Next Article in Journal
Comparison of the Zebrafish Embryo Toxicity Assay and the General and Behavioral Embryo Toxicity Assay as New Approach Methods for Chemical Screening
Next Article in Special Issue
Health Effects Associated with Exposures to Endocrine Disrupting Chemicals
Previous Article in Journal
Bird Feces as Indicators of Metal Pollution: Pitfalls and Solutions
Previous Article in Special Issue
Phthalates and Bisphenol A: Presence in Blood Serum and Follicular Fluid of Italian Women Undergoing Assisted Reproduction Techniques
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Toxics
Volume 8
Issue 4
10.3390/toxics8040125
toxics-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

The Confounder-Mediator Dilemma: Should We Control for Obesity to Estimate the Effect of Perfluoroalkyl Substances on Health Outcomes?

by
Kosuke Inoue
1,*,
Atsushi Goto
2,
Takehiro Sugiyama
3,4,5,
Cecilia Høst Ramlau-Hansen
6 and
Zeyan Liew
7,8
1
Department of Epidemiology, Fielding School of Public Health, University of California, Los Angeles, CA 90095, USA
2
Department of Health Data Science, Graduate School of Data Science, Yokohama City University, Yokohama 236-0027, Japan
3
Diabetes and Metabolism Information Center, National Center for Global Health and Medicine, Research Institute, Tokyo 162-8655, Japan
4
National Center for Global Health and Medicine, Institute for Global Health Policy Research, Bureau of International Health Cooperation, Tokyo 162-8655, Japan
5
Department of Health Services Research, Faculty of Medicine, University of Tsukuba, Ibaraki 305-8575, Japan
6
Research Unit for Epidemiology, Department of Public Health, Aarhus University, 8000 Aarhus, Denmark
7
Department of Environmental Health Sciences, Yale School of Public Health, New Haven, CT 06510, USA
8
Yale Center for Perinatal, Pediatric, and Environmental Epidemiology, Yale School of Public Health, New Haven, CT 06510, USA
*
Author to whom correspondence should be addressed.
Toxics 2020, 8(4), 125; https://doi.org/10.3390/toxics8040125
Submission received: 13 November 2020 / Revised: 14 December 2020 / Accepted: 17 December 2020 / Published: 20 December 2020
(This article belongs to the Special Issue Health Effects Associated with Exposures to Endocrine Disrupting Chemicals)
Browse Figures
Versions Notes

Abstract

:
Confounding adjustment is important for observational studies to derive valid effect estimates for inference. Despite the theoretical advancement of confounding selection procedure, it is often challenging to distinguish between confounders and mediators due to the lack of information about the time-ordering and latency of each variable in the data. This is also the case for the studies of perfluoroalkyl substances (PFAS), a group of synthetic chemicals used in industry and consumer products that are persistent and have endocrine-disrupting properties on health outcomes. In this article, we used directed acyclic graphs to describe potential biases introduced by adjusting for or stratifying by the measure of obesity as an intermediate variable in PFAS exposure analyses. We compared results with or without adjusting for body mass index in two cross-sectional data analyses: (1) PFAS levels and maternal thyroid function during early pregnancy using the Danish National Birth Cohort and (2) PFAS levels and cardiovascular disease in adults using the National Health and Nutrition Examination Survey. In these examples, we showed that the potential heterogeneity observed in stratified analyses by overweight or obese status needs to be interpreted cautiously considering collider stratification bias. This article highlights the complexity of seemingly simple adjustment or stratification analyses, and the need for careful consideration of the confounding and/or mediating role of obesity in PFAS studies.

1. Introduction

Observational studies have played an important role in making inferences about the impact of environmental exposure on health because experimental studies are often unfeasible or unethical to perform. Confounding bias is a major concern that may threaten the validity of observational studies [1]. To reduce confounding, one might think that controlling for as many variables as possible in the model is preferable, but such an approach sometimes introduces bias and reduces statistical efficiency [1,2]. For instance, the total effect of the exposure on health outcomes may be underestimated if an intermediate variable that lies within the causal pathway between the exposure and outcome is included in the statistical model, and, therefore, the estimated (total) effect would fail to capture the specific effect mediated through that variable. In addition, if there are unmeasured confounders between the intermediate variable and the outcome, adjusting for such an intermediate variable could introduce “collider bias” [3]. Some well-known examples of such bias include the “birth weight paradox” [4], where maternal smoking became protective on infant mortality risk conditioning on low birth weight status; or the “obesity paradox” [5] where favorable health risks are observed among obese individuals, conditional on surviving the study follow-up period. These examples have emphasized the potential harmfulness of conditioning or stratifying on intermediate variables affected by the exposure in statistical analyses. However, it is often challenging to determine whether variables are confounders or mediators when the time ordering of the study variables in a dataset is unclear.
Perfluoroalkyl substances (PFAS), a group of man-made fluorine-containing chemicals, has received attention over the past decade, and potential adverse health effect of human exposure to these extremely widespread and persistent compounds have been studied [6]. Although the production of some PFAS, such as perfluorooctane sulfonate (PFOS) and perfluorooctanoic acid (PFOA), has been reduced in Europe and the U.S. since 2000 [7], these compounds are still widely detected [8,9], and several other types of PFAS are reported to be increasing in use [8]. Research has demonstrated that several types of PFAS have endocrine-disturbing properties and can affect steroids and thyroid function [10] as well as cholesterols levels [11,12], leading to cardiovascular disease (CVD) [13]. These previous studies have adjusted for obesity, which is one of the metabolic disorders that has been associated with PFAS in adulthood during reproductive age and beyond [14,15,16]. However, nearly all of the studies were performed using cross-sectional data [10], and therefore, a possible bidirectional relationship between PFAS and obesity makes it difficult to determine whether obesity should be treated as a mediator or a confounder in the statistical analysis, especially if obesity is measured at the same time as PFAS assessment. Moreover, the biological half-life for common PFAS compounds is in the magnitude of years. One measurement of both PFAS and obesity might represent cumulative or chronic status years before data collection.
The goal of this study is to use PFAS research as examples to illustrate how obesity can act as either a mediator or a confounder in studies of environmental exposure and health. We used causal diagrams or directed acyclic graphs (DAGs) to explain the confounding mediator dilemma of defining the role of obesity in PFAS studies. Applying DAGs is useful to guide analysis that aims to estimate the (total) causal effect of the exposure on the outcome, by identifying a sufficient set of variables that can close all confounding paths while avoiding controlling for any mediators and colliders [17,18,19]. As an illustration, we compared results with or without controlling for body-mass-index (BMI) in two cross-sectional studies that estimated the associations of PFAS levels with maternal pregnancy thyroid function [20] and with prevalent CVD in adults [13].

