DID for cross-country

[Iris Landi]

Hello,

For my thesis I am using a difference in differences approach to examine the effect of a policy at country level. The question is: does the new policy implemented in 2003 affect the gdp per capita growth (GDPpc_g) of the country? (policy implemented in Rwanda, with Uganda as control country). I have data from 1997 to 2020 (policy implemented in 2003).

This is what I have done until now, however the model seems to not be significant. What should I do now?

• tset Country_ID Year
• gen post2003 = (Year > 2003)
• gen rwanda = (Country == "Rwanda")
• xtreg GDPpc_g i.post2003##i.rwanda, fe

note: 1.rwanda omitted because of collinearity.

Fixed-effects (within) regression Number of obs = 62
Group variable: Country_ID Number of groups = 2

R-squared: Obs per group:
Within = 0.0062 min = 31
Between = 1.0000 avg = 31.0
Overall = 0.0045 max = 31

F(2, 58) = 0.18
corr(u_i, Xb) = -0.6719 Prob > F = 0.8346

---------------------------------------------------------------------------------
GDPpc_g | Coefficient Std. err. t P>|t| [95% conf. interval]
----------------+----------------------------------------------------------------
1.post2003 | -.4010764 3.684916 -0.11 0.914 -7.777238 6.975085
1.rwanda | 0 (omitted)
|
post2003#rwanda |
1 1 | 2.584065 5.211259 0.50 0.622 -7.847402 13.01553
|
_cons | 2.748484 1.929551 1.42 0.160 -1.113932 6.6109
----------------+----------------------------------------------------------------
sigma_u | .81576852
sigma_e | 10.210224
rho | .00634307 (fraction of variance due to u_i)
---------------------------------------------------------------------------------
F test that all u_i=0: F(1, 58) = 0.09 Prob > F = 0.7660

Thank you in advance for any help

[Jared Greathouse]

You need more control units. The only way for this to be valid is is you can convincingly argue that Uganda's economy for the last 6 years is like Rwanda in every other way but for the treatment, and with one control country you rarely can make that case. Your results will have better power with a longer time series too.

Beyond that, you don't need xtreg. Normal regress will do this just fine