Methodology for Study Aim 2

The goal of Study Aim 2 is to model the predictors of change in DDE and PCBs, so as to provide a model for exposure estimation at a time of etiologic relevance to menopause (i.e. Study Aim 3). Of the 123 women with both baseline and follow-up organochlorine measures, one individual had missing pregnancy information and was excluded from this analysis.

3.4.1 Predictive model

Although changes in serum DDE and PCB levels over time are driven by complex age- and cohort-related processes, we presume the changes in organochlorine levels during the period under study could be reasonably approximated by an exponential decay, as other similar compounds and biological systems tend to operate on an exponential decay [21, 22]. Thus, the basic predictive model was:

[Follow-up] = [Baseline] ∗ exp ( - slope ∗ follow-up time) which can be re-written as

Log [Follow-up] – Log [Baseline] = - slope follow-up time

[Follow-up] and [Baseline] are the wet weight organochlorine concentrations, respectively, at follow-up and baseline, and “follow-up time” represents the years from baseline to follow- up. Slope represents change in log concentrations per unit of follow-up time. To determine whether slope (i.e. the rate of change) is constant or is affected by various determinants,

Log [Follow-up] – Log [Baseline] = - (α+β1 (X1) . . . + βk (Xk)) follow-up time

where α is the intercept and βk is the regression coefficient for the k-th predictor, Xk.

3.4.2 Potential predictors

Initial concentration, lactation duration, body mass index at baseline, percent change in fat mass, and mother’s date of birth were evaluated for their possible influence on the rate of change in DDE and PCBs from the baseline to the follow-up study. Univariate statistics, including frequencies, means and medians were generated for these variables. The

distributions of each of the variables were examined to check for missing data. One individual with missing pregnancy information was dropped from analysis. Missing observations for weight at baseline (n=3) and at follow-up (n=5) were derived or imputed from other available data. For missing baseline weight, the reported weight at the 6 week and 6 month interviews following the index pregnancy was used. For missing follow-up weight, categories of weight were available at the follow-up interview and the midpoint of each category was taken as the weight). Two individuals were missing age at menarche. One reported her first period in 5th grade and was assigned 11 as her age at menarche. Various coding schemes (continuous and categorical) were explored. Each potential

predictor was examined to obtain the fewest categories needed to adequately characterize the relationship between that variable and the slope.

Initial (baseline) DDE and PCB concentrations (ug/L serum) were log-transformed, given their log normal distributions, and were included in the model as a linear variable.

Lactation duration, as assessed from the follow-up questionnaire, was defined as the number of weeks the participant breast-fed twice or more per day. Given the diminishing role of exclusive breast-feeding as lactation proceeds, the effect of breast-feeding on the rate of organochlorine excretion could lessen later in lactation. Thus, lactation duration was treated as a continuous variable and fitted with a piecewise linear model over 3 intervals of lactation duration, with the slope of each interval estimated directly from the data. This was accomplished by defining 3 variables representing the number of weeks of lactation during the 3 intervals: 0-26th, 27th-39th, and >39th lactational week. The number of weeks of breast- feeding for each pregnancy from baseline to follow-up was split among these three variables. Additional cutpoints at the quarter (13 weeks) and the 1 year (52 weeks) time points did not change the results. The number of weeks of breast-feeding for each pregnancy from baseline to follow-up was split among these three variables. For women with multiple lactations between baseline and follow-up, the contribution from each pregnancy was summed. To illustrate, for a woman with one child whose lactation lasted 60 weeks, a value of 26, 13 and 21 weeks was assigned to the first, second and third lactation variables, respectively. If she also had a second child who was breastfed for 30 weeks, then an additional 26 weeks was assigned to the first lactation variable and 4 weeks to the second, giving final values of 52, 17, and 21 for the three lactation variables. To operationalize in modeling, the value (i.e. number of breastfeeding weeks) of each of the three variables would be multiplied by the beta coefficients obtained for the corresponding variable.

Baseline body mass index was included in the model as a categorical predictor. Due to its narrow distribution, cutpoints for body mass index were set at 20 and 23 kg/m2.

Percent change in body fat (kg) was based on a validated formula for percent body fat, which was then converted to body fat by multiplying by body weight [23]:

Body fat = [((1.46∗BMI) + (0.14∗age) – 10) / 100] ∗ weight

Usual weight prior to baseline study pregnancy and current weight were obtained, respectively, from the baseline and follow-up questionnaires. Percentage change in body fat, defined as the difference in body fat from baseline to follow-up divided by baseline body fat, was included as a linear variable.

Mother’s date of birth, a marker of secular trend in environmental DDE/PCB levels, was included as a continuous variable with unit in days (i.e. as a SAS date with reference date of January 1, 1960). Though baseline maternal age was another potential predictor, it was not included because it was highly correlated with date of birth (r=0.97).

3.4.3 Statistical analysis

For predictive modeling, linear regression models were fit using SAS (version 9.1). The statistical significance of each variable was assessed using the F test. Spearman

correlations comparing actual versus model-predicted concentrations at follow-up were used to assess the predictiveness of each model. To illustrate the model fits, model-predicted changes in organochlorine levels were plotted along with actual changes for the 122 individuals with non-missing questionnaire data.