Our overall
analytical scheme,
as well as the statistical models and tests used, is shown in Figure 1. First, we adjusted
each chemical for potential a priori defined potential confounders
in the model in addition to total lipids (for lipophilic chemicals, n = 56) and creatinine (for urinary chemicals, n = 61) to reduce the estimated variability and susceptibility to
bias.30 (link),31 (link) Specifically, we adjusted PCBs, OCPs, and
polybrominated chemicals for total lipids and all other urinary chemicals
for creatinine. All chemicals were log-transformed (x + 1), and continuous
covariates were rescaled to have a zero mean and unit variance (Figure 1A). Since the biomarkers
of chemical exposure need not follow a normal distribution, we calculated
the Spearman’s rank correlation (rs) matrices for EDCs in females, males, and couples after extracting
the residuals.
We estimated the sex-specific difference with a
paired t test (by household) after extracting the
residuals from
a linear model adjusted for age (Figure 1B). We used a similar approach to estimate
the percentage of variance explained by the shared environment. However,
sex and age variables were excluded in the adjustment step to isolate
their effects (Figure 1C). Afterward, we extracted and regressed the residuals against the
household variable to obtain the adjusted coefficient of determination
(R2).
We leverage the family based
design and diverse chemical measurements
in the LIFE Study to study household correlations of exposures. It
is important to know how generalizable our findings are to the U.S.
population and therefore we computed concordance as the
Pearson correlation coefficient (r) between the chemical
relatedness rs in this study and that
in the 2003–2004 National Health and Nutrition Examination
Survey (NHANES).9 (link) We used r as a measure of concordance to model the linear relationships between
chemicals. First, we estimated the rs of
4292 unique chemical pairs in NHANES without filtering by age, sex,
and race (median sample size was 1923). Then, we estimated the concordance
based on a total of 101 matched biomarkers between the two studies.
We chose the 2003–2004 NHANES because of the close temporal
nature with the implementation of the LIFE Study and given that the
same laboratory measured persistent EDCs.
We used all the instrument
derived concentrations for the analyses.32 (link),33 (link) For missing values, we substituted them by multiple imputation,
assuming a missing-at-random scenario.17 (link) We conducted imputation based on the information from available
demographics, previous history of clinical symptoms, and all other
chemical variables and created a total of 10 imputed data sets for
males and females separately.
We visualized the correlations
between exposures as exposome globe
using the R package Circlize (v 0.3.1).34 (link) EDCs were sorted from lipophilic to hydrophilic to aid visual interpretation
of the patterns. We combined the final estimates from imputations
using Rubin’s method35 (link) and calculated
the p values of correlations by permutation tests.
To adjust for multiple testing, we used the false discovery rate (FDR) q values unless otherwise specified. We executed all analyses
using the computing environment R (v 3.3.1).36 For reproducibility purposes, all analytic code is publicly available
on GitHub via an MIT license (https://github.com/jakemkc/exposome_variability).