[Should comparison be made between survival of responding vs. non-responding patients?]

In oncology trials, it is not uncommon to compare the survival of the patients who responded to therapy to the survival of the non-responders, while choosing the beginning of therapy as time 0. This comparison is biased because part of the survival time of the responders is the time between the start of therapy and response, which is guaranteed by the observation of the response. The survival experience of responders must be studied starting from the observation of response. A significant difference between the survival experiences of responders and non-responders is often considered as evidence of a causal relation between response and better survival, or even as demonstrating the efficacy of the treatment. A causal relation between response and survival is certainly biologically plausible but the response may as well be associated with some prognostic factors, the selection of responders leading to the selection of patients with a better prognosis, independently of the treatment. The conclusion that the efficacy of the treatment is demonstrated assumes that the non responders, which represent treatment failures, are a control group which can be compared to the responders. This is absolutely untenable, because the two groups are likely to be very different.