Archived Comments for:
Risk of breast cancer following exposure to tetrachloroethylene-contaminated drinking water in Cape Cod, Massachusetts: reanalysis of a case-control study using a modified exposure assessment
Comment on the paper by Gallagher et al.: Risk of breast cancer following exposure to tetrachloroethylene-contaminated drinking water in Cape Cod, Massachusetts: reanalysis of a case-control study using a modified exposure assessment.
John Bukowski, WordsWorld Consulting
19 October 2011
In this most recent iteration of the Cape Cod perchloroethylene (PCE) study, Gallagher et al. [1] have attempted to improve the exposure assessment used in the previous breast cancer articles [2,3]. However, these authors are still left with the same problem, trying to tease out relatively weak effects from residential exposure, when much higher occupational and laboratory exposures have failed to demonstrate them.
The results reported by Gallagher et al. are in conflict with those of occupational studies in which women were exposed to much higher levels. Assuming that the relative delivered dose (RDD) from the current paper is in grams, women on Cape Cod experienced median cumulative (ie, total) exposures of approximately 2 grams PCE [1]. Using calculations from my recent review on the epidemiology surrounding residential PCE, average occupational exposure from dry cleaning in the 1970s (which is germane to the occupational studies published in the literature) is estimated at 100-700 mg per day [4]. Assuming a 200-day work-year, dry cleaning had exposed working women to 20-140 grams PCE per year, or 400-2800 grams during a 20-year working life. Yet, Gallaher et al. acknowledge that ¿in general, null effects have been found for breast cancer¿ in this literature. In their discussion, these authors cite several isolated studies as support for an epidemiologic link between PCE and breast cancer [1]. Yet, neither authoritative reviews [5-9] nor large cohort studies [10,11] have concluded that there is any relationship, despite the aforementioned several orders of magnitude greater exposure for dry cleaners compared with women in Cape Cod. Gallagher et al. also acknowledge that laboratory studies fail to show a relationship between high-level PCE exposure and mammary cancer [1].
The results of the current study are also internally inconsistent. Gallagher et al. state that there was ¿no increase in the odds ratio until 17 and 19 years of latency.¿ But, their adjusted ORs of 1.3-1.4 for the lower latency categories are essentially the same as for those with 11 or more latent years (OR 1.3-1.5), especially given the statistical variability suggested by the confidence intervals [1]. Such findings are inconsistent with the known latency for breast cancer, in which a median latency as long as 22 years has been proposed [12]. Furthermore, none of the ORs reported by Gallagher et al. are statistically significant [1].
Gallagher et al. similarly provide no evidence of any monotonic exposure-response trend that would support a causal relationship. There is no increased risk for ever vs. never exposure, and the risks for those above either the median or 75th percentile are essentially the same as for those below the median (0.9-1.2), especially when statistical variability is considered [1]. This lack of trend would be even more pronounced if exposure groups were appropriately defined as discrete categories of increasing exposure such as quartiles or deciles. Instead, Gallagher et al. provide exposure categories nested within those below them. That is to say, those above 50% include those above 75%, which in turn include those above 90%. This sliding scale hides irregularities in exposure-response that would further detract from causal confidence. For example, adjusted ORs for those in the highest decile are 1.3-1.5, compared with 0.9-1.1 for those above the 50th or 75th percentiles. The high-risk women above the 90th percentile are included within these lower categories, so the risks from lower exposure would be greatly reduced if those with highest exposure were removed. For this reason, more appropriate categories such as 50%-75% and 75%-90% would most likely have shown OR below 1.0 (ie, protective), which is highly inconsistent with any positive trend from increasing exposure.
Gallagher et al. suggest that they have enhanced the previous Cape Cod breast cancer methodology by refining the exposure metric to reduce misclassification [1]. Although the original misclassification purportedly had a greater impact on those with lower exposure, it was still ostensibly nondifferential, meaning similarly distributed among cases and controls. Such misclassification should bias toward the null, so that improved estimates tend to elevate risks. Yet, in the current situation, most revised OR have been driven downward, including those among the highest exposure decile who were least impacted by the misclassification. Most noteworthy are risks among those with exposure below the median, for whom OR of 1.2-2.1 (from the earlier study iterations) have been completely eliminated by ostensibly decreasing nondifferential exposure misclassification (see table 5, ref. [1]). Given that such findings go against expectation, one cannot help but wonder if some degree of ¿data shopping¿ has been carried out in order to explain away these previous contrary findings.