2. Confounder-Mediator Dilemma

In general, confounders are variables that are common causes of the exposure and the outcome (i.e., the variable with arrows pointing into both the exposure and the outcome in DAG). More strictly speaking, confounders are variables that are causes of the exposure and related to the outcome, or variables that are causes of the outcome and related to (but not affected by) the exposure [19]. Mediators are variables that lie on the causal pathway from the exposure to the outcome (i.e., the variable with an arrow pointing from the exposure and an arrow pointing into the outcome). Figure 1 shows a DAG that encodes the potential causal relationship between PFAS levels and health outcomes. In this DAG, obesity “before” the PFAS measurement (Obesitypre) is a confounder for the causal relationship between PFAS and health outcomes, while obesity “after” the PFAS measurement (Obesitypost) is a mediator on the causal path from PFAS to health outcomes.
To estimate the total effect of the exposure on the outcome, only the confounders, but not the mediators, should be included in statistical analysis. Controlling for intermediate variables can introduce biased estimate mainly for two reasons (Figure 2): (i) failure to include the effect of the exposure on the outcome that is mediated through the intermediate variables, and (ii) introduction of a collider bias, which is the non-existent relationship between the exposure and the outcome through the mediator-outcome confounders created by controlling for the intermediate variables [3,21].
Most epidemiological studies of PFAS at present have information on obesity measured only once at baseline and cannot distinguish between a confounder (obesitypre) and a mediator (obesitypost) due to unknown temporality. Even when we have longitudinal data, it is often challenging to define whether obesity measurement represents the status prior to or post PFAS exposure, because both obesity and PFAS levels represent cumulative or chronic status years before the measurements. In such scenarios, it is important to carefully analyze the data and interpret the results with and without controlling for obesity or BMI (as a proxy measurement of obesity). In the following sections, we illustrate two PFAS studies (of which individual-level data were available or accessible to the authors) using BMI as an example of a variable that could be both a confounder and a mediator between PFAS and health outcomes. In these examples, we focused on PFOS and PFOA, the two most commonly used PFAS, for our exposure of interest, and discussed the results in the following discussion section.

3. Example Illustration I: The Association between PFAS and Thyroid Hormones during Pregnancy in the Danish National Birth Cohort

One of the main features of PFAS is its endocrine-disrupting properties, particularly affecting thyroid function [22,23,24]. Whether PFAS exposure during pregnancy affects maternal thyroid function—a critical endocrine function for the fetus neurodevelopment [25,26]—has been actively debated in environmental epidemiology and endocrinology [10,20,27]. A summary of findings about PFAS exposure and maternal thyroid hormones in pregnancy can be found in recent systematic review articles [10,28]. Using a sample of 1,366 pregnancies from the DNBC, a national pregnancy cohort study in Denmark established in 1996–2002 [29], we recently reported possible gestational-week-specific associations between plasma PFAS levels (continuous [per interquartile range increase] or categorical [quartiles with the lowest quartile as a reference] variables) and maternal TSH in early pregnancy [20].
In this work, the results were adjusted for pre-pregnancy BMI, assuming that BMI may affect plasma PFAS levels and maternal thyroid hormones (i.e., obesitypre in Figure 1). This assumption was based on the prior literature of prenatal PFAS exposure research. However, if PFAS measures in early pregnancy represent years of cumulative exposure, pre-pregnancy BMI can potentially be a mediator in the reported PFAS–thyroid association [30,31]. Here, we reevaluated our main findings of thyroid-stimulating hormone [TSH]—a key marker reflecting thyroid function during pregnancy [32], and compared the PFAS-TSH association with or without controlling for pre-pregnancy BMI in the regression model. Moreover, we also performed stratification by maternal overweight status (pre-pregnancy BMI ≥ 25 kg/m2). Details of the study design and data collection methods have been described in our previous article [20].
A total of 1366 pregnant women were included in the study. The median (interquartile range) value of PFOS and PFOA measured in maternal plasma collected during GW5 to GW19 (median [interquartile range] = 8 [7,8,9,10]) were 29.5 (22.6–37.7) ng/mL and 4.52 (3.38–5.80) ng/mL, respectively. In multivariable regression analyses, the effect estimates were close to null between prenatal PFOS or PFOA and TSH levels before and after adjusting for pre-pregnancy BMI (Table 1). In an analysis stratified by maternal overweight status, a tendency for positive associations between PFAS, especially for PFOA, and TSH among non-overweight women were observed. However, for overweight women, the effect estimates between PFOS and PFOA levels and TSH went in the negative direction, and for specific exposure quartiles, some estimates excluded the null (Table 2).

