Competing Risks and Multistate Models in Clinical Nephrology Research

I n cohort studies and clinical trials, outcomes are often expressed in terms of time to an event such as time to death, time to treatment, and time to disease incidence. Regression methods that handle time-to-event data are commonly called “survival analysis,” even if death is not the event of interest. Survival analyses allow us to examine not just whether an event happened, but how quickly that event happened. Survival analysis can also handle patients who are censored (i.e. those who drop out of the study or who do not experience the event of interest by the end of the follow-up period). Instead of treating these patients as having missing event data, survival analysis makes use of all of the follow-up information that is available by including those who experienced a certain amount of time without an event. Conventional survival analysis, based on Kaplan-Meier methods,

I n cohort studies and clinical trials, outcomes are often expressed in terms of time to an event such as time to death, time to treatment, and time to disease incidence. Regression methods that handle time-to-event data are commonly called "survival analysis," even if death is not the event of interest. Survival analyses allow us to examine not just whether an event happened, but how quickly that event happened. 1 Survival analysis can also handle patients who are censored (i.e. those who drop out of the study or who do not experience the event of interest by the end of the follow-up period). Instead of treating these patients as having missing event data, survival analysis makes use of all of the follow-up information that is available by including those who experienced a certain amount of time without an event.
Conventional survival analysis, based on Kaplan-Meier methods, has a set of strict assumptions about patients who are censored. The most important of these assumptions is that censoring is not informative (i.e. patients who are censored go on to have the outcome at the same rate as patients who are not censored). This assumption is often not met under real-world conditions in clinical research, where patients experience competing risks to the event of interest. 2 A competing risk is an event that changes the likelihood that an event will happen, or prevents the event entirely. For example, in a study examining progression to end-stage kidney disease (ESKD) among patients with chronic kidney disease, if the patient died prior to experiencing ESKD, death is a competing event to developing ESKD. Using traditional Kaplan-Meier methods, these patients would be censored at the time of death, and the model would assume that they had the same risk of ESKD as patients who were censored for other reasons, like loss to follow up. However, we know that this assumption is not true. The event of death actually eliminates the likelihood that the patient will develop ESKD. In the presence of this informative censoring, Kaplan-Meier results should be interpreted as the risk of ESKD among chronic kidney disease patients if the probability of death was eliminated. Though in some cases this interpretation may be of interest, for most research questions, we want to know the risk of our event in a real world scenario that includes competing risks.
A variety of survival analysis methods have arisen to handle situations of informative censoring, called competing risk analyses. Competing risks analyses rely on estimating the causespecific hazard, or the likelihood of experiencing an event from a specific cause in the presence of other causes of that event. One limitation of competing risks analysis, shared by traditional survival analysis, is that it treats all states as final. 3 This makes sense for some conditions, such as death. However, for other conditions such as diabetic ketoacidosis or graft failure after kidney transplant, the state may actually be temporary. For research questions that involve these "transitional states," multistate models can be used to approximate a patient's history of multiple events, not just their time to a single event. 4  Competing risks and transitional states are far from rare in nephrology. Any report containing time to an event as an outcome should be scrutinized as to which patients are censored and whether censoring might be informative when interpreting outcome measures. Clinical researchers interested in time to event outcomes should be familiar with competing risks and be able to identify when a competing risks approach or multistate approach is the most appropriate to answer their research question.  What is the probability that a patient will be experiencing a specific event at a given time?
What is the probability that a patient with acute kidney injury has an eGFR <15 at a given time?
Difficult to obtain longitudinal data on transitional states or outcomes CKD, chronic kidney disease; eGFR, estimated glomerular filteration rate; ESKD, end-stage kidney disease.
COMMENTARY K Ross-Driscoll and RE Patzer: Competing Risks and Multistate Models