Gallagher et al. acknowledge that there may have been residual confounding, but discount that as an explanation for results given that ¿the irregular pattern of the ACVL pipe locations¿ would make a differential association with exposure unlikely [1]. Yet, the core confounders must have been differentially distributed by exposure, because adjustment for them decreased OR by as much as 50% (see table 4, ref. [1]). Although not yet proven risk factors, other potential confounders such as bone density, use of estrogen replacement therapy, alcohol intake, diet, and obesity have relative risks similar to those associated with some of the core confounders adjusted for in the current study [13-15]. A model-building process with adjustment for a suite of these other factors (rather than one at a time) may have lowered ORs even further. Also, given that confounding covariates (like exposures) are measured imprecisely, residual confounding from risk-factor misclassification is possible even after adjustment [16, 17]. Therefore, residual confounding should be considered as at least a partial explanation for the reported results.
In conclusion, Gallagher et al. have shown different results using the EPA automated model, but have not proven that this is a superior method to the previous manual method. There was a slightly better correlation between measured and modeled values using the automated method (r=0.65) compared with the manual one (r=0.54). But as the authors point out, the measured values ¿are not a standard at all but just another view of the data¿ [1]. Therefore, given the lack of any gold standard by which to judge that the different results produced by Gallagher et al. are superior (ie, closer to the truth) to those produced previously, the current paper appears to be an academic exercise that adds little useful information to the literature on breast cancer causation. Furthermore, the information that it does provide fails at least four of Hill¿s criteria for causation. Namely, the association is weak, inconsistent with the existing human and experimental evidence, and lacking an exposure-response trend [18]. Given these limitations, the results warrant cautious interpretation and do not support a causal relationship between breast cancer and PCE exposure.
References
1. Gallagher LG, Vieiria VM, Ozonoff D, Webster TF, Aschengrau A: Risk of breast cancer following exposure to tetrachloroethylene-contaminated drinking water in Cape Cod, Massachusetts: reanalysis of a case-control study using a modified exposure assessment. Environ Health 2011, 10:47 Available from: http://www.ehjournal.net/content/10/1/47.
2. Aschengrau A, Paulu C, Ozonoff D: Tetrachloroethylenecontaminated drinking water and the risk of breast cancer. Environ Health Perspect 1998, 106 (Suppl 4):947¿953.
3. Aschengrau A, Rogers S, Ozonoff D: Perchloroethylenecontaminated drinking water and the risk of breast cancer: additional results from Cape Cod, Massachusetts, USA. Environ Health Perspect 2003, 111:167¿173.
4. Bukowsk JA: Review of the epidemiologic literature on residential exposure
to perchloroethylene. Crit Rev Toxicol 2011: 1¿12. Available early online from: http://informahealthcare.com/doi/abs/10.3109/10408444.2011.581649.
5. Weiss NS: Cancer in relation to occupational exposure to perchloroethylene. Cancer Causes Control 1995, 6:257-266
6. Lynge E, Anttila A, Hemminki K: Organic solvents and cancer. Cancer Causes Control 1997, 8:406-419
7. Institutes of Medicine: Gulf war and health. Vol. 2: Insecticides and solvents. Washington, DC: National Academies Press; 2002.
8. Mundt KA, Birk T, Burch MT: Critical review of the epidemiological literature on occupational exposure to perchloroethylene and cancer. Int Arch Occup Environ Health 2003, 76:473-491.
9. Ruder AM: Potential health effects of occupational chlorinated solvent exposure. Ann NY Acad Sci 2006, 1076: 207¿227.
10. Boice JD, Marano DE, Fryzek JP, Sadler CJ, McGlaughlin JK: Mortality among aircraft manufacturing workers. Occup Environ Med 1999, 56:581-597.
11. Blair A, Petralia SA, Stewart PA: Extended mortality follow-up of a cohort of dry cleaners. Ann Epidemiol 2003, 13:50-56.
