The draft’s critiques of studies are often uneven; studies that found no association are criticized more often than studies that found a positive association even if they had similar methodologic limitations. An example is the discussion of case-control studies on page 4-150, lines 19-31. Several of the criticized features of the case-control design that are mentioned are not inherent in the design, such as that associations may be nonlinear (this design does not require categorical exposure measures) or that duration and cumulative exposure do not address age at first exposure (this information can simply be asked of participants). Many of the studies suffered from a lack of statistical power—a common problem in studying rare cancers and exposures. However, the concern over power is uneven. On page 4-149, the absence of an association between employment in dry-cleaning and death due to lymphatic and hematopoietic cancer (Ruder et al. 2001) is attributed to lack of power. In contrast, a positive association between exposure to tetrachloroethylene and multiple myeloma in aircraft maintenance workers was based on only two deaths and is described only as noteworthy but imprecise (page 4-148, lines 6-9). There is little discussion of the potentially important limitations of proportionate-mortality studies, such as inaccuracies in death certification and the inability to adjust for potential confounders. There is some discussion of confounders in relation to the standardized mortality ratio (SMR) studies of esophageal cancer on page 4-153, but it is also unbalanced in that it focuses on adjustment for smoking but does not mention the absence of adjustment for alcohol; in addition, the effect of adjustment for smoking is derived from estimates for lung cancer and may not translate directly to esophageal cancer.
A number of errors suggest an incomplete understanding of epidemiologic and statistical methods. Such errors reduce confidence in the draft’s conclusions. For example, EPA summed observed and expected cases from studies with diverse types of end points (incidence and mortality) and, using different approaches to calculating the expected values, calculated a ratio of the summed observed and expected values. Expected numbers from different studies can be added only if they are derived from the same external rates, but mortality and incidence are different. One of the most troubling misunderstandings is related to the dismissal of the results of the 2006 study by Lynge et al. In reference to that study’s findings on non-Hodgkin lymphoma (and later on bladder cancer), EPA notes that exposure information was not available on about 20% of cases and of controls and that much of the exposure information came from next of kin. It then uses that to explain why Lynge et al. found no risk associated with tetrachloroethylene exposure and suggests an automatic bias toward the null due to misclassification. In the first instance, missing exposure data are analogous to nonresponse in that the subjects are not included in any classification group. Nonresponse will not introduce bias if it is nondifferential; if it is differential, it could bias an effect measure either toward or away from the null. In the second instance, exposure information from next of kin make it more likely that hazardous exposures will be overreported by the families of workers who developed cancer than by families of workers who did not; this would have resulted in