DNA Day TOP ARTICLES SUPPLEMENT
Total Page：16
File Type：pdf, Size：1020Kb
DNA Day TOP ARTICLES SUPPLEMENT CONTENTS RESEARCH ARTICLE: Misclassified exposure in epigenetic mediation analyses. Does DNA methylation mediate effects of smoking on birthweight? Epigenomics Vol. 9 Issue 3 REVIEW: Epigenetics and allergy: from basic mechanisms to clinical applications Epigenomics Vol. 9 Issue 4 RESEARCH ARTICLE: DNA methylation at diagnosis is associated with response to disease-modifying drugs in early rheumatoid arthritis Epigenomics Vol. 9 Issue 4 SPECIAL FOCUS y Effects of the in utero environment on the epigenome Research Article For reprint orders, please contact: [email protected] 9 Research Article 2017/02/28 Misclassified exposure in epigenetic mediation analyses. Does DNA methylation mediate effects of smoking on birthweight? Epigenomics Aims: Assessing whether epigenetic alterations mediate associations between Linda Valeri*,1,2, Sarah L environmental exposures and health outcomes is increasingly popular. We investigate Reese3, Shanshan Zhao3, the impact of exposure misclassification in such investigations. Materials & methods: Christian M Page4, Wenche 4 5 We quantify bias and false-positive rates due to exposure misclassification in mediation Nystad , Brent A Coull & Stephanie J London3 analysis and assess the performance of the simulation extrapolation method (SIMEX). 1Department of Psychiatry, Harvard We evaluate whether DNA-methylation mediates smoking–birth weight relationship Medical School, Boston, MA 02115, USA in the Norwegian Mother and Child Study birth cohort. Results: Ignoring exposure 2Psychiatric Biostatistics Laboratory, misclassification increases type I error in mediation analysis. The direct effect is McLean Hospital, Belmont, MA 02478, underestimated and, when the mediator is a biomarker of the exposure, as is true for USA 3 smoking, the indirect effect is overestimated. Conclusion: Misclassification correction National Institute of Environmental Health Sciences, National Institutes of plus cautious interpretation are recommended for mediation analyses in the presence Health, Department of Health & Human of exposure misclassification. Services, Research Triangle Park, NC, USA 4National Institute of Public Health, Oslo, First draft submitted: 17 October 2016; Accepted for publication: 22 December 2016; Norway 5 Published online: 21 February 2017 Department of Biostatistics, Harvard T.H. Chan School of Public Health *Author for correspondence: Keywords: DNA methylation • mediation analysis • misclassification [email protected] Evidence is accumulating that environmen- biased reporting occurs whereby some pro- tal exposures modify the epigenome. In portions of smokers falsely claim, on surveys, humans, the best-studied epigenetic modi- to be nonsmokers [12] . In addition, because of fication is methylation and the best-studied the well-publicized negative health impacts exposure is smoking. Smoking in adults has of maternal smoking during pregnancy been reproducibly associated with alterations on the developing fetus, pregnant women 3 in methylation at specific loci [1] . Similar under-report smoking more than nonpreg- effects have been seen in newborns whose nant smokers of reproductive age [13] . None- mothers smoked during pregnancy [2]. These theless, studies that address whether meth- smoking-methylation signals have been used ylation signatures from smoking mediate its 2016 to develop novel biomarkers of exposure [3– health outcomes have ignored the potential 6]. In addition to its value as an exposure role of measurement error in assessment of biomarker, there is great interest in the pos- smoking [7–10,14]. Given this measurement sibility that differential methylation at rel- error, evaluation of mediation may be com- evant loci mediates well-established associa- plicated by the fact that the proposed media- tions between smoking and disease, both for tors, DNA sites differentially methylated by adult [7,8] and in utero exposures [9,10]. smoking, are excellent biomarkers that may It is widely acknowledged that measure- better capture the exposure than self-reported ment of human environmental exposures, smoking [3–6]. including smoking, is prone to error [11] . In the field of mediation analysis, bias Random error exists for all exposures. How- introduced by measurement error in ever, for smoking in particular, differential the mediator variable has been investi- part of 10.2217/epi-2016-0145 © US Government Epigenomics (2016) 9(3), 253–265 ISSN 1750-1911 253 Research Article Valeri, Reese, Zhao et al. gated [15–17] . However, misclassification of the expo- weight that is mediated by the DNA-methylation level, sure variable has not been well evaluated. Reduced relative to the direct effect of smoking on birthweight