Stratified Design with Prospective Analysis Plan and Time-to-Event Endpoint

Randomized trial comparing new treatment (T) to control (C) includes both classifier positive and classifier negative patients. Presumes availability of binary classifier predictive of benefit for new treatment.


Choose one analysis plan:

Analysis plan A: Determine sample size for overall test comparing T to C for all randomized patients at reduced two-sided level alpha. If overall test is not significant, then test T vs C in classifier positive subset using (.05-alpha) significance threshold.


Analysis plan B: Determine sample size for comparing T to C in classifier positive subset at two-sided .05 level. If that is significant at .05 level, then evaluate classifier negative subset at two-sided .05 level.


Analysis plan C: First test for interaction between size of treatment effect and subset (classifier + or classifier -). If interaction is non-significant, just compare treatments overall at two-sided significance level .05. Otherwise, compare treatments within subsets at two-sided .05 level.



NIH, 2008