How to deal with patients falling out of the study more frequent in one treatment arm I'm doing a retrospective comparison of two treatments (A and B) for eye disease. Both procedures have the same most common side effect (dependent variable) but their indications differ slightly resulting in different average age of the patients treated (75 in treatment A and 69 in treatment B).
As you would expect, older patients treated with treatment A die earlier and more frequently (of unrelated causes) resulting in a shorter observation time. How can I deal with that? Is this something that requires data imputation?
 A: Read up on survival analysis. As the name suggests, it is often used to compare time until death. But it can be used to compare time to any event. Assuming the side effect you are tracking is an event (it happened or not) and not a measurement (highish level of some blood test), then you can use survival analysis to compare the times until that event occurs.
When comparing time until death (what you'll see in most texts), the event is death, and observations of the people who leave the study for some reason or who are alive at the end of the tracking period are said to be "censored". But for your study, death would not be considered to be an "event" but rather "censoring", as that person no longer can get the side effect. The terminology seems backward, but the logic makes sense.
Then you can run a Kaplan-Meier test to create the "survival" curves and a logrank test to compute a p-value testing the null hypothesis that the time until the side effect occurs is the same with the two treatments.
