Interpreting Generalized Difference in Difference outputs - Is my coefficient cumulated?

[Tobi Frohne]

Hi everyone,

[TLDR: I'm unsure how to interpret my results. Is the GDiD coefficient I get for for my log-ed variable "cumulative"? Any tips for meaningful interpretation?]

I am Tobi, currently a Master student of economics and dealing with end-of-term papers at the moment. I feel like having spent countless hours on this forum already reading and researching, and finding much helpful information and suggestions in the past. Anyway, I don't consider myself a genius with either econometrics or Stata, so a lot of what I learn is trial and error, and maybe my question suggests a lack of basic understanding of what I am dealing with - that's how I feel anyway.

To my question: For a research task I put together a panel dataset (ca. 200 countries à 11 years: 2009-2019) with a bunch of macroeconomic indicators to investigate the "effects" of a policy measure, the participation of states in the Belt and Road Initiative (BRI). I will give you just the core info, as I assume the question is much more general:
I pursued a GDiD approach as countries in my dataset entered treatment, i.e. signed some BRI participation agreement, in different years from 2013 onwards. I also have log-ed DVs.

BRIS: Dummy variable marking measurements during years of BRI participation
TREAT: Dummy variable indiciating a country's status as a BRI participant at some point in time (fun fact, no country discontinued its participation so far, good for me)
GDPPCP: GDP per capita PPP in current US$
LGDPPCP: = log(GDPPCP) as the GDP variable has a skewed distribution in my dataset and the results with it look better. I transform it back using =exp(x)-1

With that I did:

xtset Ccode YEAR, yearly

xtreg LGDPPCP i.BRIS i.TREAT i.YEAR , fe robust

And I got:

VARIABLES	LGDPPCP
1.BRIS	0.038**
	(0.016)
1o.TREAT	-
2010.YEAR	0.040***
	(0.003)
2011.YEAR	0.084***
	(0.007)
2012.YEAR	0.114***
	(0.009)
2013.YEAR	0.151***
	(0.011)
2014.YEAR	0.182***
	(0.013)
2015.YEAR	0.184***
	(0.016)
2016.YEAR	0.220***
	(0.017)
2017.YEAR	0.260***
	(0.018)
2018.YEAR	0.291***
	(0.020)
2019.YEAR	0.325***
	(0.021)
Constant	9.129***
	(0.011)
Observations	2,139
Number of Ccode	198
R-squared	0.498
Adj R-squared	0.495
F-test	64.80
Prob > F	0

So, I noticed basically the longer my pre-treatment period the bigger the coefficient (I use inrange(YEAR, a, b) and some other spiels for testing), and the YEAR outputs seem to "accumulate" as well, which is something I was totally not aware before could or should be happening, if it is indeed the case, and if indeed I didn't make any other foolish mistakes. I am really confused atm, as it makes total sense and no sense at all to me at once, and causal effects in this particular example are dubious anyway... I would be really glad for any help.

Questions:
1) Does my proceeding make sense?
2) Is the coefficient indeed cumulated and should it be like that?
3) How to interpret this? Can I just divide it by the amount of years investigated to get something like a yearly factor showing treatment effects?

Again, apologies if this is actually really simple and I really should have understood all this before bothering with such models...


Regards
Tobi

[George Ford]

If you have different treatment dates, you need to look at csdid or something like it.

It's not accumulating, I don't think. You're adding in higher GDP countries over time, possibly. i.YEAR should address inflation. Country fixed effects?

Try this to see what happens.

reghdfe LGDPPCP i.BRIS i.TREAT , absorb(country year) cluster(country)

But with different treatment dates, you've got a bigger problem.

[Jared Greathouse]

My mentor wrote a small document of the different DD commands in Stata that you might find useful. I'm attaching it. I also did the same with synthetic controls, but since you're interested in DD, here it is.