12. Olsson H, Baldetorp B, Ferno M, Perfekt R: Relation between the rate of tumour cell proliferation and latency time in radiation associated breast cancer. BMC Cancer 2003, 3:11. Available from: http://www.biomedcentral.com/1471-2407/3/11.
13. McPherson K, Steel CM, Dixon JM: Breast cancer-epidemiology, risk factors, and genetics. BMJ 2000, 321:624-628.
14. McTiernan A: Behavioral risk factors in breast cancer: Can risk be modified? Oncologist 2003, 8:326-334.
15. Santen RJ, Boyd NF, Chlebowski RT, Cummings S, Cuzick J, Dowsett M, Easton D, Forbes JF, Key T, Hankinson SE, Howell A, Ingle J: Critical assessment of new risk factors for breast cancer: considerations for development of an improved risk prediction model. Endocrine-Related Cancer 2007, 14:169-187.
16. Greenland S: The effect of misclassification in the presence of covariates. Am J Epidemiol 1980, 112:564-569.
17. Marshall JR, Hastrup JL: Mismeasurement and the resonance of strong confounders: Uncorrelated errors. Am J Epidemiol 1996, 143:1069-1078.
18. Hill AB: The environment and disease: association or causation? Proc R Soc Med 1965, 58:295-300.
Competing interests
Funding for this comment was provided by the Halogenated Solvents Industry Alliance.
Authors' Response by Lisa G. Gallagher, Veronica M. Vieira, David M. Ozonoff, Thomas F. Webster and Ann Aschengrau
Ann Aschengrau, Boston University
10 November 2011
Dr. Bukowski, writing at the request of the Halogenated Solvents Industry Association (HSIA), calls into question our results on the grounds that they conflict with occupational studies he alleges show no increased risk of breast cancer at much higher PCE exposures. We understand why the HSIA would want to weigh in on this question because it might suggest that their product, PCE, which is in widespread use and causes extensive exposure in the occupational and general community environment, is an unreasonably dangerous product. Given the size of the exposed population, even relatively small risks could result in an unacceptable breast cancer burden on society.
The majority of the occupational studies he cites examined breast cancer mortality, not incidence, as the outcome, and, as such, did not assess associations with the entire spectrum of this, often non-fatal, disease. These studies also had little adjustment for confounding factors and are confounded by socioeconomic status (SES), an important risk factor for breast cancer. This source of confounding would tend to bias the results of the occupational studies towards the null because low SES women tend to be employed in PCE-exposed occupations while high SES women have an increased risk of breast cancer. In contrast, our study considered and controlled for many potential confounding variables, including a woman's educational level. In our general population study subjects we found little evidence of confounding by SES because the irregular pattern of PCE contamination on Cape Cod often resulted in vastly different exposure levels for adjacent residents and neighborhoods. In addition, the different exposure routes in the occupational studies (e.g. mainly inhalation and dermal) may contribute to the alleged divergent findings between our study and this body of literature.
There are also several statements made by Dr. Bukowski that are incorrect. First, the author misinterpreted our exposure measure as an actual mass value in grams. As noted in our paper, our cumulative measure of PCE exposure was not a mass (a ratio measure), but an ordinal measure used to rank our subjects. In addition, the author's statement that our study results are not internally consistent inaccurately represents our data by mixing results for any exposure with those for exposure levels. Our paper noted no increases in the crude odds ratios until 17 and 19 years latency (ORs 1.3-1.4) among ever-exposed women and stated that adjusted odds ratios among ever-exposed women were null for all latent periods.
Each latency analysis also used a latency-specific exposure distribution to determine percentile cut points. Thus, varying results across latent periods may reflect different cut points (paper Table 2). The cut point for current smoothing analysis (RDD>35) is most comparable to the 90th percentile cut point used in the prior analysis. In addition, some odds ratios for the current 90th percentile may be attenuated because of the lower cut points. In general, the reduction of exposure misclassification resulted in assigning subjects who were unexposed in the prior analysis to low exposure in the current analysis. This change resulted in a cleaner referent population and was expected to strengthen the associations. However, the low exposures determined by the automated method were associated with little or no increased risk.
In summary, we maintain that the new exposure assessment method described in this paper is improved from the original method, as evidenced by an improved correlation with sampling data and reflecting a more accurate estimate of water flow in the piping network. An important contribution of this analysis was to demonstrate that the associations between breast cancer and PCE-contaminated drinking water are relatively robust to refinements in exposure modeling.