4. Example Illustration II: The Association between PFAS and Cardiovascular Disease Using the US National Health and Nutrition Examination Survey

A number of studies have suggested that higher PFAS exposure levels are associated with glucose homeostasis and cholesterol levels [11,12,33], which are strong risk factors for CVD. Moreover, a recent cross-sectional study using the NHANES 1999–2000 and 2003–2014 waves (PFAS was not measured in the 2001–2002 wave) showed that the serum PFAS levels were associated with the prevalence of CVD (the self-reported physician diagnosis) in a representative sample of U.S. adults age ≥ 20 years [13]. Previous findings of PFAS exposure and CVD have been summarized in a recent review paper [28]. In the NHANES report [13], BMI of the participants at the time of the survey was included as a confounder (i.e., obesitypre in Figure 1), and thus adjusted for in all regression models. However, if PFAS exposure increases the risk of obesity, BMI could act as a mediator (i.e., obesitypost in Figure 1) instead. Whether BMI here represents obesitypre or obesitypost in Figure 1 cannot be determined using this cross-sectional data.
Using the publicly assessable data, we re-evaluated the associations between PFAS (continuous [per interquartile range increase] or categorical [quartiles with the lowest quartile as a reference] variables) and prevalence of CVD in NHANES 1999–2000 and 2003–2014 with or without adjusting on participant’s BMI at the time of the study. Moreover, in this U.S. adult sample, we performed stratified analyses classifying participants into obese (current BMI ≥ 30 kg/m2) or non-obese (current BMI < 30 kg/m2). We used a different cut-off of BMI (i.e., 30 instead of 25 kg/m2) between the DNBC cohort and the NHANES cohort because the former included pregnant women in Denmark with a low prevalence of obesity while the latter included all adults aged ≥ 20 years in the US with a high prevalence of obesity [34]. Details of the study sample selection, the measurements of each variable, and the statistical approach are described in Text S1.
Among 7411 US adults included in this complete-case analysis, 571 participants (7.7%) had CVD. The median (interquartile range) values of PFOS and PFOA were 14.1 (7.9–24.4) ng/mL and 3.5 (2.2–5.2) ng/mL, respectively. Similar to the findings reported previously, we found that higher PFOS and PFOA levels were positively associated with the prevalence of CVD and the effect estimates did not change in the model without (Model 1) or with (Model 2) adjustment of current BMI (Table 3). In stratified analyses, a stronger positive association between PFOS or PFOA levels and CVD prevalence was observed for obese adults while the association was null for non-obese participants (Table 4).

5. Discussion

This study highlights the complex role of obesity measurement in PFAS exposure and health outcome research. Over-adjustment problems caused by conditioning on intermediate variables require attention in toxicological and environmental epidemiology given that many factors (e.g., medical conditions, biomarkers, and socioeconomic status, etc) are candidates to be potential mediators in the exposure-outcome relationship. Particularly, one measure of obesity can simultaneously act as a confounder and a mediator as well as a “collider” intersecting multiple causal and non-causal pathways, linking PFAS exposure to a health outcome. A naïve adjustment of obesity measurement in a regression model may remove confounding but also poses a risk of blocking the mediating effect of interests. Moreover, conditioning on an intermediate variable between the exposure and the outcome can also risk an induction of collider bias [3,21], which is often being amplified in stratified analyses as seen in the birth weight paradox [4,35] and obesity paradox [5]. We suggest careful consideration of the complex role of obesity in analyses of PFAS exposure on a health outcome, thus results from models adjusting or stratifying on obesity measurement need to be interpreted with caution. Having longitudinal measures of BMI and PFAS levels might help to clarify the time–orders relationship between the exposure and the covariate and appropriate statistical modeling strategies that need to be employed [36,37,38].
Animal studies have suggested that the underlying mechanisms of the PFAS–obesity association might include the activation of the peroxisome proliferator-activated receptors and other transcriptional factors [39,40], the alteration of lipid metabolism [41], and the alteration of energy metabolism and thyroid hormone homeostasis [22,23,42]. These mechanisms indicate that higher PFAS exposure may increase the risk of obesity and possibly other metabolic disorders and CVD. Meanwhile, obesity can also share several common causes with PFAS levels such as dietary and lifestyle factors, thus obesity has been often used as a proxy variable to address lifestyle confounding factors that can be unmeasured or unknown [43,44].
In our illustration examples using the DNBC (PFAS-thyroid association) and the NHANES (PFAS-CVD association), we did not find a substantial change in estimates after adjusting for BMI. In both studies, BMI might not be a strong confounder in the exposure-outcome associations. Particularly for our second example using the NHANES, several variables related to metabolic disorders highly correlated with BMI were already included in the multivariable model. Similarly, BMI might not be a strong mediator in both studies given the very slight changes in effect estimate after adjusting for BMI. However, the potential heterogeneity observed in stratified analyses by overweight or obese status could still be due to collider stratification bias, if there is the presence of other uncontrolled common causes of BMI and the health outcome (i.e., an open biasing path tracing PFAS → [BMI] ← other controlled risk factors → health outcomes in DAG). The direction and magnitude of such bias within strata of a collider can vary and are hard to predict [45,46]. In the NHANES data, measures of PFAS and BMI both can be after the participants had CVD, thus a more direct collider path is possible (i.e., an open biasing path tracing PFAS → [BMI] ← health outcomes in DAG). Pre-pregnancy BMI in the DNBC indicated weight at pre-pregnancy period before PFAS was measured in pregnancy, but PFASs are very persistent, thus one serum measure could represent cumulative exposure years prior to pregnancy. These examples highlight the complexity of seemingly simple adjustment or stratification analyses routinely performed in epidemiological analyses of PFAS.
In this article, we re-evaluated the main findings reported from two recently published articles on PFAS exposure. Our results were meant for illustration and do not reflect the full scope of considerations needed to study the potential causal effect of PFAS on pregnancy thyroid hormones or CVD risk in adults. Our illustrations were based on the two most common types of PFAS and we focused on explaining the potentially complex role of one covariate (i.e., BMI as a measurement of obesity) in analyses. Obesity or overweight measurement was based on self-reported BMI, and stratification analysis was done using a binary classification for simplicity purpose and considering the distribution of BMI in these datasets. Although, how obesity should be considered in causal analyses has been debated in epidemiology [47,48,49], we simply focused on the role of obesity measure as a confounder or a mediator in PFAS studies.
In summary, careful consideration of the confounding and/or the mediating role of obesity is critical for researchers not to misinterpret the data in PFAS studies, particularly when stratifying the study sample by individuals’ obese or overweight status. If the variable is highly likely to be a mediator of the causal pathway of interest, we recommend to refrain from conducting and interpreting the stratified analysis by obesity or overweight, because the stratum-specific effect estimates could be biased. Measuring and adjusting for common causes of PFAS and obesity such as sociodemographic, dietary and lifestyle factors that are less likely to be in the mediating pathways would be recommended to address confounding between PFAS exposure and health outcomes.

