A Meta-Analysis of Evidence for Hormesis in Animal Radiation Carcinogenesis, Including a Discussion of Potential Pitfalls in Statistical Analyses to Detect Hormesis

Abstract

A database containing 800 datasets on the incidence of specific tumor types from 262 radiation carcinogenicity experiments identified in a comprehensive literature search through September 2000 was analyzed for evidence of hormesis. This database includes lifetime studies of tumorigenic responses in mice, rats, and dogs to exposures to alpha, beta, gamma, neutron, or x-ray radiation. A J-shaped dose response, in the form of a significant decreased response at some low dose followed by a significant increased response at a higher dose, was found in only four datasets from three experiments. Three of these datasets involved the same control animals and two also shared dosed animals; the J shape in the fourth dataset appeared to be the result of an outlier within an otherwise monotonic dose response. A meta-analysis was conducted to determine whether there was an excess of dose groups with decreases in tumor response below that in controls at doses below no-observed-effect levels (NOELs) in individual datasets. Because the probability of a decreased response is generally not equal to the probability of an increased response even in the null case, the meta-analysis focused on comparing the number of statistically significant diminished responses to the number expected, assuming no dose effect below the NOEL. Only 54 dose groups out of the total of 2579 in the database had doses below the dataset-specific NOEL and that satisfied an a priori criterion for sufficient power to detect a reduced response. Among these 54, a liberal criterion for defining a significant decreases identified 15 such decreases, versus 54 × 0.2 = 10.8 expected. The excess in significant reductions was accounted for almost entirely by the excess from neutron experiments (10 observed, 6.2 expected). Nine of these 10 dose groups involved only 2 distinct control groups, and 2 pairs from the 10 even shared dosed animals. Given this high degree of overlap, this small excess did not appear remarkable, although the overlap prevented a formal statistical analysis. A comprehensive post hoc evaluation using a range of NOEL definitions and alternative ways of restricting the data entering the analysis did not produce materially different results. A second meta-analysis found that, in every possible low dose range ([0, d] for every dose, d) of each of the radiation types, the number of dose groups with significantly increased tumorigenic responses was either close to or exceeded the number showing significantly reduced responses. This meta-analysis was considered to be the more definitive one. Not only did it take dose into account by looking for consistent evidence of hormesis throughout defined low-dose ranges, it was also potentially less susceptible to limitations in experimental protocols that would cause individual animals to respond in a non-independent fashion. Overall, this study found little evidence in a comprehensive animal radiation database to support the hormesis hypothesis. However, the ability of the database to detect a hormetic effect was limited both by the small number of dose groups with doses below the range where positive effects have been found in epidemiological studies (≤ 0.1 Gy) and by the limited power of many of these dose groups for detecting a decrease in response.

Acknowledgments

The work by Kenny Crump was supported by Health Canada and by an appointment to the Research Participation Program at the National Center for Environmental Assessment, U.S. Environmental Protection Agency (EPA), administered by the Oak Ridge Institute for Science and Education through an interagency agreement between the U.S. Department of Energy and the U.S. EPA. Daniel Krewski is the NSERC/SSHRC/McLaughlin Chair in Population Health Risk Assessment at the University of Ottawa.

Notes

¹Access to the individual animal pathology data would be needed in order to conduct such tests.

²This is not always the case, since a more liberal NOEL criterion can result in some response being significantly increased in a dataset where previously there were no significant increases, and consequently making all the doses below the new NOEL candidates for inclusion in the analysis; recall that only data from experiments in which there is a significant increase at some dose are eligible for inclusion in the analysis.