Growth delay times of experimental tumors after subcurative therapy may be censored because of intercurrent death of the host animals, limitations of the follow-up period, or the number of cured tumors. Under the general assumptions of log-normally distributed data and independence of the censoring event and the therapy effect, it is shown using computer simulations that the estimate of the median growth delay according to the product limit method of Kaplan and Meier, which allows inclusion of censored data, is unbiased. Omission of censored growth delay times from incomplete accrued data often leads to biased estimates. The power of statistical tests for qualitative comparison of two therapy groups with incomplete sets of data from growth assays was also studied. In the absence of censoring, the power of the different tests is about the same. At higher censoring rates of 40%, however, tests applicable to censored data (log-rank test, Gehan-Wilcoxon test) have a markedly higher power than tests applied to the reduced set of complete observed growth delays (mu test, t test). Although complete observation of tumor regrowth should be strived for, growth delay experiments with very delicate animal tumor models can easily result in censored data. The methods presented permit quantitative and qualitative analysis of growth delay data up to a censoring rate of over 30%, if growth delays and censoring events are independent.