br Covariate information br Information on potential covaria
2.5. Covariate information
Information on potential covariates was derived from a series of mailed surveys (a baseline questionnaire completed in 1995–1996 at CTS enrollment and five follow-up surveys), as well as a survey ad-ministered by an interviewer at the time of blood draw. Factors con-sidered as potential covariates included established breast cancer risk factors (based on a review of the literature) as well as factors that prior exploratory analyses had identified as correlates to serum PBDE levels in this ITF2357 (Givinostat) study population. This initial set of potential covariates included information on: demographics (age, race, neighborhood socioeconomic status and urbanization); timing of blood draw (date and season of blood collection); behavioral factors (smoking, alcohol consumption, physical activity, use of menopausal hormone therapy (HT)); diet (total fat, fiber, vitamin D, calories, red meat, pork, fish, and total meat consumption); body mass index (BMI = weight in kg divided by square of height in meters) and changes in body weight; family history of breast cancer; and reproductive history (age at menarche, age at first full term pregnancy, lactation history, menopausal status).
2.6. Statistical analysis
All analyses were conducted in SAS Version 9.04 (SAS Institute Inc. and SAS Institute Inc., 2007). Statistical significance was defined at a p-value < 0.05. Concentrations below the LOD for which no signal was detected were estimated by single imputation from a log-normal probability distribution based on the observed distribution of quantified measurements, following the method suggested by Lubin et al. (2004). In order to minimize potential biases associated with imputing high frequency of non-detectable levels, only the three congeners with de-tection frequencies (DF) of 75% or more were included in our risk analyses. These included: 2,2′,4,4′-tetrabromodiphenyl ether (BDE-47), 2,2′,4,4′,6-pentabromodiphenyl ether (BDE-100), and 2,2′,4,4′,5,5′-hexabromodiphenyl ether (BDE-153).
Prior to conducting the risk analyses, a number of exploratory and descriptive analyses were conducted. Spearman rank correlation coef-ficients (r) between BDE congeners were calculated. Because the con-geners were highly correlated, we considered each PBDE separately in our risk analyses. The risk of breast cancer associated with each BDE congener was estimated by unconditional logistic regression using PROC LOGISTIC to generate odds ratios (OR) and 95% confidence intervals (95% CI). These models were run on the measured and imputed wet weight values (expressed as log10 [BDE, ng/mL]), adjusting for total serum lipid Environment International 127 (2019) 412–419
content by the addition of a separate term in the model (expressed as log10 [total lipids, ng/mL]), as recommended by Schisterman et al. (2005). Smoothing splines were considered in generalized additive models (using PROC GAM) and evaluated to assess potential non-line-arities in the relationship between each PBDE and the log-odds of breast cancer but no evidence of non-linearity was observed. In addition to estimating breast cancer risks for the log-linear continuous values of PBDE concentrations, we also estimated risks for quartiles of PBDE concentrations based on the distribution among controls.
Minimally-adjusted crude ORs were generated from models that included adjustment only for total serum lipids and the matching design variables (age at enrollment, race/ethnicity, and study site). Fully-ad-justed multivariable models were built via a two-step process. First, a backwards elimination approach was used, starting with a model that forced inclusion of the BDE variable, the matching design variables, and serum lipid content and retention of covariates for which the p-value for the Wald chi-square was < 0.05. We then further evaluated potential confounders by adding each of the excluded variables back into the model one at a time and evaluated the change in the estimated OR for the BDE variable. Factors that changed the estimated OR for the BDE by ≥10% were retained in our final multivariable models. While we conducted this process separately for each BDE, it resulted in the same set of covariates for all congeners. Final multivariable models included terms for: age at baseline, race/ethnicity, study site, total serum lipid content, date of blood draw, season of blood draw, BMI at baseline, long-term moderate and strenuous physical activity, family history of breast cancer, parity/age at first full term pregnancy, menopausal status/HT use, and pork consumption at baseline (see Table 1 for details of how these factors were specified in the models).