Supplementary Materials

The following are available online at https://www.mdpi.com/2305-6304/8/4/125/s1, Text S1: Methods in “Example Illustration II” using the NHANES.

Author Contributions

Conceptualization, K.I., A.G., T.S., Z.L.; methodology, K.I., Z.L.; formal analysis, K.I., Z.L.; investigation, K.I., A.G., T.S., C.H.R.-H., Z.L.; writing—original draft preparation, K.I.; writing—review and editing, K.I., A.G., T.S., C.H.R.-H., Z.L.; visualization, K.I., C.H.R.-H., Z.L.; supervision, K.I. All authors have read and agreed to the published version of the manuscript.

Funding

KI was supported by National Institutes of Health (NIH)/NIDDK grant F99DK126119, Toffler award at UCLA, and Honjo International Foundation Scholarship. ZL was supported by the NIH/NIEHS Pathway to Independence Award (K99ES026729/R00ES026729). This article does not necessarily represent the views and policies of the NIH. The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Acknowledgments

The Danish National Birth Cohort was established with a significant grant from the Danish National Research Foundation. Additional support was obtained from the Danish Regional Committees, the Pharmacy Foundation, the Egmont Foundation, the March of Dimes Birth Defects Foundation, the Health Foundation and other minor grants. The DNBC Biobank has been supported by the Novo Nordisk Foundation and the Lundbeck Foundation. Follow-up of mothers and children have been supported by the Danish Medical Research Council (SSVF 0646, 271-08-0839/06-066023, O602-01042B, 0602-02738B), the Lundbeck Foundation (195/04, R100-A9193), The Innovation Fund Denmark 0603-00294B (09-067124), the Nordea Foundation (02-2013-2014), Aarhus Ideas (AU R9-A959-13-S804), University of Copenhagen Strategic Grant (IFSV 2012), and the Danish Council for Independent Research (DFF-4183-00594 and DFF-4183-00152). The specific grant numbers relevant for this paper can be found at https://www.dnbc.dk/access-to-dnbc-data/acknowledgement.

Conflicts of Interest

All authors state that they have no conflict of interest.