birth weight is a well-established sequelae of maternal through pathways independent of DNA methyla- smoking during pregnancy [18]. Given strong evidence tion (Figure 1B). Under the counterfactual framework of differential methylation in newborns in relation to for causal inference direct and indirect causal effects smoking by the mother [2], it has been of interest to have been rigorously defined [21,22] (section A1 of the consider whether these signals mediate the effects of Supplementary material). maternal smoking on birth weight. It has recently been To validly estimate direct and indirect effects, the reported that differential DNA methylation of a single following four confounding assumptions need to be CpG site in placenta mediates up to 36% of the effect satisfied. Conditioning on covariates C, there is no of smoking on lower birth weight [9]. In another study, unmeasured confounding of the exposure–outcome differential methylation in newborn’s blood at a single relationship, the mediator–outcome relationship, the CpG site in a different gene was reported to mediate exposure–mediator relationship and there are no medi- 19–46% of the relationship between smoking and ator–outcome confounders affected by the exposure. birthweight [10] . Because self-reported smoking status See [22,23] for further discussion of these assumptions. during pregnancy is prone to misclassification, we were Furthermore, models for the outcome and mediator interested in evaluating the sensitivity of mediation need to be correctly specified. For continuous outcome analysis with methylation data to exposure misclassi- and mediator (as in the current setting of outcome birth fication bias. For this purpose, we considered a pub- weight and mediator methylation), under the assump- lished scenario in perinatal epidemiology, for which we tion of no exposure–mediator interaction in the out- have relevant data [10], as an example. come model, typically made by published applications Our study makes several contributions. First, we of mediation analysis in environmental epigenetics, if study the impact of exposure misclassification on the we specify three linear regression models: estimation of direct and indirect causal effects and ++oo+o testing of the indirect effect in methylation studies EY( Aa= , C = cc) = ii01++i ' analytically. Second, we assess the impact of misclas- sification on estimation and testing via a simulation EY( Aa= , M ==ma,)Cc= ii01++i2 m + i'c study. Third, we evaluate the ability of the SIMEX approach to adjust for exposure misclassification in EM( Aa==,)Cc= b0 +b1 a + b'c this setting. Finally we use data from the Norwegian Mother and Child Cohort Study (MoBA) [19,20] to con- then the estimators of total effect (TE), direct effect duct a mediation analysis accounting for misclassifi- (NDE) and indirect effect (NIE) take the form [24,25]: cation of self-reported smoking status using SIMEX. +o Our study provides evidence that ignoring misclassifi- TE = i1 cation can bias results of mediation analyses and shows the value of incorporating misclassification correction NDE = i1 in mediation analysis in the context of environmental +o epigenetic studies. NIE ==bi12 ii11- Methods Under the assumption of no unmeasured confound- Mediation analysis in the absence of exposure ing and that models 1–3 are correctly specified, these misclassification estimators (which are equivalent to the ones proposed With reference to our example of mediation of the by [24] in the psychology literature) can be interpreted effect of maternal smoking during pregnancy on as causal direct and indirect causal effects [22]. Esti- newborn birth weight by smoking-related differential mators for direct and indirect effects in the presence methylation, let A denote the exposure, maternal smok- of exposure-mediator interaction are given in Section ing and M denote the mediator, DNA methylation. Let A1 of the Supplementary Material. A discussion on the Y denote the outcome, birth weight and C denote a comparison between traditional and causal inference vector of covariates representing potential confound- approaches to mediation analysis is given in [26]. ers. The directed acyclic graph in Figure 1 describes the The most popular test for indirect effects is based on setting of mediation analysis. Mediation analysis can the product method, also known as the Sobel test [27]. be employed to quantify how much of the total effect This is a Wald test for the null hypothesis H0: β1 θ2 = 0 of maternal smoking on birth weight (Figure 1A) is based on the delta method standard error 2 222 2 2 explained by the indirect effect of smoking on birth vvNIE = ib21bv1 + i2 where vi2 and vb1 are the 254 Epigenomics (2016) 9(3) future science group Exposure misclassification in mediation analysis