Is att_gt sensitive to periods with few Switchers or small not-yet-treated group?

[tomasrei]

Hi! I am experiencing a curious issue. My gut feeling is that this has something todo either with there being too few units left in the not-yet-treated group or that the estimator is sensitive to having very few, or zero switchers in some years.

It can of course be a sample selection issue. Any advice?

I have data for 1968-2023.

Using 1968-2023:

Using 1968-1996:

• Note that it is not particular to 1996. I very similar results when I pick, for example 1990 (even 1980 is fairly similar).

• Note also that I havn't accounted for population growth here, so I don't think there is an actuall effect here. In fact, I am fairly certain that population groth is what is driving the result.

Number of treated in each cohort:

Share treated

As you can see, not alot happens after 1999-2000 and most units (regions) are treated by then.

Thankful for any advice!