Dendreon Corp fka DNDNQ

[ocyanblue]

Cox is an iterative procedure that continues to add/subtract individual prognostic values to correct for imbalances. So, inherent in its working is an averaging process due to the additions/subtractions. In signal processing, it is well-known that averaging is a good filter to reduce noise. That's likely what gives Cox the additional power even when the arms are perfectly balanced in all prognostic factors. That is, treatment arms can be balanced even as individual data points fluctuates, ie, noisy, and Cox averages out the noise.

Cox will worsen the p-value if treatment performs at least as well as placebo and there were significant imbalances in favor of the treatment arm. Witness the p-value of the Cox model V in the FDA's stat review of D9901. That's the case where there were only one 1 missing data point so it's closest to ITT yet the p-value worsened to .04 from log-rank .01 because there was an imbalance in Gleason scores favoring the treatment arm.