Flexible Analytical Methods for Adding a Treatment Arm Mid-Study to an Ongoing Clinical Trial

Abstract

It is not uncommon to have experimental drugs under different stages of development for a given disease area. Methods are proposed for use when another treatment arm is to be added mid-study to an ongoing clinical trial. Monte Carlo simulation was used to compare potential analytical approaches for pairwise comparisons through a difference in means in independent normal populations including (1) a linear model adjusting for the design change (stage effect), (2) pooling data across the stages, or (3) the use of an adaptive combination test. In the presence of intra-stage correlation (or a non-ignorable fixed stage effect), simply pooling the data will result in a loss of power and will inflate the type I error rate. The linear model approach is more powerful, but the adaptive methods allow for flexibility (re-estimating sample size). The flexibility to add a treatment arm to an ongoing trial may result in cost savings as treatments that become ready for testing can be added to ongoing studies.

Key Words:

• Adaptive combination test
• Adaptive design
• Adding a treatment arm
• Multi-arm clinical trials
• Phase III clinical trials

ACKNOWLEDGMENT

This work was supported by the NIH (National Institute of Neurological Disorders and Stroke) U01NS043127 and U01NS059041.

Notes

Note. POOL=t-test (pool data across cohort). LIN =linear model (adjusting for cohort as fixed effect); ICHI =inverse chi-square combination test; INORM=weighted inverse norm comb test; t _P is the time at which design change is made (as a proportion of total sample size enrolled). H _A : θ _A  ≤ 0, H _B : θ _B  ≤ 0 where θ _A  = μ _A  − μ _P and θ _B  = μ _B  − μ _P . The results are given for different sizes of intra-stage covariance {0, 1/4θ, 1/2θ, θ}, where θ = 0.38 is the true effect size. Simulations for fixed sample size of 120/group.

Note. POOL = t-test (pool data across cohort). LIN =linear model (adjusting for cohort as fixed effect); ICHI =inverse chi-square combination test; INORM =weighted inverse norm comb test. Familywise error rate is the probability of rejecting any true hypothesis from a family of hypotheses given all possible configurations of the null hypotheses (μ _A  = μ _B  = μ _P , μ _A  > μ _B  = μ _P or μ _B  > μ _A  = μ _P ). Under closed testing procedure, a particular hypothesis is rejected if all intersection hypotheses containing it are rejected (e.g., H _A is rejected if H _AB and H _A are rejected). t _P is the time at which design change is made (as a proportion of total sample size enrolled). The results are given for different sizes of intra-stage covariance {0, 1/4θ, 1/2θ, θ}, where θ = 0.38 is the true effect size.