Environmental Health
Comment on the paper by Gallagher et al.: Risk of breast cancer following exposure to tetrachloroethylene-contaminated drinking water in Cape Cod, Massachusetts: reanalysis of a case-control study using a modified exposure assessment.
19 October 2011
In this most recent iteration of the Cape Cod perchloroethylene (PCE) study, Gallagher et al. [1] have attempted to improve the exposure assessment used in the previous breast cancer articles [2,3]. However, these authors are still left with the same problem, trying to tease out relatively weak effects from residential exposure, when much higher occupational and laboratory exposures have failed to demonstrate them.
The results reported by Gallagher et al. are in conflict with those of occupational studies in which women were exposed to much higher levels. Assuming that the relative delivered dose (RDD) from the current paper is in grams, women on Cape Cod experienced median cumulative (ie, total) exposures of approximately 2 grams PCE [1]. Using calculations from my recent review on the epidemiology surrounding residential PCE, average occupational exposure from dry cleaning in the 1970s (which is germane to the occupational studies published in the literature) is estimated at 100-700 mg per day [4]. Assuming a 200-day work-year, dry cleaning had exposed working women to 20-140 grams PCE per year, or 400-2800 grams during a 20-year working life. Yet, Gallaher et al. acknowledge that ¿in general, null effects have been found for breast cancer¿ in this literature. In their discussion, these authors cite several isolated studies as support for an epidemiologic link between PCE and breast cancer [1]. Yet, neither authoritative reviews [5-9] nor large cohort studies [10,11] have concluded that there is any relationship, despite the aforementioned several orders of magnitude greater exposure for dry cleaners compared with women in Cape Cod. Gallagher et al. also acknowledge that laboratory studies fail to show a relationship between high-level PCE exposure and mammary cancer [1].
The results of the current study are also internally inconsistent. Gallagher et al. state that there was ¿no increase in the odds ratio until 17 and 19 years of latency.¿ But, their adjusted ORs of 1.3-1.4 for the lower latency categories are essentially the same as for those with 11 or more latent years (OR 1.3-1.5), especially given the statistical variability suggested by the confidence intervals [1]. Such findings are inconsistent with the known latency for breast cancer, in which a median latency as long as 22 years has been proposed [12]. Furthermore, none of the ORs reported by Gallagher et al. are statistically significant [1].
Gallagher et al. similarly provide no evidence of any monotonic exposure-response trend that would support a causal relationship. There is no increased risk for ever vs. never exposure, and the risks for those above either the median or 75th percentile are essentially the same as for those below the median (0.9-1.2), especially when statistical variability is considered [1]. This lack of trend would be even more pronounced if exposure groups were appropriately defined as discrete categories of increasing exposure such as quartiles or deciles. Instead, Gallagher et al. provide exposure categories nested within those below them. That is to say, those above 50% include those above 75%, which in turn include those above 90%. This sliding scale hides irregularities in exposure-response that would further detract from causal confidence. For example, adjusted ORs for those in the highest decile are 1.3-1.5, compared with 0.9-1.1 for those above the 50th or 75th percentiles. The high-risk women above the 90th percentile are included within these lower categories, so the risks from lower exposure would be greatly reduced if those with highest exposure were removed. For this reason, more appropriate categories such as 50%-75% and 75%-90% would most likely have shown OR below 1.0 (ie, protective), which is highly inconsistent with any positive trend from increasing exposure.
Gallagher et al. suggest that they have enhanced the previous Cape Cod breast cancer methodology by refining the exposure metric to reduce misclassification [1]. Although the original misclassification purportedly had a greater impact on those with lower exposure, it was still ostensibly nondifferential, meaning similarly distributed among cases and controls. Such misclassification should bias toward the null, so that improved estimates tend to elevate risks. Yet, in the current situation, most revised OR have been driven downward, including those among the highest exposure decile who were least impacted by the misclassification. Most noteworthy are risks among those with exposure below the median, for whom OR of 1.2-2.1 (from the earlier study iterations) have been completely eliminated by ostensibly decreasing nondifferential exposure misclassification (see table 5, ref. [1]). Given that such findings go against expectation, one cannot help but wonder if some degree of ¿data shopping¿ has been carried out in order to explain away these previous contrary findings.
