A key issue in assessing the efficacy of new drugs in oncology is getting the balance right between choosing a hard endpoint survival – and an endpoint that allows evidence to be assessed more quickly progression. Often discussions such as these are required after interactions with regulatory agencies. Progression free survival is used in many studies to assess whether drugs are providing benefit, but it comes with difficulties in analysis and interpretation. In any survival analysis (what we call any “time to-event” analysis), the statistician needs to decide on censoring – patients who “run out” of data. One of the fundamental assumptions in survival analysis is that the censoring is unrelated to the survival times. If a treatment works then you’d hope for longer survival times in one treatment group. If more patients, or more of a certain type of patients, on one of the treatment groups are censored, then this assumption will not hold.
One of the main points of contention relates to handling patients who stop randomised therapy and start a new therapy while on the study. Should these patients be censored or not? There is a good chance that patients who are changing therapies are not doing well on their current therapy, either for safety or efficacy reasons. If the patient is censored, then the negative information about the treatment is ignored, and it’s very likely that a bias will be introduced in the analysis, with resultant incorrect estimates of treatment effect. Recently, in a PhRMA sponsored study, it was found that anything other than a strict intent to treat analysis introduces bias. To avoid bias, patients should be followed up until progression, irrespective of whether they remain on the treatment being studied.
After ensuring that data collection continues at least to treatment progression, best practice would also be to choose a primary analysis that minimises bias, but to ensure that a range of sensitivity analyses are carried out to explore possible biases in the estimated treatment effect. One obvious sensitivity analysis would be to explore the treatment effect with and without censoring at the point patients switch to alternate therapies. Additionally, there is a theoretical possibility of over-estimating progression-free time due to the fact that scans are only taken at certain intervals, and the exact time of progression would be at some time between two scans. It would also be a good idea to check that the pattern of patients who are censored does not seem to be related to treatment.
For late phase oncology trials early decisions may be made using PFS, and a later decision based on overall survival. A fundamental principle in traditional clinical trials is that the chance of a false positive is kept to a minimum, usually 5%. When there are multiple decisions relating to success or failure (e.g. multiple looks at the data or multiple endpoints of interest) an adjustment needs to be made in the analysis so that the overall chance of a false positive is controlled. In the last few years, the FDA has been asking for more stringent requirements with regards to controlling the false positive rate, by requesting “strong control”. Care is required to ensure “strong control” in the design of clinical trials, especially in the case where an early decision is made using PFS, and a later decision based on overall survival.
There are many other statistical issues in oncology studies: early phase study designs – adaptive designs – continual reassessment models; accounting for treatment switching; dealing with delayed treatment effects and issues relating to biomarkers and genetics. Special care should be taken in the statistical aspects of oncology trials.
