PPML Estimator questions

[Alejandro Munoz Saez]

Good morning,

I'm performing my research in asset trade having a panel data of 50 (i) x 167 (j) x 5 years (t) and I would like to use the PPML estimator.

My model looks as follows:

X_ij,t= exp [ B_1 ln Dist * B_2ln GDP_i * B_3 ln GDP_J * B(other i and j time variant variables in levels) * B(other global time variant variables like global GDP growth in leves).

My questions are:

1) Would econometrically correct to perform the PPML without fixed effects (so I can keep time invariant variables such as distance) at least as a starting point for my stimation

The stata command that I'm using here:

ppml X ln_DIST ln_GDP_i ln_GDP_j (other i and j time variant variables) (other global time variant variables like global GDP growth) if exporter != importer, cluster(pair_id)


2) Would it be correct to do the same thing as point 1) but including pair fixed effects? All the variants would be interpretable?

ppml X ln_DIST ln_GDP_i ln_GDP_j pair_id (other i and j time variant variables) (other global time variant variables like global GDP growth) if exporter != importer, cluster(pair_id)


3) For the PPML taking into account multilateral resistance:
ppml X EXPORTER_TIME_FE* IMPORTER_TIME_FE* ln_DIST ln_GDP_i ln_GDP_j pair_id (other i and j time variant variables) (other global time variant variables like global GDP growth) if exporter != importer, cluster(pair_id)

Only distance would be interpretable, because the country time variant variables are dropped right?

4) Is there any additional thing that I should account for with panel data with T dimensions?

5) Is there any problem if I dont have values for the the independent variables for some of the countries in some years, or should I have a balance panel?

If you would be so kind as to help me with my questions professor Dr. Joao Santos Silva it would be much appreciated.

Thanks and best regards

[Joao Santos Silva]

Dear Alejandro Munoz Saez

1) That is a modelling decision, so only you can decide.

2) In that case you cannot identify the effect of distance and other bilateral variables.

3) Correct.

4) May main advice is that you use the ppmlhdfe command instead of ppml because it is much better at dealing with fixed effects.

5) You do not need a balance panel, but are those observations really missing or are they not reported because they are zero? If that is the case, you should replace the missings with zeros.

Finally, you really should discuss these questions with your adviser.

Best wishes,

Joao

[Alejandro Munoz Saez]

Dear profressor Dr Joao Santos Silva,

thanks for your quick and wise advice.

Regarding point 5) they are missing, since they are regarding the independent variables, for examples, Albania's GDP is only available from half of the sample, therefore, I include it for the period that is available, while for the rest of the sample, I include them for the full sample. Following your advice, balance panel is not needed.

Thanks and best regards

[Joao Santos Silva]

Dear Alejandro Munoz Saez

Note that if you include time-origin and time-destination fixed effects those variables are not needed and therefore you do not need to drop those observations.

Best wishes,

Joao

[Alejandro Munoz Saez]

Thanks for your clarification, therefore, I will only drop those samples with missing values in the case that time-origin and time-destination fixed effects are not included.

Again, thanks for your help!