Year*Treatment Adjustment and Matching for Staggered Difference-in-Differences
Hello everyone,
Forgive me if anything is unclear, I am still learning the nuances of this analysis.
I am working with long-form panel data regarding coups in different countries. Basic variables are:
My general idea is to run a staggered diff-in-diff using coup as the treatment and factors like ln(GDP), ln(population) and HDI as controls. To examine parallel pre-trends, I am using lgraph to show differences between the 2 groups for each factor by region, such as:
The other method I am using is a set of coup*year interactions for each variable in the form of:
(Side note: I am also unsure whether it is more proper to use coup or coup_evr for these tests.)
However, it's been brought up that I should use years adjusted around the treatment instead of strict years . To construct the adjusted year, I use:
This is much more logical, but runs into a few hurdles. First, my interactions test cannot run with negative factor values (adjyr = -1, -2... for pre-treat years). Second, my control group comparison has no treatment to center year adjustment around, making graphing on adjyr impossible with my current approach.
My professor suggested covariate matching, so I have been examining propensity score match methods, but my understanding of how to do so through Stata is foggy. I have seen some applications of matching through FD methods (Weihua An, 2016. "Combining Difference-in-difference and Matching for Panel Data Analysis," 2016 Stata Conference 21, Stata Users Group.), though the referenced command DIDMatch does not seem to exist currently, and I am not confident in my ability to manually implement the method with the time I have.
I am currently using didregress for the final analysis, and I have been looking at flexpaneldid as it seems to handle/display pre-trend matching and testing inherently, though the documentation says it requires the treatment variable to equal 1 for the full treatment period. I am wondering if I would need to set my coup variable =1 after the initial coup year, and if so, how I should go about constructing it?
All advice, general or specific, is greatly appreciated. I know this is a bit of jumble so I will try to clarify anything that I can. Thank you all!
Forgive me if anything is unclear, I am still learning the nuances of this analysis.
I am working with long-form panel data regarding coups in different countries. Basic variables are:
HTML Code:
coup = 1 if coup occurred in country in obs. year coup_evr = 1 if coup in country anytime during period iso = 3-letter country code region = numbered IMF regions
Code:
foreach x in 1 2 3 4 5 {
lgraph lnGDP year if region == `x', by (coup_evr) saving(Avg_lnGDP_R`x'.gph, asis replace)
}
Code:
foreach x in 1 2 3 4 5 {
reg HDI coup##ibn.year if region == `x'
}
However, it's been brought up that I should use years adjusted around the treatment instead of strict years . To construct the adjusted year, I use:
Code:
bysort iso (year): gen firstyr = year if coup == 1 bysort iso (firstyr): replace firstyr = firstyr[1] bysort iso (year): gen adjyr = year - firstyr
My professor suggested covariate matching, so I have been examining propensity score match methods, but my understanding of how to do so through Stata is foggy. I have seen some applications of matching through FD methods (Weihua An, 2016. "Combining Difference-in-difference and Matching for Panel Data Analysis," 2016 Stata Conference 21, Stata Users Group.), though the referenced command DIDMatch does not seem to exist currently, and I am not confident in my ability to manually implement the method with the time I have.
I am currently using didregress for the final analysis, and I have been looking at flexpaneldid as it seems to handle/display pre-trend matching and testing inherently, though the documentation says it requires the treatment variable to equal 1 for the full treatment period. I am wondering if I would need to set my coup variable =1 after the initial coup year, and if so, how I should go about constructing it?
All advice, general or specific, is greatly appreciated. I know this is a bit of jumble so I will try to clarify anything that I can. Thank you all!
Comment