Survival analysis with delayed entries (and left truncation ?)

Hello everyone
I have a database starting 01/01/1995 of insured persons some of whom are in a state of disability. I want to calibrate the survival distribution of the population in disability.
However, I can not use the data of people with disabilities between 1995 and 2000 (entered and exited over the period) because my data are not reliable, I exclude them.
So I start my study at 01/01/2000 but I want to keep people disabled before that date (entered before and still alive at 01/01/2000) as they deem to be reliable data and very informative.
I look at the duration survived in disability taking into account the duration already survived before the beginning of my study:
- for the insured who enter into disability state after 01-01-2000 and whatever the date, I consider a duration at the entry null and I observe their exit (for death or censure)
- for the insured who are already disabled on 01-01-2000, I consider their duration in the state of disability on 01-01-2000 and I make sure that they integrate my exposure to risk only when the duration of observation within the study is greater than their duration in the state.

As I understand it, this is delayed entry and there is no real "left truncation":
For those who enter disability after the start of the study, the fact of starting them in 0 and observe the duration can capture their life and makes the assumption risk stability over time.
For those who are already disabled on entry, integrating them into the risk set at time T (which is their duration already passed in the state) makes it possible to impact the right elapsed time when an event occurs ( for example, for an insured who enter disability on 15/06/1996 and dies on 15/01/2000, I want his death to impact the survival time of the 4th year and not the first year if my observation starts in 2000) .

Is my method OK? or is there a bias that can be corrected?

thank you in advance


Less is more. Stay pure. Stay poor.
If I follow you, it seems like your access/viewing of the older data is potentially biasing you. If the old data is questionable in any regards, why use any of it at all? So your desired sample is disabled individuals and the outcome is death, correct. Am I missing anything? What variables are you going to control for?
Not all the old data are questonable, people who survived after 2000 (for whom we didn't experienced exit before 2000) are OK as their ral survival time will be known (exit recorded after 2000). Thus, I want to keep them because they give me some valuable information.
But as I'm keeping only some of the observation and excluding others, I'm wondering what are the bias.


Less is more. Stay pure. Stay poor.
If data quality follow a pattern and aren't questionable at random, you have some form of systematic bias. It is up to you to figure out the mechanism and whether it biases results or if you can generalize results just to a particular subset.