Stochastic models of tumorigenesis have been developed to investigate the implications of experimental data on tumour induction in wild-type and p53-deficient mice for tumorigenesis mechanisms. Conventional multistage models in which inactivation of each p53 allele represents a distinct stage predict excessively large numbers of tumours in p53-deficient genotypes, allowing this category of model to be rejected. Multistage multipath models, in which a p53-mediated pathway co-exists with one or more p53-independent pathways, are consistent with the data, although these models require unknown pathways and do not enable age-specific curves of tumour appearance to be computed. An alternative model that fits the data is the 'multigate' model in which tumorigenesis results from a small number of gate-pass (enabling) events independently of p53 status. The role of p53 inactivation is as a rate modifier that accelerates the gate-pass events. This model implies that wild-type p53 acts as a 'caretaker' to maintain genetic uniformity in cell populations, and that p53 inactivation increases the probability of occurrence of a viable cellular mutant by a factor of about ten. The multigate model predicts a relationship between the time pattern of tumour occurrence and tumour genotype that should be experimentally testable. Stochastic modelling may help to distinguish 'gatekeeper' and 'caretaker' genes in other tumorigenic pathays.