A popular class of univariate carcinogenicity tests for non-palpable tumors is generalized to allow for non-independence among different types of tumors. These univariate tests require non-informative censoring and representativeness conditions for lethal and incidental tumors, respectively. The restrictions implied for the joint distribution of tumors are characterized. A general multivariate model is presented, and parameters of the model are estimated to simulate critical levels of test statistics. Useful applications include omnibus tests for carcinogenicity and adjustment of the largest observed treatment effect for the number of statistical tests performed.
- multiple tests
- structural dependence