While quality of life, toxicity and cost are often accepted as important secondary outcomes, the common assumption in most cancer RCTs seems to be that the new treatment should be adopted as the new standard for all patients if statistical assessment of relevant time-to-event HR is significantly better than the standard control treatment.
However, this is a false assumption, as the value of a HR can arise from numerous scenarios. For example a HR of 0.75 will be generated if, in an RCT:
the survival of all patients in the new treatment group is increased by 25%, or
25% of patients in the new treatment group experience an approximate 3-fold survival benefit, but the remaining 75% have no survival benefit, or
25% of patients in the new treatment group experience an approximate 4-fold survival benefit, but the remaining 75% experience a 10% detriment,
This creates a major dilemma, as it appears impossible to tease out the components of a HR, and distinguish which new treatments should be introduced into routine clinical practice for all patients, and which might actually be detrimental to the majority of patients. None of the possible solutions seem to help: modeling suggests that the survival plots resulting from these various scenarios are virtually indistinguishable, this uncertainty is not ameliorated by increasing the sample size (thus meta-analyses are equally unhelpful), and if predictive factor analyses are undertaken and a subgroup of patients is found that benefits from the new treatment, it is not possible to tell whether that subgroup in turn may need to be subdivided further.
Outcomes such as response can identify the impact of treatment on individual patients, but simply comparing the numbers of patients who respond in an RCT does not overcome the underlying problems, as:
the RCT alone does not tell us which specific patient subgroups benefit
different subgroups of patients may benefit from different treatments,
response rates of combination therapy cannot differentiate between the effectiveness of the individual drugs.
Stewart and Kurzrock  have highlighted many of the problems with RCTs in trying to identify ‘who benefits?’ and argued that we need to identify predictive biomarkers for response in phase I and II studies, and use this information to enrich RCTs. Whilst this increases the chances of a clearer outcome, it does not guarantee that all patients will benefit, and does not negate the need to explore other factors over and above the target biomarker. Indeed, if a clear benefit is found in phase I and II studies, there seems little point in running a large expensive RCT.