Statistical simplicity in long-term animal feeding studies of suspect carcinogens

The statistical interpretation of long-term animal carcinogenesis experiments seems complicated at first sight, because the early death of the treated animals can do two things. First, if it is early enough in the animal's lifespan, premature death may prevent the occurrence of a tumour that would have arisen had the animal lived on, thereby artificially diminishing the net tumour yield and suggesting (to any unwary person who merely looks at the percentage of animals that ever develop tumours in the whole course of the experiment) that treatment inhibits the processes of carcinogenesis. Second, if the early death takes place after the tumour is present in an internal organ but before that tumour would normally have been detectable, it will lead to earlier detection of that tumour (suggesting, to anybody incautious enough to use life-table methods on 'incidental' tumours, that the treatment enhances the processes of carcinogenesis). Fortunately, simultaneous avoidance of both these pitfalls is possible by the use of quite simple statistical methods, which are described in the Statistical Annex to the International Agency for Research on Cancer's 1980 monograph entitled 'Long-term and short-term screening assays for carcinogens: a critical appraisal (IARC Monographs, Suppl. 2.). These methods are straightforward enough to be understood and used confidently by non-statisticians and, although a computer program is freely available on request to perform them, it is possible to do them on hand calculators. The IARC monograph includes some real-life examples where serious errors would arise if these or similar methods were not used. The only drawback is that some estimate of the 'context of observation' of each tumour is needed (i.e., fatal, probably fatal, probably incidental, or incidental). The practical difficulties in this are reviewed and, in practice, contexts of observation can be estimated reasonably confidently in over 90% of cases. Although there are a few instances where it is impossible to tell whether or not a particular tumour was the underlying cause of death of the host, these are not in practice sufficiently numerous to cause any material difficulties of interpretation of the overall results. A small change in standard laboratory practice is, however, necessary: at gross necropsy, the 'contexts of observation' of any putative tumours should be estimated. The extra work is negligible, and this one extra piece of information may be extremely useful, while attempting to record it can obviously do no harm.