Issues using PPML Fixed Effects Gravity Model

[Ross Connolly]

Hello all,

I am a masters student using the gravity model to try and understand trade between the former Soviet Union States between 1992-2014. I've built a composite dataset using gravity data from Head and Mayer, trade data from the Correlates of War Project, and democracy data from both Polity5 and v-Dem.

As I've researched, I discovered that the PPML estimator is commonly used to account for zero trade flows (among other reasons). I am attempting to run a PPML regression with fixed-effects dummies. These include 'exporter-time', 'importer-time' and country pair dummies. Unfortunately I am encountering a problem whereby the coefficient for distance is estimated. This should be dropped the regression due to collinearity with the dummies. Can anyone help address this? I can provide any details regarding data and commands I have used.

ppml Flow_Trade origin_time_* destination_time_* Pair_* revised_polity_o revised_polity_d eu_o eu_d CIS_o CIS_d log_distance, cluster(panel_id)

I have read as much as I could find already on Statalist about this issue, have tried running the regression with distance in various positions, however Stata is dropping multiple fixed effects dummies instead. I found this strange because the OLS regression I am also using drops the distance variable.

Thanks in advance for any assistance.

[Ross Connolly]

I know Joao Santos Silva is usually very helpful in this area.

[Joao Santos Silva]

Dear Ross Connolly,

If two variables are collinear, either can be dropped and Stata does not always drop the one you want. I suggest you use the ppmlhdfe command because it is much better at dealing with fixed effects.

Best wishes,

Joao

[Ross Connolly]

Dear Joao,

Thanks very much for the response! I have redone some of my regressions using 'ppmlhdfe' and indeed, it does seem to be working in the way that I need it to now. I initially used the following code;

ppmlhdfe Flow_Trade revised_polity_o revised_polity_d log_gdp_o log_gdp_d eu_o eu_d CIS_o CIS_d log_distance, absorb(Pair_* origin_time_* destination_time_*) vce(cluster panel_id)

But unfortunately using these particular fixed effects, Stata drops both democracy and GDP figures due to multicollinearity - I am assuming this is because of the country-time fixed effects I am using. Is it the case that I cannot use these fixed effects because I am investigating a country-specific factor at a given time? The following code works perfectly;

ppmlhdfe Flow_Trade revised_polity_o revised_polity_d log_gdp_o log_gdp_d eu_o eu_d CIS_o CIS_d log_distance, absorb(Pair_*) vce(cluster panel_id)

My only concern would be that I am not controlling for enough unobserved heterogeneity.

A further question; my sample only includes the fifteen former soviet states' trade with one another between 1992-2014. I know in another thread that you said gravity models are not optimal for smaller country sample sizes, however would it be valid in this framework that if a negative coefficient is found for the democracy variable, this can simply be interpreted that more democratic states have traded less amongst their former Soviet neighbours (whilst not precluding them from trading more with countries outside the sample)?

Many thanks,

Ross

[Ross Connolly]

Sorry Joao Santos Silva, I meant to tag you in that previous post. Still getting the hang of Statalist!

Best wishes,

Ross

[Joao Santos Silva]

Dear Ross Connolly,

Indeed, in the first equation the effect of variables such as GDP cannot be identified. Whether or not to include the fixed effects is a modelling decision, so it is up to you.

You are correct on the second question; the results have to be interpreted in the context of your sample.

Best wishes,

Joao

[Ross Connolly]

Dear Joao Santos Silva,

Many thanks for all your help!

Best Wishes,

Ross

[Ann Steele]

Hi can I apply this to any topic? like health economics?

[Joao Santos Silva]

Yes!