Gallagher et al. acknowledge that there may have been residual confounding, but discount that as an explanation for results given that ¿the irregular pattern of the ACVL pipe locations¿ would make a differential association with exposure unlikely [1]. Yet, the core confounders must have been differentially distributed by exposure, because adjustment for them decreased OR by as much as 50% (see table 4, ref. [1]). Although not yet proven risk factors, other potential confounders such as bone density, use of estrogen replacement therapy, alcohol intake, diet, and obesity have relative risks similar to those associated with some of the core confounders adjusted for in the current study [13-15]. A model-building process with adjustment for a suite of these other factors (rather than one at a time) may have lowered ORs even further. Also, given that confounding covariates (like exposures) are measured imprecisely, residual confounding from risk-factor misclassification is possible even after adjustment [16, 17]. Therefore, residual confounding should be considered as at least a partial explanation for the reported results.
In conclusion, Gallagher et al. have shown different results using the EPA automated model, but have not proven that this is a superior method to the previous manual method. There was a slightly better correlation between measured and modeled values using the automated method (r=0.65) compared with the manual one (r=0.54). But as the authors point out, the measured values ¿are not a standard at all but just another view of the data¿ [1]. Therefore, given the lack of any gold standard by which to judge that the different results produced by Gallagher et al. are superior (ie, closer to the truth) to those produced previously, the current paper appears to be an academic exercise that adds little useful information to the literature on breast cancer causation. Furthermore, the information that it does provide fails at least four of Hill¿s criteria for causation. Namely, the association is weak, inconsistent with the existing human and experimental evidence, and lacking an exposure-response trend [18]. Given these limitations, the results warrant cautious interpretation and do not support a causal relationship between breast cancer and PCE exposure.
References
1. Gallagher LG, Vieiria VM, Ozonoff D, Webster TF, Aschengrau A: Risk of breast cancer following exposure to tetrachloroethylene-contaminated drinking water in Cape Cod, Massachusetts: reanalysis of a case-control study using a modified exposure assessment. Environ Health 2011, 10:47 Available from: http://www.ehjournal.net/content/10/1/47.
2. Aschengrau A, Paulu C, Ozonoff D: Tetrachloroethylenecontaminated drinking water and the risk of breast cancer. Environ Health Perspect 1998, 106 (Suppl 4):947¿953.
3. Aschengrau A, Rogers S, Ozonoff D: Perchloroethylenecontaminated drinking water and the risk of breast cancer: additional results from Cape Cod, Massachusetts, USA. Environ Health Perspect 2003, 111:167¿173.
4. Bukowsk JA: Review of the epidemiologic literature on residential exposure
to perchloroethylene. Crit Rev Toxicol 2011: 1¿12. Available early online from: http://informahealthcare.com/doi/abs/10.3109/10408444.2011.581649.
5. Weiss NS: Cancer in relation to occupational exposure to perchloroethylene. Cancer Causes Control 1995, 6:257-266
6. Lynge E, Anttila A, Hemminki K: Organic solvents and cancer. Cancer Causes Control 1997, 8:406-419
7. Institutes of Medicine: Gulf war and health. Vol. 2: Insecticides and solvents. Washington, DC: National Academies Press; 2002.
8. Mundt KA, Birk T, Burch MT: Critical review of the epidemiological literature on occupational exposure to perchloroethylene and cancer. Int Arch Occup Environ Health 2003, 76:473-491.
9. Ruder AM: Potential health effects of occupational chlorinated solvent exposure. Ann NY Acad Sci 2006, 1076: 207¿227.
10. Boice JD, Marano DE, Fryzek JP, Sadler CJ, McGlaughlin JK: Mortality among aircraft manufacturing workers. Occup Environ Med 1999, 56:581-597.
11. Blair A, Petralia SA, Stewart PA: Extended mortality follow-up of a cohort of dry cleaners. Ann Epidemiol 2003, 13:50-56.
12. Olsson H, Baldetorp B, Ferno M, Perfekt R: Relation between the rate of tumour cell proliferation and latency time in radiation associated breast cancer. BMC Cancer 2003, 3:11. Available from: http://www.biomedcentral.com/1471-2407/3/11.
