Among the full cohort of 5,752 women, the mean ± SD duration of employment was 7.7 ± 4.5 years. A total of 1,413 (25%) had died by the study end date. We obtained completed questionnaires for 3,952 (69%) of the cohort, 80% from living respondents and 20% from next of kin of deceased cohort members. The primary reason for nonresponse was our inability to determine a correct address (21%). A total of 500 former workers and proxy respondents (8.7%) refused to complete the questionnaire, and 112 (2%) failed to respond after repeated attempts.
We found 281 incident breast cancer cases in the full cohort. Although mortality follow-up of the cohort was exhaustive, nonfatal incident cases occurring before initiation of the cancer registries or occurring in states other than Massachusetts, New York, Indiana, Florida, and California could be ascertained only via questionnaire. Approximately 20% of the cohort did not return a questionnaire and were not known to be deceased at end of follow-up; incident cases occurring outside the registry search may have been missed in this group, particularly in the early years of follow-up.
Of the full cohort, we identified 147 women by plant records and death certificates as being of races other than white. When we used the questionnaire responses instead, the number of workers (both cases and noncases) identified as being of races other than white rose to 282. This group includes workers identifying as members of specific races other than white (black, American Indian/Alaskan Native, Asian, or Pacific Islander), women who identified only as “other,” and women identifying as multiracial. Because the number of workers identifying as other than white was small, we kept this group together for all further analyses and refer to it herein as “nonwhite,” although some women in fact selected a multiracial identity where one of the races was white.
Of the 281 cases in the full cohort, we identified only eight as nonwhite from plant or death certificate data. However, among workers with complete questionnaire data on the covariates of a priori interest (restricted cohort), we identified 14 of the 145 cases as nonwhite. Twelve of these were from plant 2. Of the 268 noncase workers identified as nonwhite on the questionnaire, 67 identified as African American, and 201 identified as “other.” Most workers (n = 150) in the last group did not further specify their race/ethnicity; of those who did, 90% (n = 55) identified themselves solely as Cape Verdean or as Cape Verdean and some other ethnicity.
provides selected demographic and exposure characteristics of the breast cancer cases and noncases in the entire cohort (n = 5,752) and the questionnaire subcohort (n = 3,952). In the entire cohort, on average, cases (n = 281) were born earlier than noncases and were exposed longer; these differences were highly statistically significant. Cases also had higher cumulative exposures. Breast cancer in situ was reported for 8 of the 281 cases (2.8%).
Characteristics of breast cancer cases and noncases.
More data were available for the questionnaire subcohort, including demographic and lifestyle characteristics. Again, cases (n = 201) were born significantly earlier than noncases and were exposed longer. Cases had higher mean cumulative exposure than did noncases, but the difference was not statistically significant. Cases were significantly more likely to have a first-degree relative with breast cancer than were noncases. Average age at first live birth showed a slight but statistically significant difference. There were no statistically significant differences between case and noncases in the average age at menarche, body mass index (BMI), parity, number of children, use of hormone replacement therapy, age at first use of hormone therapy, or in the proportion of nonwhites or ever-smokers.
The plants differed in a number of ways. Most striking were the contrasts in exposure distributions. Mean ± SD estimated exposure for the workforce of plant 2 (0.80 ± 1.06, 1,000 unit-years) was more than twice that in plant 1 (0.34 ± 0.71) and was an order of magnitude greater than that in plant 3 (0.078 ± 0.10). The medians reflected a similar pattern (plant 2 = 0.39, plant 1 = 0.11, and plant 3 = 0.04). Although all three facilities had nonwhite populations < 4% according to records-based data, we classified > 15% of workers at plant 2 as nonwhite using questionnaire data. The percentages of nonwhite workers for plants 1 and 3 increased only slightly when we used questionnaire data.
In the aggregate, white and nonwhite workers in the restricted subcohort differed on a number of demographic and lifestyle factors (). Statistically significant differences included higher mean and median cumulative exposures, fewer pack-years of smoking, greater number of live births, and a greater percentage with a first-degree female relative with breast cancer among nonwhite workers. Mean exposure durations for the two groups were similar. We saw more striking differences among cases: The median exposure for nonwhite cases was nearly double the 75th percentile of exposure for white cases; in contrast, for noncases, the mean was only slightly higher in nonwhites and the median was actually slightly higher in whites.
Selected characteristics of restricted cohorta by race.
