Prenatal exposure to trans fatty acids and head growth in fetal life and childhood: triangulating confounder-adjustment and instrumental variable approaches

Setting and participants

This study was embedded in the population-based Generation R cohort. Pregnant women with an expected delivery date between April 2002 and January 2006 in Rotterdam were eligible [14]. The Generation R Study was approved by the Medical Ethics Committee of the Erasmus Medical Center, and written informed consent was obtained from adult participants.

A total of 8633 live singletons were born to women recruited in pregnancy. Among these children, 1710 were excluded due to missing data on maternal plasma TFA concentration during pregnancy. After further exclusion of those without ultrasound data on HC in the second or third trimester, 6900 children constituted the study population. Of these, 2933 children underwent a structural brain magnetic resonance imaging (MRI) session at age 9–11 years [15], and 2354 had usable data after quality control (see Figure S2 for the flow-chart).

Maternal TFA concentration during pregnancy

As previously described [16], maternal plasma fatty acids concentrations were assessed in mid-gestation (mean 20.6 weeks, SD = 1.1) using gas chromatography. The concentrations of individual fatty acids are expressed as weight percentage (%, wt:wt) of all glycerophospholipid fatty acids detected with a chain length between 14 and 22 carbon atoms. For the current study, total TFA concentration was calculated by summing the concentrations of t16:1, t18:1, and tt18:2 isomers. In our sample, the interquartile range (IQR) of total TFA concentration was 0.27–0.41%, and the t18:1 isomer (IQR 0.16–0.28%) accounted for the majority (on average 63%) of total TFAs. Absolute concentration (i.e., mg/L) of TFA was used in a supplementary analysis.

Head growth in fetal life and childhood

Fetal HC was measured with ultrasound during each trimester of pregnancy by technicians at the Generation R Study research center [14]. First trimester ultrasounds were used for pregnancy dating. In our study population, fetal HC data were available in 6792 children in the second trimester (mean = 20.6 weeks, SD = 1.2), and 6625 children in the third trimester (mean = 30.4 weeks, SD = 1.1). The intra- and inter-observer correlation coefficients for fetal HC were 0.995 and 0.988, suggesting excellent reliability. In addition, we calculated fetal HC growth rate (in cm/week) from the second to third trimester as the difference between HC divided by the difference in gestational age in 6517 children. HC at birth measured using standardized procedures was available in 3752 children [17].

Brain morphology in childhood was assessed using MRI at age 9–11 years. All images were acquired on a 3-Tesla GE Discovery MR750w MRI System (General Electric, Milwaukee, WI, USA) scanner using an 8-channel head coil. High-resolution T1-weighted sequences were obtained using a 3D coronal inversion recovery fast spoiled gradient recalled (IR-FSPGR, BRAVO) sequence with 1-mm isotropic resolution, and images were rated for quality control both during and after the MRI acquisition [15]. Volumetric segmentation and cortical reconstruction were performed with FreeSurfer v.6.0.0 (http://surfer.nmr.mgh.harvard.edu/), and the standard reconstruction stream was applied. The quality of FreeSurfer output was visually inspected, and data with insufficient quality were eliminated. For the current study, total brain volume, and volumes of cortical gray matter and cerebral white matter were used to quantify global brain size.

Covariates

We included the following maternal and child characteristics as covariates based on prior literature [13, 18]: child sex and (gestational) age at head assessment, maternal ethnicity, age at enrollment, pre-pregnancy body mass index (BMI), psychopathology, marital status, parity, educational level, total energy intake, diet quality, smoking and alcohol use during pregnancy, and family income. Crown-rump length or bi-parietal diameter was used to determine gestational age. Maternal psychopathology was assessed using the Brief Symptom Inventory (scores range from 0 to 4, with higher scores indicating more clinically relevant psychological symptoms) [19]. Maternal diet quality was assessed in early pregnancy using a semi-quantitative 293-item food frequency questionnaire and quantified by an overall score ranging from 0 to 15, with higher score reflecting better adherence to Dutch dietary guidelines [20].

Additional variables

Information on calendar time of maternal TFA assessment during pregnancy was used as an IV for the IV analysis. The recruitment of pregnant women approximately concurred with the initiative on TFAs reduction (see Fig. 1 and Figure S1).

Fig. 1figure 1

Maternal trans fatty acid concentration during pregnancy per calendar time of assessment. Maternal TFA concentration (%, wt:wt, mean ± se) per calendar time of TFA assessment. Maternal TFA concentration was assessed in the plasma at 20.6 (1.1) weeks of gestation. The Dutch initiative to further reduce TFA content in food took effect in 2003. Abbreviation: TFA, Trans fatty acid

Given the evidence suggesting a relation between prenatal HC and childhood cognitive development [7, 8], child nonverbal IQ assessed at age 5–9 (mean = 6.2) years was used in a secondary analysis [21]. We examined the association of HC in utero with cognitive abilities in childhood in our sample to highlight the clinical implications of our findings.

