To illustrate the strategy, if not the complete implementation, of an evaluation of sources of error in epidemiologic studies, the first major epidemiologic study on persistent organochlorides and breast cancer by Wolff and colleagues (1993) is examined. The hypothesis they considered was that persistent organochloride compounds, including the pesticide DDT, its metabolite dichlorodiphenyldi-chloroethane (DDE), and the industrial pollutant, polychlorinated biphenyls
(PCBs), might increase the risk of developing breast cancer. A major motivation for such inquiry is the experimental evidence of carcinogenicity of these compounds and the postulated effects of such compounds on estrogenic activity in humans and other species (Davis et al., 1993). Prior to 1993, studies in humans had generally been small and were based largely on comparisons of normal and diseased breast tissue rather than on an evaluation of exposure levels in women with and without breast cancer. Because the report by Wolff et al. (1993) was a major milestone in the literature and stood essentially in isolation, it provides a realistic illustration of the interpretive issues surrounding a specific epidemio-logic study. The fact that a series of subsequent evaluations have been largely negative (Hunter et al., 1997; Moysich et al., 1998; Millikan et al., 2000) does not detract from the methodologic issues posed at the time when the initial study was first published and evaluated.
In order to evaluate the possible association between exposure to persistent organochloride compounds and breast cancer, Wolff et al. (1993) identified over 14,000 women who had been enrolled in a prospective cohort study between 1985 and 1991 that included collection of blood samples for long-term storage. From this cohort, all 58 women who developed breast cancer and a sample of 171 controls who remained free of cancer had their sera analyzed for levels of DDT, DDE, and PCBs. After adjustment for potential confounders (family history of breast cancer, lifetime history of lactation, and age at first full-term pregnancy), relative risks for the five quintiles of DDE were 1.0 (referent), 1.7, 4.4, 2.3, and 3.7. Confidence intervals were rather wide (e.g., for quintile 2, approximately 0.4-6.8 as estimated from the graph, and for quintile 5, 1.0-13.5).
The focus here is on the critical interpretation of these results in terms of epi-demiologic methods, but the contribution of this study to expanding interest in the potential environmental influences on breast cancer generally is a notable achievement with implications yet to be fully realized. The first step in examining these data is to define the result that is to be scrutinized for potential error. Although PCBs were examined as well as DDT and DDE, we will focus on DDE and breast cancer, for which the evidence was most suggestive of a positive association. An entirely different set of criticisms might arise in evaluating the validity of the measured absence of association (or very small association) identified for PCBs.
There were three main calculations undertaken for DDE: a comparison of means among cases versus controls (of dubious value as a measure of association), adjusted odds ratios calculated across the five quintiles (as provided above), and an estimated adjusted odds ratio for increasing exposure from the 10th to 90th percentile of 4.1 (95% confidence interval: 1.5-11.2), corresponding to an assumed increase from 2.0 ng/mL to 19.1 ng/mL. Although the latter number smoothes out the irregularities in the dose-response gradient that were seen across the quintiles, and may mask non-linearity in the relationship, it provides a convenient single number for scrutiny. The question we focus on is whether changing a woman's serum DDE level from 2.0 to 19.1 ng/mL would actually cause her risk of breast cancer to rise by a factor of 4.1.
What are the primary sources of uncertainty in judging whether the reported association accurately reflects the causal relationship between DDE and the development of breast cancer? We ask first whether the association between the study variables was likely to have been measured accurately, deferring any consideration of whether the association is causal. The underlying study design is a cohort, in which healthy women were identified and followed prospectively over time for the occurrence of breast cancer. Given the identification of all cases and appropriate sampling of controls from within this well-defined cohort, selection bias is unlikely. The constitution of the study groups being compared is thus not likely to have distorted the measure of association other than by having drawn an aberrant sample of the cohort to serve as controls, which is accounted for in the measures of precision. Although there is always some degree of laboratory error in the assays of DDE given the technical challenges in measuring the low levels of interest, the masking of case-control status suggests that such errors would be similar for cases with breast cancer as for controls without breast cancer. As discussed at length in Chapter 8 and elsewhere (Kleinbaum et al., 1982), nondifferential misclassification of this nature is most likely to be associated with some shift in the relative risk toward the null value. Furthermore, quality control procedures described in the manuscript make laboratory error an unlikely source of major distortion.
Random error is an important concern, as reflected by the wide confidence intervals. Based on the confidence interval reported for the point estimate of a relative risk of 4.1, 1.5 to 11.2, true values of 3 to 6 or 7 could readily have yielded the observed estimate of 4.1 through random error. The data are not likely to have arisen however, under assumptions about more extreme values that would markedly change the substantive interpretation of the study, such as the null value or relative risks of 10 or 15.
Accepting the observed association as a reasonable if imprecise estimate, the possibility of an association being present without reflecting a causal relation between DDE and breast cancer must be considered. Two key concerns are as follows:
1. Is there some metabolic consequence of early breast cancer that increases the serum level of DDE among cases? Given that serum was collected in the six months or more prior to breast cancer diagnosis, latent disease may have affected the balance between fat stores and serum levels of DDE in a manner that artifactually raised (or lowered) the serum DDE level of cases. A detailed evaluation of the metabolism of DDE in serum is beyond the scope of this discussion, but any such effect on cases would directly distort the measured relative risk given that the controls did not experience the disease of concern. Assessment of the validity of this hypothesis requires examination of the literature on metabolism, storage, and excretion of persistent organochlorides and an understanding of the physiologic changes associated with the early stages of breast cancer. Independent of this study, examining patterns of association for cases with varying stages of disease might help to evaluate whether such bias occurred, with the expectation that the bias would result in stronger influence among cases with more advanced disease and little or no influence among cases with carcinoma in situ of the breast (Millikan et al., 1995). Such a bias might also be expected to be strongest for cases diagnosed close to the time of serum collection (when latent disease is more likely to be present) as compared to cases diagnosed later relative to serum collection.
