Stratifying by year of treatment may suppose two relevant aspects to consider…
On the one hand, with the passage of time the baseline hazard changes, new therapies are introduced, etc., which leads me to think that it is healthy to stratify by year of treatment in observational registries.
However, in diseases with a long evolution, the rate of censored data increases progressively in years close to the present.
In relation to this, I have had the following experience.
In one of my databases I have patients from 2000 to 2019 and I have divided them into four strata for every five years (2000-2004; 2005-2010… etc).
As expected, the censoring rate increases substantially from the strata furthest from the present (22%) to the strata with the closest years (50%).
In the analysis by strata I find that the treatment effect tends to be beneficial in the invervals <2010 (HR 0.75), but not in the strata >2010.
Of note, the bulk of patients were treated after 2010.
When I fit a non-stratified Cox PH model the result leads to no effect, with HR very close to one.
However when I stratify by the year of treatment, the protective effect of the first years predominates, and I get a HR of around 0.82.
Considering both problems I exposed above, how is it more correct to make the survival analysis in these cases, stratified by year or not stratified?