Archives

  • 2018-07
  • 2019-07
  • 2019-08
  • 2019-09
  • 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • 2021-03
  • br Multipollutant analysis br We used Classification and Reg

    2020-07-06


    2.6.2. Multipollutant analysis
    We used Classification and Regression Tree (CART) methods to explore whether there are combinations of pollutants that may be more or less harmful for breast cancer than would be expected based on exposure to only a single pollutant in isolation, while also addressing co-pollutant confounding (Breiman et al., 1984; Lemon et al., 2003). CART is a forward-selection approach that selects, at each step, the variable with the strongest association with the outcome and performs a binary split according to that variable. Further partitions occur within the previous partitions, which forms “branches” that characterize the expected risk in strata of the variables on the branch. Risk is estimated for each individual by taking the average outcome among the partici-pants in the end of each branch in which that participant is partitioned. 
    We included the 29 air toxics, age at baseline, BMI, and physical ac-tivity in the classification trees. Age, BMI, and physical activity were included to parallel the single pollutant analyses. The most common splitting criterion for classification trees, and the one used in this analysis, is the Gini index (Lemon et al., 2003). For the stopping cri-teria, we specified that the tree grow to a maximum depth of five levels and that the minimum number of total observations in a Relebactam be 12 (leading to at least five cases per node). Cost-complexity pruning (Breiman et al., 1984) was used to create a tree with a total of 11 terminal nodes. These specifications were used to control the size of the tree so that it wouldn't be too large and lose interpretability, while still identifying meaningful groups. The final tree was chosen by balancing model fit and interpretability across combinations of maximum depth, node size, and number of terminal nodes. CART was conducted using SAS function PROC HPSPLIT.
    3. Results
    3.1. Population characteristics
    In our study population, the mean census tract estimated air toxic concentrations ranged from 2.30 × 10−7 μg/m3 for benzidine to 2.21 μg/m3 for toluene (Table 1). Quintile cutpoints for each air toxic are also provided in Table 1.
    3.2. Results from single pollutant models
    Methylene chloride was associated with an increased risk of overall (HRquintile 4vs1 = 1.21; 95% CI = 1.07–1.38) (Table 2) and ER+ in-vasive breast cancer (HRquintile 4vs1 = 1.28; 95% CI = 1.08–1.52) (Table 3). However, the associations were non-monotonic. For ER+ breast cancer, HRs were elevated for all quintile-based categories of exposure relative to the lowest category. The positive association with methylene chloride was also observed for postmenopausal breast cancer, and was maintained across sensitivity analyses when restricting, separately, to invasive cases, non-Hispanic whites, those who enrolled in 2005 or later, cases confirmed by medical record, and when adjusted Relebactam for region (data not shown). Methylene chloride was also identified as a variable of importance in the classification tree (Fig. 1). The combi-nation of these results indicates that, among the 29 air toxics we con-sidered, methylene chloride was most consistently associated with elevated breast cancer risk.
    Table 1 Mean concentrations and quintile cutpoints of the individual hazardous air pollutants (μg/m3) among the 49,718 Sister Study women in hair bulb study.
    Air toxic Mean Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5
    a Air toxic categorized with a concentration of (9.9 × 10−9 for benzidine) as the referent and remaining values split based on tertiles.
    Table 2
    HRs (95% CIs) for the associationsabetween individual hazardous air pollutants and overall breast cancer risk, the Sister Study.
    Air toxic Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5 p-Trend
    a Models adjusted for race, residence type, education, and smoking status.
    b Air toxic categorized with a concentration of (9.9 × 10−9 for benzidine) as the referent and remaining values split based on tertiles.
    Table 3
    HRs (95% CIs) for the associationsabetween hazardous air pollutants and ER+ invasive breast cancerb, the Sister Study.
    Air toxic Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5 p-Trend
    a Models adjusted for race, residence type, education, and smoking status.
    c Air toxic categorized with a concentration of (9.9 × 10−9 for benzidine) as the referent and remaining values split based on tertiles.
    Fig. 1. Classification tree of hazardous air pollutants and breast cancer, the Sister Study.
    categories above the referent.
    When we considered single pollutant associations stratified by me-nopausal status, the results for postmenopausal breast cancer were si-milar to those for overall breast cancer, while many, but not all, of those for premenopausal breast cancer were slightly attenuated and less precise (data not shown). Tests of heterogeneity revealed no differences by menopausal status.