Statistical analysis

For descriptive purposes, continuous variables were presented as mean (SD) and categorical variables as number (%). In a non-response analysis, we describe the characteristics of children with and without MRI data at 9–11 years, using analysis of variance (ANOVA) or Wilcoxon test for continuous variables, and chi-square tests for categorical variables.

We investigated the association of maternal TFA concentration with HC in second and third trimesters, fetal HC growth rate, and global brain volume (i.e., total brain volume, cortical gray matter volume, and cerebral white matter volume) at age 9–11 years using linear regression. We also examined maternal TFA concentration in relation to HC at birth in a supplementary analysis. These regression analyses were performed using multi-stage models with different levels of confounder adjustment, to explore how adjustment for measured confounding affects estimates. In the initial model, only child sex and (gestational) age at outcome assessment were adjusted for. Using a change-in-estimate criterion of 5% [22], maternal ethnicity, age at enrollment, educational level, diet quality, smoking during pregnancy, and family income were selected from the covariate pool as confounders and were additionally adjusted for in a second model. In a third model, we additionally corrected for concentrations of maternal essential fatty acids (EFAs, i.e., α-linoleic acid and linoleic acid) and long-chain polyunsaturated fatty acids (LC-PUFAs), because they can be either confounders (i.e., reductions of TFA may increase PUFAs in body fat composition) or intermediates (i.e., TFA may interfere with EFAs and reduce the synthesis of LC-PUFAs) [23,24,25]. Further, models with MRI measures were weighted by inverse probability weights to account for attrition (see supplementary methods for details). We examined non-linear relations by comparing the linear model and the model including splines using likelihood ratio test.

To explore implications on long-term neurocognitive development, we investigated prenatal exposure to TFA, and HC and HC growth rate in utero in relation to nonverbal IQ at age 6 years in a secondary analysis.

Next, for head outcomes that were substantially associated with maternal TFA concentration using confounder-adjustment approaches, we performed an IV analysis using two-stage least squares (TSLS) estimation [26]. Calendar time of maternal TFA assessment was proposed as the IV given the initiative to reduce food TFA content in the Netherlands since 2003. In addition, we assessed the basic assumptions of IV analysis (see supplementary methods for details) to evaluate whether calendar time of maternal TFA assessment could be utilized as an IV. For the ‘relevance’ assumption, we examined the association between calendar time of maternal TFA assessment and maternal TFA concentration, both with and without adjusting for covariates. In addition, we performed an F-test on the instrument in the first stage of the TSLS regression. Since the ‘exclusion restriction’ and ‘exchangeability’ assumptions cannot be verified, we explored our rich observed data in ways that might falsify these assumptions. These analyses were performed to minimize the risk that a covariate with a similar time trend as maternal TFA concentration underlies the relation between calendar time of maternal TFA assessment and child head measures. First, we related calendar time of maternal TFA assessment to each covariate using separate linear or logistic regression model, and covariates that did not vary by calendar time of maternal TFA assessment were excluded. Second, the mean ± standard error (se) of each remaining covariate over calendar time of maternal TFA assessment was plotted to inspect time trends. In the primary IV analysis, we adjusted for no covariates, and the sandwich estimator was used because it is robust to heteroscedastic errors. In the secondary analysis, we conducted two TSLS regressions to rule out bias caused by any measured nuisance variables: first, we adjusted for covariates with a potential monotonic time trend; second, we adjusted for all covariates regardless their variation over calendar time. Since maternal TFA concentration was the dependent variable in the first-stage model of the IV analysis, we log-transformed the raw values to obtain a normal distribution. Finally, because we are interested in seeing how the IV analysis results complement the confounding-adjustment regression results, we interpret our IV analyses as primarily testing a causal null hypothesis [27].

We performed two sensitivity analyses to test the robustness of our findings. First, we re-ran the analysis in participants of Dutch national origin only. Second, we examined the trans 18:1 fatty acid, the primary TFA isomer, as exposure of interest.

Missing data on covariates were accounted for by multiple imputation. We generated 20 imputed datasets with 20 iterations, and report pooled results. Statistical significance was set as α < 0.05 (two-sided), and a false discovery rate (FDR) correction was performed in primary analyses [28]. All statistical analyses were performed using R version 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria), including the ‘ivpack’ package for IV analysis.

留言 (0)

沒有登入
gif