Bilateral trade gravity model specification

[Ilona Mazmanidou]

Dear community,

I want to estimate a gravity model to examine the impact of interstate wars on bilateral trade flows for the period 1870-2014. I will provide a little explanation of the model to give an insight into what I am trying to do.

I specified a model as follows: ln(trade) = a + ln(dist) + war + wardur + countwar + yearcounter + power + alliance + contig + colborder + ln(cincA) ln(cincB)
All the variables are in country-pairs, except for ln(cincA) and ln(cincB), which are the national capabilities of country A and country B, respectively. The rest are specified as:

• war is a dummy variable equals 1 if country pair is in conflict at time t, 0 otherwise
• wardur is the duration of the war at time t (if war started in 1942, and we are at time t=1945, then wardur=3)
• countwar is the number of wars that the country pair has been involved at time t
• yearcounter is just a counter for the years for each country pair (not sure if I need this)
• power is a categorical variable (=1 if both countries are major powers;=2 if major/minor power;=3 if minor/minor power)
• alliance is a dummy variable which varies over time; if country-pair is in an alliance agreement at time t
• contig;colborder are time-invariant dummy variables per country-pair (I also include few other dummies not presented here)

I have searched the forum and papers for a while, but I feel more confused than ever. So my question is:
1. Are there problems with the variable specifications for the gravity model? that some are country-pair specified and some country specified?
2. Is it ok to only have country-pairs instead of origin and destiny country for the bilateral trade estimation?
3. Since fixed effects drop time-invariant variables but I would like to account for them, can I use the ppml command?

So far I have made a panel data:



And run a ppml regression, but I account for trade in levels, and not logtrade as I read here and in referred papers:

ppml trade lndist war wardur countwar yearcounter power alliance contig colborder lncinc1 lncinc2

I am totally confused which type of fixed effects I should use, country-pairs fixed effects, time fixed effects, origin and destination fixed effects?

Any help will be highly appreciated! Thank you!
P.S: Sorry for not using the Stata output, I had issues to find how to do this.

Best,
Ilona

[Joao Santos Silva]

Dear Ilona Mazmanidou,

1. Only you can know if you are included the variables that are needed to answer your research question, but it is fine to have both bilateral variables and country specific variables.
2. Again, it is up to you to decide on the fixed effects you need. However, the standard in these cases is to include origin-time, destination-time, and pair fixed effects. This means you can only identify the effect of bilateral variables that vary over time, such as war.
3. There is some confusion here, the fact that variables are dropped when you include fixed effects does not mean you are not accounting for them; indeed you still account for them and for anything else that is perfectly collinear with the fixed effects, you just do not have coefficients for them. Although I wrote the ppml command, I strongly advise you to use the ppmlhdfe command if you intend to use fixed effects.

Best wishes,

Joao

[Ilona Mazmanidou]

Dear Professor Silva,

Thank you very much for getting back to me. I will then read the corresponding papers and guides for the plhdfe to make sure that I'm on the right side.

Best wishes,

Ilona