13. McPherson K, Steel CM, Dixon JM: Breast cancer-epidemiology, risk factors, and genetics. BMJ 2000, 321:624-628.
14. McTiernan A: Behavioral risk factors in breast cancer: Can risk be modified? Oncologist 2003, 8:326-334.
15. Santen RJ, Boyd NF, Chlebowski RT, Cummings S, Cuzick J, Dowsett M, Easton D, Forbes JF, Key T, Hankinson SE, Howell A, Ingle J: Critical assessment of new risk factors for breast cancer: considerations for development of an improved risk prediction model. Endocrine-Related Cancer 2007, 14:169-187.
16. Greenland S: The effect of misclassification in the presence of covariates. Am J Epidemiol 1980, 112:564-569.
17. Marshall JR, Hastrup JL: Mismeasurement and the resonance of strong confounders: Uncorrelated errors. Am J Epidemiol 1996, 143:1069-1078.
18. Hill AB: The environment and disease: association or causation? Proc R Soc Med 1965, 58:295-300.
Competing interests
Funding for this comment was provided by the Halogenated Solvents Industry Alliance.
Authors' Response by Lisa G. Gallagher, Veronica M. Vieira, David M. Ozonoff, Thomas F. Webster and Ann Aschengrau
10 November 2011
Dr. Bukowski, writing at the request of the Halogenated Solvents Industry Association (HSIA), calls into question our results on the grounds that they conflict with occupational studies he alleges show no increased risk of breast cancer at much higher PCE exposures. We understand why the HSIA would want to weigh in on this question because it might suggest that their product, PCE, which is in widespread use and causes extensive exposure in the occupational and general community environment, is an unreasonably dangerous product. Given the size of the exposed population, even relatively small risks could result in an unacceptable breast cancer burden on society.
The majority of the occupational studies he cites examined breast cancer mortality, not incidence, as the outcome, and, as such, did not assess associations with the entire spectrum of this, often non-fatal, disease. These studies also had little adjustment for confounding factors and are confounded by socioeconomic status (SES), an important risk factor for breast cancer. This source of confounding would tend to bias the results of the occupational studies towards the null because low SES women tend to be employed in PCE-exposed occupations while high SES women have an increased risk of breast cancer. In contrast, our study considered and controlled for many potential confounding variables, including a woman's educational level. In our general population study subjects we found little evidence of confounding by SES because the irregular pattern of PCE contamination on Cape Cod often resulted in vastly different exposure levels for adjacent residents and neighborhoods. In addition, the different exposure routes in the occupational studies (e.g. mainly inhalation and dermal) may contribute to the alleged divergent findings between our study and this body of literature.
There are also several statements made by Dr. Bukowski that are incorrect. First, the author misinterpreted our exposure measure as an actual mass value in grams. As noted in our paper, our cumulative measure of PCE exposure was not a mass (a ratio measure), but an ordinal measure used to rank our subjects. In addition, the author's statement that our study results are not internally consistent inaccurately represents our data by mixing results for any exposure with those for exposure levels. Our paper noted no increases in the crude odds ratios until 17 and 19 years latency (ORs 1.3-1.4) among ever-exposed women and stated that adjusted odds ratios among ever-exposed women were null for all latent periods.
Each latency analysis also used a latency-specific exposure distribution to determine percentile cut points. Thus, varying results across latent periods may reflect different cut points (paper Table 2). The cut point for current smoothing analysis (RDD>35) is most comparable to the 90th percentile cut point used in the prior analysis. In addition, some odds ratios for the current 90th percentile may be attenuated because of the lower cut points. In general, the reduction of exposure misclassification resulted in assigning subjects who were unexposed in the prior analysis to low exposure in the current analysis. This change resulted in a cleaner referent population and was expected to strengthen the associations. However, the low exposures determined by the automated method were associated with little or no increased risk.
In summary, we maintain that the new exposure assessment method described in this paper is improved from the original method, as evidenced by an improved correlation with sampling data and reflecting a more accurate estimate of water flow in the piping network. An important contribution of this analysis was to demonstrate that the associations between breast cancer and PCE-contaminated drinking water are relatively robust to refinements in exposure modeling.
Competing interests
We are the authors of the article.