Heterogeneous Difference in difference - public policy analysis - staggered/multiperiod/continuous...help!

[Luna Viana]

Dear all,

I need help to get my diff-diff model on track.

- my dependent variable is a continuous variable that measures civil service professionalization

- the treatment is a staggered multiperiod measured as funding allocation by year at the municipality level

- it is a not-yet-treated model, since eventually all municipalities get treated. (most get treated in the first year)

- i want to "weight" the results since the volume of funding is fundamental (like a "dose" in a continuous treatment)


I have reached this initial function:

hdidregress twfe (patronage lmunrev ppoverty) ( firsttreat ) [aweight = stn], group(ibge7) time(Year) controlgroup(notyet) cohortvar(treat, replace) vce(cluster ibge7)


I have a few questions that I hope you can help me with:

1 - since I´m using "not-yet-treated", it is not clear how to define the tvar (in case firsttreat), because all units get eventually treated.

2—Is there a minimum balance between treated and not-yet-treated municipalities in each year period? In the last years, a few municipalities have still not been treated, and in the last year, everyone has been treated.

3—Is the weighting that I´m suggesting correct? Its objective is to weigh municipalities by the volume of resources they received.


Thank you!
Luna Viana

[Jared Greathouse]

You gotta email Clement de Chaisemartin. He's an expert on DID. Continuous DID is such an involved, and indeed unsolved, topic that I'm pretty confident nobody here would have any idea on how to really think about this. Clement is the econometrician who knows about all this, so I'd imagine he'd have some thoughts.

[Luna Viana]

Thank you Jared! I hoped it would be my ignorance, so things are indeed really complicated. I still hope that others also see and comment on this thread.

Since I got stuck with the hdidiregress (which seemed simpler to explain in my paper, because it considers mostly interactions following Wooldrige 2021), I tried the csdid (which is more difficult to explain Callaway and Sant´Anna 2023).

The csid ran and gave me theory interesting results.

csdid patronage lgovfundingammendp munrev ppoverty , ivar(ibge7) time (Year) gvar(firsttreat) notyet method(reg) wboot rseed(08052021)

patronage = level of professionalization
lgovfundigammendp = log of transferred resources
munrev/ppoverty= control variables
unit = ibge7 - identify municipalities
first treated = the first year that received amendments
notyet
i used regression because it is something that Ican explain

however, a few questions remain:

- is it ok to keep in the dataset municipalities that are treated all years?
- is it ok to keep municipalities that are treated intermittently?
- how to test strong parallel trends?
- placebo test?
- what econometric equation describes the csdid calculation (model identification)?
- Did I lose my weight? All municipalities are being treated equally independently of the volume of allocated resources.


Thank you all