Hi all
For the past week or two I've been so confused on what regressions to run for my data. I'm running a model to compare the effect of foreign aid on developing countries. My dependent variable is GDP capita growth, with my independent variables being initial gdp per capita, aid/gdp, trade openness(of gdp %), foreign direct investment, population growth, violence. Aid/gdp is the main variable with the rest being controls.
Initial gdp per capita is the initial capita for the start of 5 periods. Every 5 periods it switches so 1970-1974 would have the 1970 GDP, 1975-1979 would have 1975 and so on. I wanted to just apply 1970 to all of it but it came up with the issue of multicollinearity. Every piece of literature that includes this has the variable as significant and negative, so it's important that I try and reach the same. This is because of the idea of economic convergence
The number of countries is 40, with the period being 44 years (1970-2014).
-----------------------------------------
Through a hausman test it indicated that the
fixed effects model is preferred. Running xtreg fe brings up some nice results akin to much of the literature (all significant bar violence; aid/gdp being negative and initial gdp per capita being negative). However running '
xttest3' suggests that the data is heteroskedastic, and hence I must use other models.
My first problem is this: From my research and notes I've been met with the following:
xtpcse, xtreg fe robust, xtgls and xtscc. I'm not sure on what to use. Through
xtserial, 'correlate' amongst lagged variables and
actest no autocorrelation was found.
I don't know about xtscc. From what I've read on the forums it seems that it needs to be have some level of autocorrelation with high N but low T. With mine they're medium and roughly equivalent. At the same time, running the '
xtcsd pesaran abs' command (which, according to university notes tests for contemporaneous correlation), yields pr = 0.00000, suggesting that driscoll and kraay standard errors should be used in this instance. Nonetheless, they unfortunately make all my variables insignificant bar aidgdp which yields p>|t| as 0.042
Code:
xtgls x y, panels(heteroskedastic)/
xtpcse x y
I've seen are both good. They both follow the original xtreg fe command.
unfortunately makes my initial gdp insignificant, although aid/gdp is still significant alongside FDI and trade.
---------------------------------------------
My second problem is this: from my university notes, running the
testparm for both i.year and for i.country suggest that country/time-fixed effects need to be used as pr > 0.0000. Now, running
Code:
(1) xtreg x y i.year, fe robust
(2) xtreg x y i.countryid , fe robust
(1) makes all my variables insignificant bar aid at 1% (2) omits all my country IDs but says fdi is 0.01% significant
Applying both year and countryID in one command makes my initial gdp per capita negatively significant but aid itself insignificant, which is quite confusing.
Code:
(1) xtgls x y i.year, panels(heteroskedastic)
(2) xtgls x y i.countryid panels(heteroskedastic)
both of them support the previous original xt reg fe and the original xtpcse/xtgls (where aid and initial gdp are negatively significant)
combine them and they show that aid is insignificant, yet again
Code:
xtpcse x y i.country
xtpcse x y i.year
both show negative initial gdp, although year has aidgdp at significant at 0.1 whereas country has it for 0.05.
Combining the two has aid insignificant but initial as negatively significant
So what should models would be most appropriate to use? Do I need to control for time-fixed effects? I'm just so confused and really desperate for a response
Thanks a lot