SIR and SRR results reflect follow-up time and cases occurring in 1970 or later, because of availability of SEER comparison data. For the full cohort, the unlagged SIR was 0.81 [95% confidence interval (CI), 0.72–0.92; n = 257] and using a 10-year lag was 0.81 (95% CI, 0.71–0.92; n = 251 cases). SRRs by exposure level showed no trend using lagged or unlagged data. However, the SIRs and SRRs differed somewhat by race. For women identified as white by plant records or death certificates, the SIRs were 0.80 (95% CI, 0.70–0.90; n = 250) and 0.80 (95% CI, 0.70–0.90; n = 244) with 0- and 10-year lags, respectively. For women identified as nonwhite by records (all facilities), the SIRs had nonsignificant elevations of 1.87 (95% CI, 0.75–3.84; n = 7) and 1.94 (95% CI, 0.77–3.99; n = 7) for 0- and 10-year lags respectively. SIRs for women with questionnaire data were similar to those for the full cohort in white women (unlagged SIR = 0.84; 95% CI, 0.72–0.97; n = 176; 10-year lag SIR = 0.84; 95% CI, 0.72–0.97; n = 172) but showed a more modest elevation in nonwhite women (unlagged SIR = 1.14; 95% CI, 0.59–1.99; n = 12; 10-year lag SIR = 1.17; 95% CI, 0.06–2.03; n = 12) than that seen in the full cohort.
SRR results differed depending on dose cut points and lag (data not shown), particularly for white women, with no consistent patterns observed. In nonwhite women, trends were always positive and sometimes statistically significant, depending on cut points and lag. However, elevations were limited to the highest dose categories, with deficits observed in some intermediate categories.
shows results of main effects exposure–response analyses via Cox regression using the full, questionnaire, and restricted cohorts. To account for secular trends in breast cancer incidence, all models included variables for birth cohort (before 1920 as a referent, 1920–1934, and ≥ 1935). Although the exposure metrics were right skewed, the untransformed metrics fit the data best. For the full cohort, neither of the exposure metrics achieved statistical significance. Cumulative exposure was not associated with elevations in breast cancer incidence for the full cohort, the questionnaire subcohort, or the restricted subcohort; exposure duration showed a statistically significant association with risk only in the restricted cohort. Results for external and internal analyses did not differ greatly with the inclusion or exclusion of in situ cases.
Main effects exposure–response results for breast cancer incidence with a 10-year lag by subcohort: hazard ratio (95% CI).
Because the SRR results suggested a difference in risk by race, we evaluated the effects of race in the regression analyses and then conducted further evaluations of potential confounders and effect modifiers in the restricted cohort separately for nonwhite and white women. provides the results of these analyses. Although the unlagged model fit best for most exposure metrics (cumulative, duration) in white women, a 10-year lag fit better in nonwhite women. We had doubts about the biologic plausibility of a zero lag and chose the 10-year lag for further modeling.
Full models for breast cancer incidence with a 10-year lag, restricted subcohorta by race.
Among white women, cumulative exposure and duration of exposure had little effect on breast cancer risk. Birth cohort was significant, with an increased risk for the group born during the period 1920–1934 and a larger increase compared with baseline for those born after 1934. Beyond birth cohort, the covariates retained in the model were parity, family history of breast cancer, and self- versus proxy questionnaire completion. Plant did not strongly affect the relation between the exposure metrics and breast cancer risk, and we eliminated this variable from the model. Use of hormone replacement therapy, although technically a confounder, had a lesser effect on risk, and we also eliminated this variable from the final model in the interest of parsimony. Menopausal status was not a confounder in this group.
In contrast, among nonwhite women, the effects of increasing exposure were positive and statistically significant. Both cumulative exposure (continuous) and duration of exposure were highly associated with breast cancer risk. Categorical exposure was associated with elevated risk at the two highest levels, but not in the intermediate category. The variables usually considered breast cancer risk factors (family history, parity, use of hormone replacement therapy, menopausal status) had little effect on risk in this group (data not shown). The most important covariates for this cohort were ever-smoking, source of questionnaire data, and birth cohort (limited to those women born after 1934). In univariate analyses, nonwhite cases smoked more than noncases (46 vs. 29 pack-years), in part because they smoked for more years (42.6 vs. 32.9). We did not observe these patterns among white workers. To assess whether smoking might be a proxy for alcohol use, we examined models with smoking but not alcohol, with alcohol but not smoking, and with both alcohol and smoking and found that only smoking had significant explanatory value.
To further explore the effects of smoking in nonwhite workers, we considered current smoking status (at time of death for deceased workers or time of questionnaire completion for living workers) and pack-years. Adding pack-years or current smoking status to a model that included ever smoking had little effect. The interaction between current smoking status and cumulative exposure nearly attained statistical significance (p = 0.06).
Because the mean and median PCB exposure estimates were significantly higher among nonwhite workers, we considered the possibility that the greater risk observed in nonwhite women was attributable to a high-exposure effect. We reevaluated the exposure risk in white women using the same cut points employed for nonwhite women and found the same lack of trend we observed with the original quartile-based cut points. The lack of exposure–response in white women does not appear to be related to the lower exposure levels in this group.