Re: analysis of dioxin cancer threshold.

Mackie et al. (2003) present an exploratory Monte Carlo meta-regression of selected cohort mortality and exposure data for three dioxin-exposed cohorts [National Institute for Occupational Safety and Health (NIOSH), Ranch Hand, and Seveso] as a critique of our earlier publication of a similar analysis (Kirman et al. 2000). Although we appreciate that our preliminary examinations of the cohort mortality dose–response data have been followed by more detailed analyses of these data (e.g., Crump et al. 2003), we disagree with a number of Mackie et al.’s procedures, assertions, and conclusions. Mackie et al. (2003) assert that no weighting of the individual cohort data points is necessary and that a linear regression of the unweighted dose–response data is appropriate. We disagree. The data used in the regression analysis include observations obtained for groups ranging from 19 individuals (Ranch Hand Group R4) to 15,000 individuals (Seveso Zone R females). In an unweighted regression, dose–response information from these two groups is weighted equally, which is clearly inappropriate. Weighting by sample size, while relatively simplistic, eliminates the sensitivity of the regression results to either study size or data-grouping decisions of the study authors. For example, the unweighted fits by Mackie et al. would have been quite different if the results for women in Seveso Zone R had been reported for 789 groups of 19 individuals rather than one group of 15,000. The data are probably best analyzed using a Poisson regression. Estimators produced by maximizing the likelihood function in a Poisson regression are generally preferred over those produced by least squares based on their statistical properties, and the data are automatically weighted based on the expected number of deaths (highly correlated with population size). The results of a Poisson regression would be expected to be nearly identical to those obtained using a population-weighted least-squares regression (ongoing work confirms this). The broad conclusion of Mackie et al. (2003) that the data provide no evidence of a threshold is based on analysis of only three of the five cohorts with available dose–response data. Further, the data used in their analysis do not reflect the most current mortality information available for two of the three included cohorts [NIOSH cohort, updated by Steenland et al. (2001); Seveso cohort, updated by Bertazzi et al. (2001)]. Mackie et al.’s conclusion of no evidence of a threshold seems unreasonable on the basis of this incomplete analysis. Crump et al. (2003) noted that the dose–response data from the manufacturing cohorts do not show any positive trend until groups with estimated doses well above current background exposures are included. Data from the Seveso cohort are consistent with this observation. Finally, data from the occupational cohorts indicate an underlying elevation in cancer mortality independent of 2,3,7,8tetrachlorodibenzo-p-dioxin (TCDD) exposure [standard mortality ratio between 115 and 120 at zero dioxin exposure (Crump et al. 2003; Starr 2001)], consistent with possible confounding by exposure to other chemicals in the workplace, elevated smoking rates, or other factors. In contrast, the data from the Seveso and Ranch Hand populations, with more appropriate control populations, do not show this underlying elevation at zero exposure. Combining these two groups of populations into a single meta-regression may result in spurious effects on the apparent shape of the dose–response curve unless this elevation is accounted for. In this respect, we agree with Mackie et al. (2003) that the epidemiologic data must be analyzed carefully using appropriate statistical and commonsense approaches, and with an awareness of the strengths and weaknesses of the underlying data. However, we think additional work remains to be done to fully appreciate the implications of the available epidemiologic data for the cancer dose– response of TCDD in humans. The authors declare they have no conflict of interest.