In cancer clinical trials new regimens are typically tested for antitu
mor activities in patients with advanced disease. The promising ones a
re then compared to the standard treatment in a randomized study, some
times performed on patients with earlier-stage disease. When there are
multiple promising regimens, it may not be possible to compare all of
them to the control group because of the prohibitive sample size and
study length requirements. We propose a design that uses the Cox regre
ssion model to select a best treatment based on survival before the ra
ndomized comparison. Sample sizes for an asymptotically correct select
ion probability of .90 are presented for Weibull survival distribution
s with parameters in a range we consider to be of practical interest.
Simulations verify that the asymptotic approximations to the correct s
election probabilities are quite satisfactory. Simulations also indica
te that the procedure is reasonably robust to the proportional hazards
assumption. In contrast to the two-stage screening design recommended
by Schaid, Wieand, and Therneau (1990, Biometrika 77, 507-513), our d
esign has the advantage of fitting naturally to a progression of cance
r trials where the selection and comparison phases are carried out on
different populations of patients. When the population of interest sta
ys the same, our design can be more conservative on the average but of
fers the opportunity to base the comparative trial on the experience g
ained during the selection phase.