Consistent with our earlier findings, the current analyses using an automated method for exposure assessment showed a modest association with breast cancer among highly exposed women. This method identified subjects with low levels of exposure among those previously considered unexposed and identified more highly exposed subjects among those previously considered to have low or moderate exposure levels. Based on comparison with historically measured samples, the automated exposure assessment appears to be more accurate than those derived from the earlier manual method. Thus, exposure misclassification in our prior analyses occurred mainly among subjects with low exposure levels and the current exposure distribution is shifted downward. It was not surprising that comparisons made using the percentile categories of RDD distribution shifted the elevated risks from women whose exposures were above the 75th percentile (in the prior analysis) to women whose exposures were above the 90th percentile (in the current reanalysis). Minimizing exposure misclassification among the unexposed most affected the referent group in the prior analysis, but the increased risk of breast cancer remained among more highly exposed women.
Given the current distribution of exposure, smoothing analyses were an important way to identify a meaningful cut point for "high" exposure. Similar or slightly stronger associations with breast cancer were seen when high exposure was defined at this new cutpoint (RDD > 35). Defining high exposure in the current analysis at an RDD of 35 was most comparable to the 90th percentile RDD in the prior analysis (Table ). While the majority of subjects classified above the 90th percentile in current analysis also had an RDD above 35 (85-100%, depending on latency), differences in RDD cut points may account for the attenuation of some of the 90th percentile odds ratios.
Because peak and duration of exposure are incorporated in the cumulative exposure measure, all exposure measures were highly correlated. In fact, the Spearman rank correlation coefficient between these measures ranged from 0.93 to 0.99 depending on latency period (p < 0.0001). Thus, while analyses of breast cancer risk using cumulative, peak, and duration of exposure were in good agreement, these analyses could not distinguish effects of intensity or duration of exposure from cumulative exposure.
Depending on the time interval between the installation of an ACVL pipe and the move-in date of a subject, varying the leaching rate should result in different exposure estimates. For example, a subject who moved into a location with ACVL pipe a few years after installation would be unexposed with a fast leaching rate (because the PCE would disappear quickly) or exposed with a slow leaching rate (because the PCE would mostly be remaining). Overall, we found that RDDs were smaller if the leaching rate was faster, and that RDDs were larger if the leaching rate was slower. However, varying the leaching rates did not affect the results of our epidemiological analyses when we used the 90th percentile to define the highest category of exposure because the exposures of the study population still maintained the same rank order.
The manual method's simplified flow estimation was a potential source of exposure misclassification in the prior breast cancer analyses, and we found this to be true for some residences when we compared manual estimates to those with the automated method. The manual method used the tools available at the time and simplified modeling water flow by addressing small sections of the distribution system piping around each residence. Using EPANET to incorporate complex conditions of water flow appears to predict PCE concentrations more accurately compared to measured water samples.
While incorporating EPANET in the exposure model addressed many complexities of the water distribution system that the Webler-Brown flow model could not, some of the same limitations remain. In EPANET, the modeled flow pattern and distribution system conditions were used to represent a wide range of time periods and water usage conditions, but the EPANET assessment still assumed the predicted steady-state flow pattern in the system was typical of any given time of the day, year, or season. Other studies have created more sophisticated models using information on current or historical conditions, such as tank levels, water account data, or pressure data recorded at hydrants, to characterize and validate the EPANET model [18
]. In the absence of reliable historical data for Cape Cod, we used EPANET to characterize water flow patterns to provide a reasonable method for ranking exposure for subjects in our epidemiological analyses.
In addition, the EPANET assessment did not determine if all land parcels had residences, and therefore water use, during the exposure period. An analysis of one study town (Mashpee) found that limiting water use to parcels occupied in 1980 reduced the number of users by one-half, which changed magnitude of exposure, but did not significantly affect exposure categories because water flow direction remained similar. Nevertheless, other towns with different patterns of development may have different results. Residential build year information could be used to see if improving this aspect of the EPANET model would provide still better exposure information.
Predicting detectable concentrations using measured samples from 1980 was improved with the automated method. Again, this improvement was most evident in areas where applying the manual method was difficult, including complicated configurations, at the middle or initial segment of a pipe, and in higher flow areas. The modeled estimates using the automated method, in addition to incorporating more physical conditions and principles, are better correlated with the water samples from 1980 (p = 0.65 vs. 0.54).
It is important to recognize that there were also likely inaccuracies in the measured concentrations. The water samples were collected in the 1980s to obtain a rough determination of the scope of the PCE leaching problem and begin remediation [5
]. Areas with ACVL pipe that were anticipated to have high levels like low flow, dead-end pipes, were preferentially sampled. In the absence of written protocols it is possible there were inconsistencies in water sampling. Aeration from hydrant sampling may have introduced errors depending on sampling procedure and head space in hydrant lines may have caused loss of PCE due to volatilization. Thus, many samples below the detection limit may have been "false negatives," particularly in sample locations in low flow areas near recently installed pipe. Several towns had only one or two samples with detectable levels, suggestive of measurement error, although there are circumstances (e.g. high flow, older pipe) where levels below the detection limit would be expected. Variation in pipe drying times and initial PCE concentrations may have contributed to the low levels.
