PurposeTo improve the understanding of the appropriate design of phase II oncology clinical trials, we compared error rates in single-arm, historically controlled and randomized, concurrently controlled designs.Patients and MethodsWe simulated error rates of both designs separately from individual patient data from a large colorectal cancer phase III trials and statistical models, which take into account random and systematic variation in historical control data.ResultsIn single-arm trials, false-positive error rates (type I error) were 2 to 4 times those projected when modest drift or patient selection effects (eg, 5% absolute shift in control response rate) were included in statistical models. The power of single-arm designs simulated using actual data was highly sensitive to the fraction of patients from treatment centers with high versus low patient volumes, the presence of patient selection effects or temporal drift in response rates, and random small-sample variation in historical controls. Increasing sample size did not correct the over optimism of single-arm studies. Randomized two-arm design conformed to planned error rates.ConclusionVariability in historical control success rates, outcome drifts in patient populations over time, and/or patient selection effects can result in inaccurate false-positive and false-negative error rates in single-arm designs, but leave performance of the randomized two-arm design largely unaffected at the cost of 2 to 4 times the sample size compared with single-arm designs. Given a large enough patient pool, the randomized phase II designs provide a more accurate decision for screening agents before phase III testing.
True and false positive rates for different criteria of evaluating statistical evidence from clinical trials