Difference-in-differences estimation

[Aisha Khan]

I have a data set on around 3752 farmers over three time periods and I am trying to implement a Difference-in-differences estimation to understand of dissemination of technical information impacted crop yields.
The dataset is on farmers (Respondents). They are located within villages (41 villages), located within higher geographical units called Halqa (4 in total), located within higher geograohical units called tehsils (2 in total). I have this data for 3 time periods (wave1 in 2020, wave 2 in 2021 and wave 3 in 2023). Out of the 4 Halqa 2 are treated and 2 are control. Both tehsils have treated and control Halqa

My question is what is the correct way to implement a DID with three time periods. I see the following approaches
-> I see papers that drop the midline and use only the baseline and the endline.
-> The thread here answered by Jeff Wooldridge says that a fixed effects model can be used as

xtreg Yields treatment i.time controls, fe

-> Another thread here points out that we can use an interaction between the time period and the treatment

xtreg Yields i.time##i.treatment controls, fe

-> Or the following command can be used

 xtdidregress (Yields controls) (treatment ), group(qh) time(wave)

I have read many threads on this question but am not very clear which one is appropriate in my case.

[George Ford]

What year are they treated?

[Aisha Khan]

George Ford thank you for your reply.
The treatment occurs in October to November 2020 and then in October to November 2021.
The baseline was conducted from Sept to Oct 2020, the midline in October to December 2021 and endline in November to December 2023.

[George Ford]

So all treated are treated at the same time, but twice?

[Aisha Khan]

George Ford. thank you for your reply.
Yes. There are 4 'Halqa' in the data. Out of these 4 Halqas 2 are treated and 2 are control.
(Halqa are geographical units nested within villages).
The treated Halqas receive treatment twice.

[George Ford]

So the treatment is removed at some point?

[Aisha Khan]

George Ford thank you for your reply.
the treatment is 'technical demonstrations' which are provided to the treated sample at two points of time . these technical demonstrations are on a 'crop harvester technology'.

[George Ford]

I'm thinking the demonstration in post1 may affect outcomes in later periods even absent the treatment.

as a guess, you might start here. some of it may be unecessary (post4)

g treated = any treated farmer (could also define on village or some level of aggregation that captures just the treated)
g post0 = dummy for 1st period, 1 during and all periods after
g post1 = dummy for first treated period (if gap, make it 0)
g post2 = gap between 1st and 2nd treatment
g post3 = dummy for second treated period (if data after, make it 0)
g post4 = period after treated

reghdfe Yield c.treated#(c.post1 c.post2 c.post3 c.post4) , absorb(respondent time) cluster(respondent)

this gives you unique DID coefficients for each treatment periods and gaps. if the first treatment gives full effect, then all DID coefficients will be equal.

reghdfe Yield c.treated#(c.post0 c.post2 c.post3 c.post4) , absorb(respondent time) cluster(respondent)

The post2/3/4 dummies measure the difference between the effect of first exposure.

[Aisha Khan]

George Ford Thank you for your reply!

I have one related question.

The technology demonstrations (i.e., treatment) were on the 'crop harvester' technology. The adoption of the 'crop harvester' technology is hypothesized to lead to improvements in yields.
Should I also include a dummy variable to denote whether the 'crop harvester' is adopted in the equation?

[George Ford]

Are you trying to determine the effect of a demonstration or the harvester, or both? You think the demonstration will have an additional effect to just using the harvester?

[George Ford]

If the harvester has an effect, then that's a treatment. The demonstration is a second treatment. So what you're interested in is whether the harvester had an effect, and then whether the demonstration make the effect larger.

Now you have two treated variables: treated1 (harvester) treated2 (demonstration and harvester).

[Aisha Khan]

George Ford Thank you for your reply.

My main hypothesis is that the ‘technology demonstrations’ will results in an improvement in crop yields. The mechanism by which the ‘technology demonstrations’ improve yields is the adoption of ‘crop harvester’. This is because the ‘technology demonstrations’ showed and motivated the farmers on the use of the ‘crop harvester' machines.
So, the ‘technology demonstrations’ will results in an improvement in crop yields via adoption of the ‘crop harvester’

[George Ford]

Can you have a harvester without a demo, or a demo without a harvester?

[Aisha Khan]

No. the demonstrations were specifically to show the use and motivate the farmers for the 'crop harvester technology'

[George Ford]

So all demos result in a harvester use? And no harvester use without a demo?