References

  1. Rothman, K.; Grenland, S.; Lash, T.L. Modern Epidemiology; Wolters Kluwer Health/Lippincott Williams & Wilkins: Philadelphia, PA, USA, 2008. [Google Scholar]
  2. Schisterman, E.F.; Cole, S.R.; Platt, R.W. Overadjustment Bias and Unnecessary Adjustment in Epidemiologic Studies. Epidemiology 2009, 20, 488–495. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  3. Cole, S.R.; Platt, R.W.; Schisterman, E.F.; Chu, H.; Westreich, D.; Richardson, D.; Poole, C. Illustrating bias due to conditioning on a collider. Int. J. Epidemiol. 2010, 39, 417–420. [Google Scholar] [CrossRef] [PubMed]
  4. Hernández-Díaz, S.; Schisterman, E.F.; Hernán, M.A. The Birth Weight “Paradox” Uncovered? Am. J. Epidemiol. 2006, 164, 1115–1120. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  5. Preston, S.H.; Stokes, A. Obesity Paradox. Epidemiology 2014, 25, 454–461. [Google Scholar] [CrossRef] [PubMed]
  6. US EPA. Per- and Polyfluoroalkyl Substances (PFAS). Available online: https://www.epa.gov/pfas (accessed on 24 July 2020).
  7. WHO State of the Science of Endocrine Disrupting Chemicals—2012. Available online: http://apps.who.int/iris/bitstream/10665/78101/1/9789241505031_eng.pdf (accessed on 25 April 2018).
  8. Bjerregaard-Olesen, C.; Bach, C.C.; Long, M.; Ghisari, M.; Bossi, R.; Bech, B.H.; Nohr, E.A.; Henriksen, T.B.; Olsen, J.; Bonefeld-Jørgensen, E.C. Time trends of perfluorinated alkyl acids in serum from Danish pregnant women 2008–2013. Environ. Int. 2016, 91, 14–21. [Google Scholar] [CrossRef] [PubMed]
  9. Kato, K.; Wong, L.-Y.; Jia, L.T.; Kuklenyik, Z.; Calafat, A.M. Trends in exposure to polyfluoroalkyl chemicals in the U.S. Population: 1999–2008. Environ. Sci. Technol. 2011, 45, 8037–8045. [Google Scholar] [CrossRef] [PubMed]
  10. Ballesteros, V.; Costa, O.; Iñiguez, C.; Fletcher, T.; Ballester, F.; Lopez-Espinosa, M.-J. Exposure to perfluoroalkyl substances and thyroid function in pregnant women and children: A systematic review of epidemiologic studies. Environ. Int. 2017, 99, 15–28. [Google Scholar] [CrossRef] [Green Version]
  11. Nelson, J.W.; Hatch, E.E.; Webster, T.F. Exposure to polyfluoroalkyl chemicals and cholesterol, body weight, and insulin resistance in the general U.S. population. Environ. Health Perspect. 2010, 118, 197–202. [Google Scholar] [CrossRef] [Green Version]
  12. Geiger, S.D.; Xiao, J.; Ducatman, A.; Frisbee, S.; Innes, K.; Shankar, A. The association between PFOA, PFOS and serum lipid levels in adolescents. Chemosphere 2014, 98, 78–83. [Google Scholar] [CrossRef]
  13. Huang, M.; Jiao, J.; Zhuang, P.; Chen, X.; Wang, J.; Zhang, Y. Serum polyfluoroalkyl chemicals are associated with risk of cardiovascular diseases in national U.S. population. Environ. Int. 2018, 119, 37–46. [Google Scholar] [CrossRef]
  14. Cardenas, A.; Hauser, R.; Gold, D.R.; Kleinman, K.P.; Hivert, M.-F.; Fleisch, A.F.; Lin, P.-I.D.; Calafat, A.M.; Webster, T.F.; Horton, E.S.; et al. Association of Perfluoroalkyl and Polyfluoroalkyl Substances with Adiposity. JAMA Netw. Open 2018, 1, e181493. [Google Scholar] [CrossRef] [PubMed]
  15. Braun, J.M. Early-life exposure to EDCs: Role in childhood obesity and neurodevelopment. Nat. Rev. Endocrinol. 2017, 13, 161–173. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  16. Andersen, C.S.; Fei, C.; Gamborg, M.; Nohr, E.A.; Sørensen, T.I.A.; Olsen, J. Prenatal exposures to perfluorinated chemicals and anthropometry at 7 years of age. Am. J. Epidemiol. 2013, 178, 921–927. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  17. Pearl, J. Causal Diagrams for Empirical Research. Biometrika 1995, 82, 669–688. [Google Scholar] [CrossRef]
  18. Greenland, S.; Pearl, J.; Robins, J.M. Causal Diagrams for Epidemiologic Research. Epidemiology 1999, 10, 37–48. [Google Scholar] [CrossRef]
  19. VanderWeele, T.J. Principles of confounder selection. Eur. J. Epidemiol. 2019, 34, 211–219. [Google Scholar] [CrossRef] [Green Version]
  20. Inoue, K.; Ritz, B.; Andersen, S.L.; Ramlau-Hansen, C.H.; Høyer, B.B.; Bech, B.H.; Henriksen, T.B.; Bonefeld-Jørgensen, E.C.; Olsen, J.; Liew, Z. Perfluoroalkyl Substances and Maternal Thyroid Hormones in Early Pregnancy; Findings in the Danish National Birth Cohort. Environ. Health Perspect. 2019, 127, 117002. [Google Scholar] [CrossRef]
  21. Greenland, S. Quantifying biases in causal models: Classical confounding vs collider-stratification bias. Epidemiology 2003, 14, 300–306. [Google Scholar] [CrossRef]
  22. Weiss, J.M.; Andersson, P.L.; Lamoree, M.H.; Leonards, P.E.G.; van Leeuwen, S.P.J.; Hamers, T. Competitive binding of poly- and perfluorinated compounds to the thyroid hormone transport protein transthyretin. Toxicol. Sci. 2009, 109, 206–216. [Google Scholar] [CrossRef] [Green Version]
  23. Yu, W.-G.; Liu, W.; Jin, Y.-H. Effects of perfluorooctane sulfonate on rat thyroid hormone biosynthesis and metabolism. Environ. Toxicol. Chem. 2009, 28, 990–996. [Google Scholar] [CrossRef]
  24. Long, M.; Ghisari, M.; Bonefeld-Jørgensen, E.C. Effects of perfluoroalkyl acids on the function of the thyroid hormone and the aryl hydrocarbon receptor. Environ. Sci. Pollut. Res. Int. 2013, 20, 8045–8056. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  25. Burrow, G.N.; Fisher, D.A.; Larsen, P.R. Maternal and fetal thyroid function. N. Engl. J. Med. 1994, 331, 1072–1078. [Google Scholar] [PubMed] [Green Version]
  26. Korevaar, T.I.M.; Medici, M.; Visser, T.J.; Peeters, R.P. Thyroid disease in pregnancy: New insights in diagnosis and clinical management. Nat. Rev. Endocrinol. 2017, 13, 610–622. [Google Scholar] [CrossRef] [PubMed]