2. Has lactation or childbearing confounded the measured association between serum DDE and breast cancer? The investigators reported that lactation was associated with a decreased risk of breast cancer (Wolff et al., 1993) as reported by others, and that adjustment for lactation markedly increased the relative risk. Lactation is known to be a major pathway to eliminating stored organochlorides and thus causes lower measured levels of these compounds in the body. Whatever exposure level was truly present prior to the period of lactation, the level measured after lactation would be lower. If adjustment affected the reported relative risk for the comparison of 10th to 90th percentile of DDE to the same extent as it affected their categorical measure of relative risk of DDE, the odds ratio without adjustment for lactation would have been around 2.4 instead of 4.1. Thus, the validity of the lactation-adjusted estimate warrants careful scrutiny (Longnecker & London, 1993).
If early-life DDE levels are etiologically important, lactation presumably has artificially lowered later-life serum levels and introduced error relative to the exposure of interest (prelactation levels). If lactation reduced the risk of breast cancer (independent of its DDE-lowering influence), then lactation would be expected to introduce positive confounding and falsely elevate the relative risk (Longnecker & London, 1993). Lactation would lower the measure of exposure and lower breast cancer risk, so that failure to adjust for lactation would result in a spuriously elevated relative risk for DDE and breast cancer, and adjustment for lactation would therefore lower the relative risk. The reason for the opposite effect of adjustment for lactation is not clear (Dubin et al., 1993), but it suggests that lactation history was associated with a higher level of DDE rather than a lower level of DDE in this population. The high proportion of nulliparous (and thus never-lactating) women in the Wolff et al. (1993) study may influence the observed impact of lactation in comparisons of those with and without such a history.
On the other hand, focusing on lactation as a means of reducing body burden (unfortunately, through exposure to the infant), if later-life DDE levels are critical, and lactation's beneficial impact on breast cancer risk is mediated by reduced DDE levels, then adjustment for lactation is inappropriate given that it is an exposure determinant but not a confounder. The preferred relative risk for estimating the causal effect of DDE would not include adjustment for lactation history. Lactation would be no different than working on a farm or consuming DDT-contaminated fish in that it affects merely the DDE levels but has no independent effects on breast cancer.
Resolution of these uncertainties regarding the role of lactation in the DDE/breast cancer association requires further evaluation of the temporal relationship between exposure and disease, improved understanding of the epidemiology of lactation and breast cancer, and a methodological appreciation of the subtleties of confounding and effect measure modification. If we were able to have measurements available from both early life (e.g., prereproduction and lactation) as well as later life but prior to the development of disease, we could empirically assess the relationship of those measurements to one another and to the risk of breast cancer. The resolution of the role of lactation and breast cancer is also complex (e.g., Newcomb et al., 1994; Furberg et al., 1999), but is an active area of investigation.
Each of these issues could affect the true (unknown) measure of the relative risk in comparison to the observed value of 4.1. We would like to be able to assign probabilities to these alternative scenarios given that they have implications for the interpretation of the study results. If these potential biases were incorporated, the distribution of values around the point estimate would not necessarily be symmetrical, as is presumed for random error, but may take other shapes. For example, metabolic effects of early disease seem more likely to artificially elevate case serum DDE levels relative to controls rather than lower them, so that the confidence interval might be weighted more on the lower relative risk end. Lactation may require several curves to address its potential role according to the alternative hypotheses. Insofar as it reflects a true confounder of the DDE/breast cancer association, more refined measurement and adjustment for the relevant aspects of lactation might be predicted to further elevate the DDE/breast cancer association in the Wolff et al. (1993) study (Greenland & Robins, 1985; Savitz & Barón, 1989). As a marker only of reduced body burden of DDE, it should not have been adjusted and thus the smaller relative risks reported without adjustment may be more valid, making true values below 4.1 more compatible with the observed results than values above 4.1. On the other hand, since the confounding influence of lactation was counter to the expected direction (Longnecker & London, 1993), we may wish to raise questions about the assessment of lactation or DDE, and spread the probability curve more broadly in both directions.
Evaluation of results through specifying and working through the consequences of a series of potential biases, focusing on two principal ones in some detail, has not answered the question of whether the measured association of DDT/DDE and breast cancer was accurate, but it helped to refine the question. Instead of asking whether the study's results are valid, we instead ask a series of more focused and answerable questions that bear on the overall result. Does preclinical breast cancer distort measured levels of serum DDE, and if so, in which direction? Is lactation inversely related to breast cancer, independent of DDE? Is serum DDE level a more accurate reflection of early-life exposure among non-lactating women? Some of these questions point toward research outside of the scope of epidemiology, but other approaches to addressing these questions would involve identifying populations in which the threat to validity is much reduced. The lactation issue could be examined in a population in which breastfeeding is absent, not resolving the questions about lactation, DDE, and breast cancer, but addressing DDE and breast cancer without vulnerability to distortion by lactation. These refined questions are, in principle, testable and would help to resolve the questions raised by the Wolff et al. (1993) study. The critical evaluation of study results should enhance intellectual grasp of the state of the literature, help us judge the credibility of the measured association, and identify testable hypotheses that would clarify a study's results and advance knowledge of the issue.
Was this article helpful?
Learning About 10 Ways Fight Off Cancer Can Have Amazing Benefits For Your Life The Best Tips On How To Keep This Killer At Bay Discovering that you or a loved one has cancer can be utterly terrifying. All the same, once you comprehend the causes of cancer and learn how to reverse those causes, you or your loved one may have more than a fighting chance of beating out cancer.