Country-pair fixed effects in gravity

[Nina Dolmer]

Hi all,

I'm investigating exports to the Central and Eastern European countries from the rest of Europe. I have panel data over the last 20 years in my survey.
I have conducted a hausman test which tells me that I should use fixed effect (as expected).
However fixed effects omits my time-invariant variables, which I am interested in keeping in the equaiton, so I have tried to do country-pair fixed effects, because I have read that it is then possible keep the time-invariant variables.

But, when I try to conduct it in Stata it still omits my variables (But it works with OLS, Poisson and Random effects).

I grouped it:
egen pairid = group(country_ceec partner)

So far I have tried (and only used distance as a time-invariant variable):

xtreg imports_1000 lgdp_ceec lgdp_partner ldist offlang year*, fe robust

xtreg limports lgdp_ceec lgdp_partner ldist year_* exporter_*, fe cluster(pairid)

xtreg limports lgdp_ceec lgdp_partner ldist offlang year* exporter_*, fe vce(robust)

Do anyone know what I am missing in this, since distance keeps being omitted? I would appreciate if anyone had a solution.

Thanks in advance,

Nina

[Marcos Almeida]

Welcome to the Statra Forum / Statalist.

Please take a look at the FAQ, particularly on how to share command and data. Thanks.

That said, and considering you know a fixed-effects model will do towards time-invariant variables precisely what you have, I wonder whether a between-effects model would fit in.

[Nina Dolmer]

Sorry for not reading the FAQ.
As I understand a country specific fixed effects would still include the time invariant variables, is that not correct?
I have actually not considered the between-effects model, I will look into that, Thanks.

[Tom Zylkin]

Hi Nina,

Just to clear up any confusion, it is not possible to include time-invariant variables (such as log distance etc) in the presence of pair fixed effects. These fixed effects absorb all time-invariant variation that is specific to each pair, which would typically include distance.

Hope the above is helpful.

Regards,
Tom

[Sebastian Kripfganz]

Time-invariant regressors in panel models have been discussed many times here on Statalist. The forum search (or a Google search) should direct you there. As a starting point, my comment in the following Statalist discussion and the links therein (not just about system GMM but fixed-effects models in general) might be useful:
How to treat time-invariant variables in system GMM?

In addition, you might find the discussion in my working paper with Claudia Schwarz useful. As an application, we consider a dynamic gravity model with a special focus on how to estimate the coefficient of the time-invariant distance variable:

• Kripfganz, S. and C. Schwarz (2015). Estimation of linear dynamic panel data models with time-invariant regressors. ECB Working Paper 1838, European Central Bank.