PPML estimation and Fixed-effects in gravity model

[Madhuri Kumari]

I am trying to see the impact of two RTAs ( both between two countries only) on trade. There are a total of six RTA dummy variables including the import and export diversion RTA dummies. The dataset has about 5 million observations. Will it be ok to do include only Exporter-product-time and Importer-product-time fixed effects, and skip Exporter-Importer-product FEs?
When I try to include Exporter-Importer-product FEs, the results turn insignificant.

[Joao Santos Silva]

Dear Madhuri Kumari,

A first point I want to make is that we should not choose the model specification based on how convenient the results are. Anyway, the exp-imp fixed effects are generally used to account for the fact that trade agreements are potentially endogenous, so they are generally included in these models. Another point to note is that if you are using separate dummies for each RTA, they may have very few observations equal to 1, and that will make inference unreliable.

Best wishes,

Joao

[Madhuri Kumari]

Thank you Joao Santos Silva for this answer.
I would still like to ask you for my further confusion.
I am trying to analyze the diversion effects. I have the following result with this command for my aggregate trade. The time period of the study is 1995 to 2019. Here, fta1 was signed in 2010, and fta2 was signed in 2011.

ppml_panel_sg trade fta1 imp_fta1_rest_ exp_fta1_rest_ fta2 imp_fta2_rest_ exp_fta2_rest_ , ex( iso3_o ) im( iso3_d ) y( year )
Initializing...
Checking for possible non-existence issues...
note: imp_fta1_rest_ omitted because of collinearity over lhs>0 (creates possible existence issue)
note: exp_fta1_rest_ omitted because of collinearity over lhs>0 (creates possible existence issue)
note: imp_fta2_rest_ omitted because of collinearity over lhs>0 (creates possible existence issue)
note: exp_fta2_rest_ omitted because of collinearity over lhs>0 (creates possible existence issue)
Iterating...

---------------------------
Variable | active
-------------+-------------
ijcepa | .1002447
ikcepa | .0374814
---------------------------
iterations: 39
tolerance: 1.000e-12
Computing standard errors

******* PPML Panel Structural Gravity Estimation **********

Number of obs = 486,021
Log likelihood = -6.489e+09 R-squared = 0.9935
-------------------------------------------------------------------------------------
trade | Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------------+----------------------------------------------------------------
fta1 | .1002447 .1038665 0.97 0.334 -.1033298 .3038192
fta2 | .0374814 .0955838 0.39 0.695 -.1498594 .2248222
-------------------------------------------------------------------------------------
Fixed Effects included: iso3_o-year, iso3_d-year, iso3_o-iso3_d
Clustered standard errors, clustered by iso3_o-iso3_d (default)
944 obs. dropped because they belong to groups with all zeros or missing values


My diversion dummy variables are getting omitted. I am unable to understand why this is happening?
(i) is this happening due to the fewer observations equal 1?
(ii) even if so is the case, how can i solve this issue?
(iii) Is there anything wrong with my model?

[Joao Santos Silva]

Dear Madhuri Kumari,

They are getting omitted because they are perfectly collinear with the fixed effects. By the way, the command you are using is obsolete, you should use ppmlhdfe.

Best wishes,

Joao