Testing by the DEP suggested the laboratory's use of head space analysis could also have underestimated PCE levels in the water samples by as much as 80% [13
]. A DEP memorandum indicated that this methodology was considered a qualitative, not quantitative, method that required less analysis time than the more accurate purge and trap method and so allowed for a rapid response and remediation [13
]. Up to two-fold fluctuations in water concentrations were observed in a sampling study that measured concentrations at the same location and time on consecutive days [14
]. Wacholder et al. have used the term "alloyed gold standard" to describe these type of data [22
], although the addition of even more sources of measurement error (such as sampling personnel and season) suggest that the historical measurements are not a standard at all but just another view of the data.
As discussed in our prior publication on breast cancer and PCE exposure, this reanalysis is unlikely to be affected by selection bias. The Massachusetts Cancer Registry was the source of all breast cancer cases, and had nearly complete reporting according to a Department of Public Health comparison with other state cancer registries [4
]. Demographic characteristics, follow-up and interview rates were similar among cases and controls, and demographic characteristics were similar among participants and non-participants [4
]. While interviewers were not blinded to a woman's disease status, observation bias was also unlikely to affect the results. The closed-ended questions were carefully written and pre-tested, and interviewers were trained in appropriate interviewing techniques. Also, proxy interviews for deceased cases and controls resulted in comparable information quality [4
]. Lastly, the exposure assessments using EPANET were conducted without knowledge of the participant's disease status.
Core confounding variables of age at diagnosis or index year, vital status at interview, family history of breast cancer, personal history of prior breast cancer, age at first live birth or stillbirth, and occupational exposure to PCE were controlled in adjusted analyses. Other potential confounders did not change the core-adjusted ORs by more than 10%, so they were not included in the models. Unmeasured factors, environmental or otherwise, may have resulted in residual confounding, but this is an unlikely explanation for these findings because these factors would need to be strong risk factors for breast cancer and tightly correlated with PCE exposure. The latter is unlikely given the irregular pattern of the ACVL pipe locations.
Animal and epidemiological studies have suggested that PCE exposure is associated with several types of cancer; but in general null effects have been found for breast cancer. Experiments have shown increases in the incidence of liver tumors in mice exposed to PCE orally or by inhalation, and increases in incidence of leukemia and kidney cancer in rats with inhalation exposure [2
]. There has been no evidence of mammary tumors stemming from PCE exposure in animal assays, although other organic solvents have shown this effect [7
]. Epidemiological evidence has been provided primarily by occupational studies of dry cleaning workers and people working with solvents in metal industries. Exposure assessments in these studies are difficult, relying on job title or duration of work to define exposure, and so potentially misclassify subjects' exposure. In many studies, there were multiple chemical exposures occurring at the same time, relatively small numbers of women, and missing information on potential confounders. Reviews of the overall weight of evidence, however, prompted the International Agency for Cancer Research (IARC) to classify PCE as a probable human carcinogen and the National Toxicology Program (NTP) to classify PCE as reasonably anticipated to be carcinogenic to humans [23
Studies of the association between PCE and breast cancer incidence have produced inconsistent results. An 11% decreased incidence of breast cancer was seen in a large Scandinavian cohort (n = 23,714), but exposure to PCE was uncertain among this group of laundry and dry cleaning workers [26
]. Compared to women in the general population, laundry and dry cleaning workers in a U.S.-Canada population-based study had a lower incidence of breast cancer (mean annual age-standardized rates were, 77.4 vs. 100.3 per 100,000 person-years) [27
]. A case-control study in British Columbia examined post-menopausal women whose usual occupation was in dry cleaning industry and found a 4.9-fold increased risk of breast cancer (95%CI 1.3-18.7) after controlling for important risk factors such as family history of breast cancer [28
]. Chemically similar to PCE, studies on the solvent trichloroethylene (TCE) and breast cancer show similarly mixed results.
A study of Finnish workers found a decreased risk of breast cancer for women who were biologically monitored for exposure to TCE, PCE and another halogenated hydrocarbon trichloroethane (SIR = 0.84, 95% CI, 0.44-1.48) [29
]. A study in Taiwan of electronics factory workers exposed to chlorinated solvents that likely included PCE and TCE found a small but significantly elevated incidence of breast cancer (SIR = 1.19, 95% CI, 1.03-1.36) after adjusting for age and calendar year [30
Our current analyses made use of technological advances to investigate a limitation of our earlier exposure assessment method and potential source of misclassification. We used GIS software in conjunction with a modification of the open source water distribution modeling software, EPANET, to create a more detailed exposure model that could easily be manipulated for sensitivity analyses and other improvements. EPANET has been used in other epidemiological studies to conduct exposure assessments of distribution system contamination, including simulations to study the extent and severity of source contamination, as well as assess water treatment and system issues such as trihalomethane exposure [18
]. This unique application of EPANET software was made possible by availability of its open source code, which could be adapted for our use, and by GIS software which provided the necessary platform to bring together the different types of data needed to improve the drinking water flow model. In this study, GIS software made it possible to create detailed maps with pipe characteristics, land parcels to represent water users, and participants' residential locations. These maps formed the basis for the EPANET schematics used to simulate water flow and the dispersion of PCE in the exposure model. With this automated method, we were also able to perform sensitivity analyses on the leaching rate.