PFAS examined.

Conclusion

Overall, we detected a small but statistically significant BW deficit (β= −7.9 g per each ln-unit PFHxS increase) across 27 studies including some studies where results were re-expressed from natural scale units to ln-unit and combined across sexes. Although we saw comparable magnitudes of BW deficits in both medium and high confidence studies as well as the early pregnancy group compared to mid- and late-pregnancy, when we further stratified the early samples to six studies with the earliest biomarker centrality estimates, the magnitude of the BW differences were closer to the null value. Future efforts would benefit from increased standardization of data reporting and better gestational age estimates of biomarker sampling.

Key message:

What is already known on this topic:

Limited systematic efforts exist for characterizing the relationship between PFHxS exposure and mean birth weight.

What this study adds:

We found a statistically significant relationship between PFHxS exposure and mean birth weight but saw important distinctions across different subgroups that may be partly explained by pregnancy hemodynamics.

How this study might affect research, practice or policy:

There is an increasing need to determine safe levels in environmental media to address clean-up of environmental media such as contaminated soil and water; these data can directly inform risk assessment efforts for developmental effects related to PFHxS exposures.