I see no evidence favoring either atezolizumab or chemotherapy in that survival analysis. “Milestone” or “landmark” analyses are very problematic (as discussed here) and the numbers at risk are way too small to make any valid inferences after the 12 month cutoff (when those K-M curves start to separate). Also, keep in mind that they used a stratified Cox model which can address the problem of proportionality if the stratification is done by the covariate(s) that do not satisfy the PH assumption (although one can validly claim that an appropriately covariate-adjusted unstratified Cox model would achieve the same). In addition, there is a cost to testing the PH assumption and then switching to, e.g., a RMST model in that it generates too narrow CI (see here). If anything, I’d say that the results of the presented stratified Cox model are uninformative with the data being compatible with a HR of 0.63 and 1.21 (the data are compatible with clinically meaningful effect size differences at opposite directions).
6 Likes