The purpose of this paper is to address the very important problem of
accurate statistical analysis of certain types of cancer inhibition/pr
omotion (IP) experiments. These experiments are routinely used by the
National Cancer Institute to test the effects of potential chemopreven
tative agents. The statistical analysis is difficult since there is Ty
pe I censoring. In the IP experiments under investigation, laboratory
animals (rats) are injected with a single dose of either a direct or i
ndirect acting carcinogen. In the mammary tumor system, animals in the
control group generally develop 5-7 tumors and typical experiments ar
e usually terminated after 4-6 months. Animals are sacrificed at the e
nd of the experiment and all observed tumors are confirmed. The two mo
st common response variables are the number of observed tumors per ani
mal and the rate of tumor development. The difficulty in analyzing the
se experiments occurs because experiments are terminated before all in
duced tumors have been observed. Fewer observed tumors in one group co
mpared to another could be the result of a decreased number of induced
tumors, a decrease in growth rate, or a combination of both. It is es
sential for the experimenter to distinguish between these two differen
t biological actions. Present statistical techniques do not account fo
r this confounding and since they rely primarily on nonparametric proc
edures, do not present an accurate description of potential IP agents.
In this paper we introduce a parametric procedure that explicitly ack
nowledges the confounding present in experiments of this nature. The a
nalysis is based on the comparison of the mean number of tumors per gr
oup (lambda) and the mean time to tumor appearance (mu). A longer mean
lime to development is believed to indicate a slower tumor growth rat
e. Hypothesis tests are developed to determine if there is an overall
experiment effect, to isolate which groups are contributing to an obse
rved experiment effect, and to isolate factors (tumor number and/or gr
owth rate) contributing to an observed group difference. Confidence re
gions for (lambda, mu) are also generated, This analysis leads to a be
tter understanding of how potential IP agents function.