Analysis of survival outcomes in haematopoietic cell transplant studies: Pitfalls and solutions

Series Editor Introduction: The final article in our Statistics Series by de Wreede and colleagues deals with the important issue of survival analyses in general and in recipients of haematopoietic cell transplants specifically. At first glance analyzing survival should be simple. The endpoint is clear with rare exception, the subject is either alive or dead. Compare this to other less well defined transplant-related outcomes such as who has acute graft-versus-host disease (GvHD) and of what grade or what is the cause of interstitial pneumonia. There is also the complexity of composite endpoints when one analyzes outcomes such as event-free (EFS) or relapse-free survival (RFS). Here you’re either alive or dead. Period. Alas, as it turns out things are not so simple. As the authours point out: it takes time to observe time. It is almost never possible to wait long enough for everyone in a study to die. (Some people who are cured by a transplant will outlive their physician and statistician.) Other subjects may not be followed until the end of the study, lost to follow-up or withdraw consent to participate. Often these are non-random events, muddy the water and make what seems a simple analysis of survival not so. Fortunately, de Wreede and colleagues discuss the issues of informative and non-informative censoring and time-dependent co-variates. And there are other nasty complexities such non-proportional hazards of death say when initially there is a survival disadvantage to transplants from transplant-related mortality followed in 1–2 years by a survival benefit. They emphasize the danger of considering only Hazard Ratio in this setting. Lastly, the authours discuss how to compare interventions such as conventional therapy versus a haematopoietic cell transplant when the endpoint of interest is survival. We think this article will be of considerable interest to readers of BONE MARROW TRANSPLANTATION and suggest you study it carefully. Survival analyses, seemingly simple, are a potential minefield. You don’t want to step on one. This article and the entire Statistics Series are available online at https://www.nature.com/collections/ejhigdbeeh. Robert Peter Gale MD, PhD & Mei-Jie Zhang PhD. Abstract: The most important outcome of many studies of haematopoietic cell transplants is survival. The statistical field that deals with such outcomes is survival analysis. Methods developed in this field are also applicable to other outcomes where the occurrence and timing are important. Analysis of such time-to-event outcomes has special challenges because it takes time to observe time. The most important condition for unbiased estimation of a survival curve—non-informative censoring—is discussed along with methods to account for competing risks, a situation where multiple, mutually-exclusive endpoints are of interest. Techniques to compare survival outcomes between groups are reviewed, including the instance where it is unknown at baseline to which group a subject will belong later during follow-up (time-dependent covariates).