  27. Preston, E.V.; Webster, T.F.; Oken, E.; Claus Henn, B.; McClean, M.D.; Rifas-Shiman, S.L.; Pearce, E.N.; Braverman, L.E.; Calafat, A.M.; Ye, X.; et al. Maternal Plasma per- and Polyfluoroalkyl Substance Concentrations in Early Pregnancy and Maternal and Neonatal Thyroid Function in a Prospective Birth Cohort: Project Viva (USA). Environ. Health Perspect. 2018, 126, 027013. [Google Scholar] [CrossRef] [PubMed]
  28. Kirk, M.; Smurthwaite, K.; Bräunig, J.; Trevenar, S.; D’Este, C.; Lucas, R.; Lal, A.; Korda, R.; Clements, A.; Mueller, J.; et al. The PFAS Health Study: Systematic Literature Review; The Australian National University: Canberra, ACT, Australia, 2018. [Google Scholar]
  29. Olsen, J.; Melbye, M.; Olsen, S.F.; Sørensen, T.I.; Aaby, P.; Andersen, A.M.; Taxbøl, D.; Hansen, K.D.; Juhl, M.; Schow, T.B.; et al. The Danish National Birth Cohort—Its background, structure and aim. Scand. J. Public Health 2001, 29, 300–307. [Google Scholar] [CrossRef]
  30. Houde, M.; Martin, J.W.; Letcher, R.J.; Solomon, K.R.; Muir, D.C.G. Biological monitoring of polyfluoroalkyl substances: A review. Environ. Sci. Technol. 2006, 40, 3463–3473. [Google Scholar] [CrossRef]
  31. Olsen, G.W.; Burris, J.M.; Ehresman, D.J.; Froehlich, J.W.; Seacat, A.M.; Butenhoff, J.L.; Zobel, L.R. Half-life of serum elimination of perfluorooctanesulfonate, perfluorohexanesulfonate, and perfluorooctanoate in retired fluorochemical production workers. Environ. Health Perspect. 2007, 115, 1298–1305. [Google Scholar] [CrossRef]
  32. Alexander, E.K.; Pearce, E.N.; Brent, G.A.; Brown, R.S.; Chen, H.; Dosiou, C.; Grobman, W.A.; Laurberg, P.; Lazarus, J.H.; Mandel, S.J.; et al. 2017 Guidelines of the American Thyroid Association for the Diagnosis and Management of Thyroid Disease During Pregnancy and the Postpartum. Thyroid 2017, 27, 315–389. [Google Scholar] [CrossRef] [Green Version]
  33. Lin, C.-Y.; Chen, P.-C.; Lin, Y.-C.; Lin, L.-Y. Association among serum perfluoroalkyl chemicals, glucose homeostasis, and metabolic syndrome in adolescents and adults. Diabetes Care 2009, 32, 702–707. [Google Scholar] [CrossRef] [Green Version]
  34. CDC Obesity Is a Common, Serious, and Costly Disease. Available online: https://www.cdc.gov/obesity/data/adult.html (accessed on 7 November 2020).
  35. VanderWeele, T.J. Commentary: Resolutions of the birthweight paradox: Competing explanations and analytical insights. Int. J. Epidemiol. 2014, 43, 1368–1373. [Google Scholar] [CrossRef]
  36. Daniel, R.M.; Cousens, S.N.; Stavola, B.L.D.; Kenward, M.G.; Sterne, J.A.C. Methods for dealing with time-dependent confounding. Stat. Med. 2013, 32, 1584–1618. [Google Scholar] [CrossRef] [PubMed]
  37. VanderWeele, T.J. Mediation Analysis: A Practitioner’s Guide. Annu. Rev. Public Health 2016, 37, 17–32. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  38. Inoue, K.; Yan, Q.; Paul, K.; Walker, D.; Jones, D.; Ritz, B. Air pollution and adverse pregnancy and birth outcomes: Mediation analysis using metabolomic profiles. Curr. Environ. Health Rep. 2020, in press. [Google Scholar] [CrossRef] [PubMed]
  39. Rosen, M.B.; Lee, J.S.; Ren, H.; Vallanat, B.; Liu, J.; Waalkes, M.P.; Abbott, B.D.; Lau, C.; Corton, J.C. Toxicogenomic dissection of the perfluorooctanoic acid transcript profile in mouse liver: Evidence for the involvement of nuclear receptors PPAR alpha and CAR. Toxicol. Sci. 2008, 103, 46–56. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  40. Heuvel, J.V.; Thompson, J.; Frame, S.; Gillies, P. Differential Activation of Nuclear Receptors by Perfluorinated Fatty Acid Analogs and Natural Fatty Acids: A Comparison of Human, Mouse, and Rat Peroxisome Proliferator-Activated Receptor-α, -β, and -γ, Liver X Receptor-β, and Retinoid X Receptor-α. Toxicol. Sci. 2006, 92, 476–489. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  41. Yu, N.; Wei, S.; Li, M.; Yang, J.; Li, K.; Jin, L.; Xie, Y.; Giesy, J.P.; Zhang, X.; Yu, H. Effects of Perfluorooctanoic Acid on Metabolic Profiles in Brain and Liver of Mouse Revealed by a High-throughput Targeted Metabolomics Approach. Sci. Rep. 2016, 6, 23963. [Google Scholar] [CrossRef] [Green Version]
  42. Kirkley, A.G.; Sargis, R.M. Environmental Endocrine Disruption of Energy Metabolism and Cardiovascular Risk. Curr. Diab. Rep. 2014, 14, 494. [Google Scholar] [CrossRef] [Green Version]
  43. Schaider, L.A.; Balan, S.A.; Blum, A.; Andrews, D.Q.; Strynar, M.J.; Dickinson, M.E.; Lunderberg, D.M.; Lang, J.R.; Peaslee, G.F. Fluorinated Compounds in U.S. Fast Food Packaging. Environ. Sci. Technol. Lett. 2017, 4, 105–111. [Google Scholar] [CrossRef]
  44. Susmann, H.P.; Schaider, L.A.; Rodgers, K.M.; Rudel, R.A. Dietary Habits Related to Food Packaging and Population Exposure to PFASs. Environ. Health Perspect. 2019, 127, 107003. [Google Scholar] [CrossRef]
  45. Liew, Z.; Olsen, J.; Cui, X.; Ritz, B.; Arah, O.A. Bias from conditioning on live birth in pregnancy cohorts: An illustration based on neurodevelopment in children after prenatal exposure to organic pollutants. Int. J. Epidemiol. 2015, 44, 345–354. [Google Scholar] [CrossRef] [Green Version]
  46. VanderWeele, T.J.; Mumford, S.L.; Schisterman, E.F. Conditioning on intermediates in perinatal epidemiology. Epidemiology 2012, 23, 1–9. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  47. Hernán, M.A.; VanderWeele, T.J. Compound Treatments and Transportability of Causal Inference. Epidemiology 2011, 22, 368–377. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  48. Hernán, M.A.; Taubman, S.L. Does obesity shorten life? The importance of well-defined interventions to answer causal questions. Int. J. Obes. 2008, 32 (Suppl. 3), S8–S14. [Google Scholar] [CrossRef] [Green Version]
