Addressing Informative Censoring Bias in Clinical Trials

John M. Burke, MD

Hematologist and Medical Oncologist

Rocky Mountain Cancer Centers

Associate Chair

US Oncology Hematology

Research Program

Aurora, CO

All of us oncologists celebrate a “positive” clinical trial. We pick up our favorite oncology journal and read that treatment X improves progression-free survival compared with treatment Y. Often, the overall survival data are reported as "immature" with longer-term follow-up to be reported later. Treatment X is celebrated as a superior therapy, gets approved by the FDA, is added to NCCN guidelines as a standard of care, and starts to be used in practice. The problem with this pattern is that the celebrated progression-free survival benefit may have been caused by excessive toxicity with treatment X, not by a real improvement in outcomes in patients treated with X. How is this possible? The answer has to do with a form of bias called informative censoring.

What is informative censoring? The problem begins with the use of progression-free survival as the primary end point for a randomized trial. If a patient stops treatment because of toxicity from the treatment, the patient may not be followed to determine disease progression from the time of the cessation of treatment. The patient is then censored from the progression-free survival analysis. If treatment X is more toxic than treatment Y, more patients treated with X will discontinue therapy due to toxicity, and such patients will be censored prior to undergoing response assessment. This censoring, in turn, leads to an analysis of progression-free survival that is not by intent to treat. That is, the analysis does not include all patients who were randomly assigned to X; it excludes those who dropped out due to toxicity. If such censored patients have a higher risk of progression than the overall study population, the results become biased in favor of X. This bias can make the Kaplan-Meier progression-free survival curve for X better than that for Y, simply because X was more toxic than Y.

To give credit where it is due, I have become familiar with the concept of informative censoring from reading several articles about it, often coauthored by Ian Tannock, MD, PhD, emeritus professor of medicine and medical biophysics, Princess Margaret Cancer Centre, University of Toronto in Canada. In some of these articles, the authors list clinical trials in oncology that have contained informative censoring bias. I have found numerous additional examples on my own, even in articles published in the most prestigious medical journals. Informative censoring is not being called out by editors at journals, by NCCN when it is adding regimens to guidelines, or even by the FDA when it is approving treatments for clinical use.

How should we identify informative censoring? First, when analyzing the results of a trial, look at the rate of discontinuation for toxicity in the treatment arms. If there is a higher rate of discontinuation for toxicity in one arm compared with the other, that difference is going to introduce informative censoring bias. This pattern exists in many randomized clinical trials comparing 2 different treatments for patients with advanced cancer, but it may be even more prevalent in adjuvant trials comparing a drug with placebo, which, of course, has no toxicity.

Once we have identified the potential for informative censoring bias, what can be done to determine whether a treatment offers clinical benefit? One option is to place more emphasis on the overall survival end point, which is not going to have informative censoring bias. A second option is to look for a sensitivity analysis in which it is assumed that all patients who have dropped out for toxicity have experienced a progression event, which is creating a worst-case scenario for the investigational treatment in question. (However, the truth is, I almost never find this type of sensitivity analysis in trial reports.) Another option, which is similar to the second one, is to perform an analysis of time-to-treatment failure, in which an event is defined as death, progression, or treatment interruption for any reason. If all progression-free survival “benefit” is caused by informative censoring, then time-to-treatment-failure curves should not show a benefit.

The take-home point here is that improvement in progression-free survival does not always mean that a treatment is providing clinical benefit for a patient. We need to remember that there are only 2 ways to help our patients: make them live longer (improve overall survival) and improve their quality of life. When we analyze clinical trial results to determine whether a treatment has achieved these goals, we need to look beyond progression-free survival at end points such as toxicity rates, overall survival, quality of life, and time-to-treatment failure.