I have to do a diff-in-diff estimator in order to evaluate the impact of the policy. The problem is that the policy is gradually implemented and in the first year of implementation, it affects 25% of the observations, then 50%,75% and 100% for the remaining years in the sample. The easiest thing to do is to remove the years in which the impact was gradual. However, I still have to somehow account for this gradual policy implementation.

I am wondering how can I do that? One thing that comes to mind is to guess, which observations were impacted by the policy in each of the years(so guess the 25% in year 1, and assign them the treatment and remove the other observations(that are not the control group) and do the same for each of the other years in which the impact was not 100%. Is that a legit thing to do and is there something else that I can do? (I do not observe which observations were impacted and have to guess based on very limited demographic information).

I would really appreciate some input.