  49. Pearl, J. Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes. J. Causal Inference 2018, 6. [Google Scholar] [CrossRef]
Toxics 08 00125 g001 550
Figure 1. A causal diagram that illustrates the potential causal relationship between PFAS exposure, the measure of obesity, and health outcomes. Obesitypre represents a measure of obesity status before the PFAS measurement and Obesitypost represents a measure of obesity after PFAS measurement. Obesitypre affects PFAS levels and health outcomes (i.e., a confounder), while Obesitypost is affected by PFAS and affects health outcomes (i.e., a mediator). In this scenario, we should adjust for Obesitypre in addition to other confounding variables such as the common causes of exposure and outcomes to unbiasedly estimate the total effect of PFAS exposure on health outcomes. Adjustment for Obesitypost is unnecessary and should be avoided. However, the exact timing for Obesitypre is hard to determine because serum or plasma PFAS levels could represent cumulative exposure. For persistent PFAS compounds, obesity may need to be measured at least a few years ahead of PFAS measurements to be considered as Obesitypre.
Figure 1. A causal diagram that illustrates the potential causal relationship between PFAS exposure, the measure of obesity, and health outcomes. Obesitypre represents a measure of obesity status before the PFAS measurement and Obesitypost represents a measure of obesity after PFAS measurement. Obesitypre affects PFAS levels and health outcomes (i.e., a confounder), while Obesitypost is affected by PFAS and affects health outcomes (i.e., a mediator). In this scenario, we should adjust for Obesitypre in addition to other confounding variables such as the common causes of exposure and outcomes to unbiasedly estimate the total effect of PFAS exposure on health outcomes. Adjustment for Obesitypost is unnecessary and should be avoided. However, the exact timing for Obesitypre is hard to determine because serum or plasma PFAS levels could represent cumulative exposure. For persistent PFAS compounds, obesity may need to be measured at least a few years ahead of PFAS measurements to be considered as Obesitypre.
Toxics 08 00125 g001
Toxics 08 00125 g002 550
Figure 2. Collider bias due to conditioning on the measure of obesity affected by PFAS when estimating the effect of PFAS exposure on health outcomes. In this DAG, [variable] represents the adjustment of the variable. Uncontrolled common causes of obesity and health outcomes are variables that were not included in the model. Measured confounders (common causes between the exposure and outcome) and obesitypre should be adjusted to remove confounding (backdoor) paths between PFAS and health outcomes. However, conditioning on obesitypost will (i) close a causal mediating path between PFAS exposure on health outcomes, and (ii) induce a biasing path via uncontrolled common causes of obesitypost and health outcomes (i.e., via a spurious association between PFAS and uncontrolled common causes of obesitypost and health outcomes as indicated by the dash line in this DAG). The latter is also known as “collider bias”.
Figure 2. Collider bias due to conditioning on the measure of obesity affected by PFAS when estimating the effect of PFAS exposure on health outcomes. In this DAG, [variable] represents the adjustment of the variable. Uncontrolled common causes of obesity and health outcomes are variables that were not included in the model. Measured confounders (common causes between the exposure and outcome) and obesitypre should be adjusted to remove confounding (backdoor) paths between PFAS and health outcomes. However, conditioning on obesitypost will (i) close a causal mediating path between PFAS exposure on health outcomes, and (ii) induce a biasing path via uncontrolled common causes of obesitypost and health outcomes (i.e., via a spurious association between PFAS and uncontrolled common causes of obesitypost and health outcomes as indicated by the dash line in this DAG). The latter is also known as “collider bias”.
Toxics 08 00125 g002
Table
Table 1. Associations between plasma PFAS levels (ng/mL) and maternal thyroid hormone levels with or without adjustment of pre-pregnancy body mass index (BMI; n = 1366).
Table 1. Associations between plasma PFAS levels (ng/mL) and maternal thyroid hormone levels with or without adjustment of pre-pregnancy body mass index (BMI; n = 1366).
PFASRelative % Difference of TSH (95% CI) 1
Model 1 2Model 2 3
PFOS
 Per IQR increase1.06 (0.96, 1.16)1.04 (0.96, 1.14)
 Quartile 1RefRef
 Quartile 20.85 (0.68, 1.07)0.86 (0.69, 1.06)
 Quartile 30.93 (0.75, 1.16)0.96 (0.78, 1.17)
 Quartile 41.03 (0.84, 1.27)1.01 (0.83, 1.22)
PFOA
 Per IQR increase1.02 (0.94, 1.11)1.01 (0.93, 1.10)
 Quartile 1RefRef
 Quartile 20.95 (0.76, 1.19)0.96 (0.78, 1.19)
 Quartile 31.01 (0.81, 1.25)1.02 (0.83, 1.25)
 Quartile 41.09 (0.86, 1.39)1.08 (0.86, 1.36)
IQR, interquartile range; TSH, thyroid-stimulating hormone; BMI, body mass index. 1 Multivariable linear regression models were employed to estimate the relationships between PFAS levels and ln-TSH. 2 Adjusted for maternal age, parental socio-occupational status, parity, maternal smoking, and birth year. 3 Adjusted for pre-pregnancy BMI (<18.5, 18.5 to <25, 25 to <30, and ≥30 kg/m2) in addition to variables included in Model 1.
