I am new to survival analysis and am having difficulty choosing how to analyze my dataset.
The background is that I have a pretty big dataset that includes ~22,000 subjects, with slightly over 65K observations over five years.
The 'final state' of the patients can be: Death (due to disease), Death (due to other causes), Cure, Completed Treatment (but did not cure), Fail Treatment (did not complete treatment for some reason), Default (Lost-to-follow up), and Continuing Treatment.
All told there are about ~2700 deaths, ~5000 cures, and ~7400 completed treatment patients
Some complicating factors
- During the course of follow-up (six to nine months for a simple disease, two years for complicated patients), the patients are monitored irregularly and can get zero, one or two different tests that can determine if the patient is 'cured'. I think of the results of these tests as 'states' because they indicate if you are doing well (results=negative) and on the road to cure vs not so well (results=positive) and are more likely to die. Of course these tests are not cheap so not everyone gets them and they also may (or may not) get them very frequently.
- My dataset is long enough (five years) that some patients who failed treatment (or who were cured) actually come back, sick again.
- As mentioned in (1), disease can be multi-faceted (e.g. several levels of severity depending on the strain of infection). This is something I want to eventually analyze.
Because of 1 & 2, I started with multi-state models because the test results can flip flop between negative and positive and some (but not very many) patients came back years later with new infections. I wound up getting overflow errors (see post Numerical Overflow so now I am seeking to simplify my analysis.
So now to the questions which is how to simplify...
(A) Simplify my dataset by completely discarding Death (due to other causes), and lumping together Cure+Complete and also Fail+Default. Often times simplifying helps but it stinks to 'loose' resolution of my data.
(B) Discard the patients who got sick again and stratify my analysis into a several competing risk models. Each model would take the patients with a test results (none/pos/neg) and then model the competing risks of the outcome.
(C) Move to Cure Models (which I am just now finding out about)