2.5. Statistical Analysis

Descriptive statistics were calculated for the exposure, outcomes, and covariates of interest. Concentrations below the limit of detection for TT, fT, E3, and missing values for E3 were replaced by the LOD divided by the squared root of 2. Sex steroid hormones and cadmium concentrations were not normally distributed and were therefore log transformed. We examined correlations between cadmium and hormone concentrations using Spearman correlation. We also calculated a two-way mixed-effect intraclass correlation coefficient (ICC) to assess reliability in the urinary cadmium measurements [58]. For our primary analyses, to capitalize on the repeated measurements across pregnancy, we fit linear mixed models (LMMs) for each hormone with an unstructured correlation matrix, a fixed effect for cadmium, a smooth function for gestational age, and a random effect for each participant. We expressed results as the mean percent difference (∆%) in hormone concentrations associated with a ln-unit increased in cadmium. We evaluated potential confounders in multiple ways: a priori based on previous research, using DAGs, and assessing changes in effect estimates in bivariate models. Final covariates for TT and fT models included maternal age, gestational age at visit, maternal race, maternal ethnicity, highest maternal education, and pre-pregnancy BMI. Models for E1, E2, and E3 included those variables and were additionally adjusted for fetal sex, parity, maternal smoking during pregnancy. Secondarily, we fitted trimester-specific models, consisting of crude and adjusted linear regression models, examining the association between cadmium and each sex steroid hormone. In light of the substantial literature suggesting sex-specific impacts of cadmium on perinatal outcomes [59,60], we examined interaction terms between cadmium and fetal sex. Additionally, we stratified by fetal sex to better assess this potential effect modification [61]. Finally, we conducted two sensitivity analyses excluding the following (1) participants who reported any smoking in pregnancy; and (2) participants who had any major pregnancy complications (preeclampsia, gestational diabetes, preterm birth). Probability values (p-values) < 0.05 were considered statistically significant. Analyses were performed in SAS version 9.4.