Table
Table 2. Associations between plasma PFAS levels (ng/mL) and maternal thyroid hormone levels stratified by maternal overweight status indicated by body mass index (BMI).
Table 2. Associations between plasma PFAS levels (ng/mL) and maternal thyroid hormone levels stratified by maternal overweight status indicated by body mass index (BMI).
PFASRelative % Difference of TSH (95% CI) 1
Non-Overweight, BMI < 25
(n = 1002)		Overweight,
BMI ≥ 25
(n = 364)		P for Interaction2
PFOS
 Per IQR increase1.08 (0.97, 1.21)0.94 (0.82, 1.09)0.19
 Quartile 1RefRef-
 Quartile 20.87 (0.67, 1.13)0.79 (0.56, 1.12)0.83
 Quartile 31.05 (0.83, 1.33)0.64 (0.42, 0.98)0.15
 Quartile 41.03 (0.83, 1.28)0.95 (0.66, 1.35)0.84
PFOA
 Per IQR increase1.04 (0.94, 1.15)0.95 (0.82, 1.10)0.24
 Quartile 1RefRef-
 Quartile 21.09 (0.84, 1.42)0.68 (0.50, 0.92)0.03
 Quartile 31.15 (0.90, 1.46)0.75 (0.53, 1.05)0.04
 Quartile 41.12 (0.86, 1.48)1.02 (0.72, 1.43)0.52
IQR, interquartile range; TSH, thyroid-stimulating hormone; BMI, body mass index. 1 Multivariable linear regression models were employed to estimate the relationships between PFAS levels and ln-TSH adjusting for maternal age, parental socio-occupational status, parity, maternal smoking, and birth year. 2 Tests of heterogeneity were also performed by assessing the p-value of the interaction term for each PFAS and overweight status in the regression models.
Table
Table 3. Associations between serum PFAS levels (ng/mL) and cardiovascular diseases in models with or without adjustment of body mass index (BMI; n = 7411).
Table 3. Associations between serum PFAS levels (ng/mL) and cardiovascular diseases in models with or without adjustment of body mass index (BMI; n = 7411).
PFASPrevalence Ratio of Cardiovascular Diseases (95% CI)
Model 1 aModel 2 b
PFOS
 Per IQR increase1.06 (1.03, 1.09)1.06 (1.03, 1.09)
 Quartile 1RefRef
 Quartile 21.08 (0.76, 1.52)1.07 (0.76, 1.51)
 Quartile 31.19 (0.84, 1.67)1.18 (0.84, 1.67)
 Quartile 41.18 (0.85, 1.64)1.19 (0.86, 1.66)
PFOA
 Per IQR increase1.04 (1.01, 1.08)1.04 (1.01, 1.08)
 Quartile 1RefRef
 Quartile 21.01 (0.75, 1.36)1.00 (0.75, 1.34)
 Quartile 31.14 (0.82, 1.58)1.14 (0.83, 1.56)
 Quartile 41.30 (0.99, 1.70)1.31 (1.00, 1.71)
IQR, interquartile range; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; BMI, body mass index. a Multivariable modified Poisson regression models were employed to estimate the relationships between PFAS levels and CVD prevalence adjusting for age, sex, race/ethnicity, education status, income, marital status, smoking, systolic blood pressure, eGFR, HbA1c, statin prescriptions, and history of cancer. b Adjusted for BMI (<18.5, 18.5 to <25, 25 to <30, and ≥30 kg/m2) in addition to Model 1.
Table
Table 4. Associations between serum PFAS levels (ng/mL) and cardiovascular diseases stratified obesity status indicated by body mass index (BMI).
Table 4. Associations between serum PFAS levels (ng/mL) and cardiovascular diseases stratified obesity status indicated by body mass index (BMI).
PFASPrevalence Ratio of Cardiovascular Diseases (95% CI) a
Obese, BMI < 30
(n = 4796)		Non-Obese, BMI ≥ 30
(n = 2615)		P for Interactionb
PFOS
 Per IQR increase1.09 (1.04, 1.14)1.04 (1.01, 1.09)0.18
 Quartile 1refref-
 Quartile 20.89 (0.58, 1.35)1.34 (0.78, 2.29)0.28
 Quartile 31.01 (0.69, 1.50)1.51 (0.90, 2.54)0.29
 Quartile 41.08 (0.72, 1.60)1.39 (0.85, 2.27)0.64
PFOA
 Per IQR increase1.01 (0.95, 1.08)1.07 (1.04, 1.11)0.10
 Quartile 1refref-
 Quartile 20.77 (0.53, 1.11)1.38 (0.85, 2.23)0.04
 Quartile 31.01 (0.65, 1.55)1.35 (0.87, 2.09)0.35
 Quartile 41.03 (0.73, 1.46)1.79 (1.15, 2.77)0.05
IQR, interquartile range; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; BMI, body mass index. a Multivariable modified Poisson regression models were employed to estimate the relationships between PFAS levels and CVD prevalence adjusting for age, sex, race/ethnicity, education status, income, marital status, smoking, systolic blood pressure, eGFR, HbA1c, statin prescriptions, and history of cancer. b Tests of heterogeneity were also performed by assessing the p-value of the interaction term for each PFAS and obesity status in the regression models.
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Inoue, K.; Goto, A.; Sugiyama, T.; Ramlau-Hansen, C.H.; Liew, Z. The Confounder-Mediator Dilemma: Should We Control for Obesity to Estimate the Effect of Perfluoroalkyl Substances on Health Outcomes? Toxics 2020, 8, 125. https://doi.org/10.3390/toxics8040125

AMA Style

Inoue K, Goto A, Sugiyama T, Ramlau-Hansen CH, Liew Z. The Confounder-Mediator Dilemma: Should We Control for Obesity to Estimate the Effect of Perfluoroalkyl Substances on Health Outcomes? Toxics. 2020; 8(4):125. https://doi.org/10.3390/toxics8040125

Chicago/Turabian Style

Inoue, Kosuke, Atsushi Goto, Takehiro Sugiyama, Cecilia Høst Ramlau-Hansen, and Zeyan Liew. 2020. "The Confounder-Mediator Dilemma: Should We Control for Obesity to Estimate the Effect of Perfluoroalkyl Substances on Health Outcomes?" Toxics 8, no. 4: 125. https://doi.org/10.3390/toxics8040125

APA Style

Inoue, K., Goto, A., Sugiyama, T., Ramlau-Hansen, C. H., & Liew, Z. (2020). The Confounder-Mediator Dilemma: Should We Control for Obesity to Estimate the Effect of Perfluoroalkyl Substances on Health Outcomes? Toxics, 8(4), 125. https://doi.org/10.3390/toxics8040125

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop