by Steve Simon, PhD
There are two features of survival models.
First is the process of measuring the time in a sample of people, animals, or machines until a specific event occurs. In fact, many people use the term “time to event analysis” or “event history analysis” instead of “survival analysis” to emphasize the broad range of areas where you can apply these techniques.
Second is the recognition that not everyone/everything in your sample will experience the event. Those not experiencing the event, either because the study ended before they had the event or because they were lost to follow-up, are classified as censored observations.
Some examples of time to event data
The methods for survival analysis were developed to handle the complexities of mortality studies, but they can be used for so much more.
You can study the “death” of mechanical devices, though the term “failure” is probably a better word to use for something that was never truly alive.
You can also study other health related events like relapse or re-hospitalization. The events do not even need to be events that you’d like to avoid. Survival models are used to model the time to pregnancy for couples treated for fertility problems.
Censoring in time-to-event data
One of the hallmarks of survival analysis is censoring. You are measuring the time until a certain event occurs in a sample of people, animals, or machines, and some of those in your sample never experience the event, at least not while you were studying them.
Consider a hypothetical experiment involving the survival times of a sample of 25 fruit flies. You watch these flies daily and whenever a fly drops to the bottom of the cage, you give it a proper burial and record the number of days it was alive.
Suppose that you’ve done this for 15 of the flies, but on day 70 of the experiment, you carelessly leave the cage open and the 10 flies who are still alive bug out.
You might think that your experiment is ruined, but not so fast. You can still estimate the median survival time, because the median fly (#13) died before your gaffe occurred.
Why you have to worry about censoring?
Censored observations are not missing observations. You know something about these ten flies. They were the senior citizens in your sample and last longer than all of the other flies. You don’t ignore this information because ignoring information about the toughest ten flies in your sample seriously biases your outcome.
There’s a story about a statistician who was studying various careers, and found out that the most dangerous career of all wasn’t being a police officer (average age at death 50) or a logger (average age at death 48) or working on an Alaskan fishing boat (average age at death 44). It was being a high school student (average age at death 16).
This is just a made-up story, but it shows that the bias that you can get if you ignore censoring (all those high school students who graduated before dying) can be huge.
The censored values contribute valuable information up to the time of censoring. They represent data that is only partially missing and they can and should be incorporated into your statistical analysis.
Leave a Reply