PPML estimation without exporter-year, importer-year fixed effects

[Arthur Carvalho Brito]

Hello,

I need some help in dealing with the ppml_panel_sg command to estimate a gravity equation using the PPML method.

The issue is that my dataset on trade flows is focused on a specific country, say the US. In that sense, I only have a panel of US exports to Germany and France, and of US imports to Germany and France across the years, not on trade flows between France and Germany.

This of course precludes me from using exporter-year and importer-year fixed effects.

As an alternative, I ran the following sort of regression by OLS:

 reghdfe log_flow ln_tariff , absorb(i j) vce(robust)
        outreg2         ln_tariff  using sigma_ols.xls, append dec(3)

In which i is the exporter, j is the exporter.

Is it possible to do something similar using the ppml_panel_sg command? It appears it actually requires these rich fixed effects, so when I run the following:

    ppml_panel_sg flow ln_tariff , exporter(i) importer(j) year(year) nopair robust
    outreg2         ln_tariff     using sigma_ppml.xls, append dec(3)

I get, as expected, the following error

Initializing...
Checking for possible non-existence issues...
insufficient observations

Is there any way to bypass this?

Thank you

[Andrew Musau]

Look at ppmlhdfe from http://scorreia.com/software/ppmlhdfe/ which may allow you to achieve what you want.

[Arthur Carvalho Brito]

Andrew Musau thanks a lot, this was helpful!

[Doreen Asaba]

> ow tolerance)
Converged in 10 iterations and 38 HDFE sub-iterations (tol = 1.0e-08)

HDFE PPML regression No. of obs = 829
Absorbing 2 HDFE groups Residual df = 19
Statistics robust to heteroskedasticity Wald chi2(1) = 16.52
Deviance = 2478116374 Prob > chi2 = 0.0000
Log pseudolikelihood = -1239064851 Pseudo R2 = 0.8541

Number of clusters (importer)= 20
(Std. err. adjusted for 20 clusters in importer)
---------------------------------------------------------------------------------
| Robust
exports | Coefficient std. err. z P>|z| [95% conf. interval]
----------------+----------------------------------------------------------------
ln_tariff | 0 (omitted)
ln_gdp_pc | 0 (omitted)
ln_tot | 0 (omitted)
ln_hhi_ip | .2869129 .0706 4.06 0.000 .1485395 .4252864
ln_distance | 0 (omitted)
fta | 0 (omitted)
coastal_impor~r | 0 (omitted)
_cons | 17.42176 .1075355 162.01 0.000 17.211 17.63253
---------------------------------------------------------------------------------

Absorbed degrees of freedom:
------------------------------------------------------------+
Absorbed FE | Categories - Redundant = Num. Coefs |
--------------------+---------------------------------------|
importer_id#year | 293 293 0 *|
product_id | 6 1 5 |
------------------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation
how do i fix this

[Doreen Asaba]

i have one exporter and multiple importers so i cant use exporter FE. importer FE are also failing

[Andrew Musau]

These omitted variables should be collinear with your fixed effects. For example, GDP per capita is usually specific to a country and is reported in yearly intervals. So this will not vary with importer-year observations given that importers are situated in a single country. In addition, 20 clusters is somewhat low, so the reported standard errors may be biased. You should go back to the drawing board and consider other models and estimation techniques appropriate for your research questions and data.