fixed effect with gravity model: insignificant variables

[aylin ayla]

Dear all,

I have to do a gravity model in Stata using fixed effect for my thesis. My director asked me to use time fixed effect and country fixed effect in my regression. So, i have used this following regression on Stata:

xtreg lntotexp lnpibass lnpopass lnce agoa_dum2retarde sra_dum2 omc_dum2 i.year i.country1, fe

I want to estimate the effect of these different variables on the total export from sub-saharan Africa countries to the United states.
In my result on Stata, the country fixed effect are omitted and other variables are really insignificant.
So I tried to do it again without the country fixed effects:

xtreg lntotexp lnpibass lnpopass lnce agoa_dum2retarde sra_dum2 omc_dum2 i.year, fe

Again, all the variables are insignificant (except the lnpibass). So, i would like to know how to correct this insignificance problem and omitted variables problem when I use year and country fixed effects in my regressions?

There is three dummy variables in regression, and AGOA is delayed by one year (because agreements take time to translate into increased exports) by using this code:

gen agoa_dum2retarde=agoa[_n-1]

Data goes from 1992 to 2017 and take account 49 countries.

Best regards,

Aylin

[Clyde Schechter]

Well, given that you got any results at all out of your -xtreg- command, you must have at some point -xtset- your data. I'm guessing you did so as -xtset country year- or perhaps just -xtset country-. Is that right?

If so, then when you run -xtreg- with the -fe- option, the country effects are automatically accounted for, and your attempt to also specify them explicitly with i.country is redundant. Stata recognizes the redundancy and, helpfully, removes those variables for you.

Next, your code

gen agoa_dum2retarde=agoa[_n-1]

is incorrect, because when it applies to the first observation for any country, it is setting the delayed variable to the value of agoa for the preceding country in its final year. That's clearly not what you want. In fact, there is no need to calculate this lagged variable explicitly anyway. The error-proof way to do this, and shorter at that, is:

xtset country year
xtreg lntotexp lnpibass lnpopass lnce L1.agoa sra_dum2 omc_dum2 i.year, fe

The L1 operator tells Stata to use the one time-unit lagged value of the variable agoa. It does this "on the fly" without your having to actually create a separate variable. Read -help tsvarlist- for other useful operators that apply to panel data.

Again, all the variables are insignificant (except the lnpibass). So, i would like to know how to correct this insignificance problem

It's not a "problem." It is what it is. If the variables do not have statistically significant coefficients, then the data simply do not identify the coefficients precisely enough to give you much confidence about their signs. It's a fact, not a problem to be fixed. In fact "fixing" this problem would fall under the heading of "torturing the data until they confess" and is not science, but, in the eyes of some, a form of scientific misconduct.