[Richa Khurana]

Dear Sir,

Thank you so much for your help!

Best regards
Richa

[michael cheuk]

Hi Joao,

Thanks for your response.

I am still struggling with the fixed effects dummies, I have tried to create them for year and countries, yet when I run my regressions using something like...

ppml exports ldist lgdp_swe lgdp_part comlang border bothin dimp* dyear*, cluster(importer)

...the time constant factors such as distance, comlang and border are still not dropping out of my regression. Hence I must be generating the fixed effects wrong.

Are you able to provide me with Stata code on how to generate these year and country dummies, I am currently using -

qui tab importer, g(dimp)
qui tab year, g(dyear)

Kind regards,

Michael

[Joao Santos Silva]

Dear Michael,

The fact that those variables are not dropped does not mean that you are doing something wrong; maybe you are just not understanding your results. Consider the following example.

cls
clear all
sysuse auto
reg price mpg rep78 i.rep78
reg price mpg i.rep78 rep78

The variables dropped in the two cases are different, but it does not matter which is dropped because the only coefficient that matters is the same in both cases.

Best wishes,

Joao

[michael cheuk]

Thanks again Joao! It is much appreciated.

Michael

[Richa Khurana]

Hello Sir,

I am still getting different results from the estimation using ppml and ppml_panel_sg.
The ppml does not drop the the trade creation and diversion variables that I have introduced while ppml_panle_sg does, while estimating using origin and destination time varying fixed effects.
Sir will it be possible for you to have a look at my results via email?

Thank you very much.

Best regards
Richa

[Joao Santos Silva]

Dear Richa,

It would be better if you could post your results here so that others can comment as well.

Best wishes,

Joao

[Richa Khurana]

Dear Sir,
Please find attached the link to my results. The file was too big to upload.
I am sorry for the inconvenience.

Thank you
Best regards
Richa

[Joao Santos Silva]

Dear Richa,

Making results available on external sources is not a good practice; you could just paste here the relevant parts. Anyway, it looks like you are not including "border" in the -ppml- estimation.

Best wishes,

Joao

[Richa Khurana]

Dear Sir,
I am extremely sorry for this. I will post the relevant parts of the results here.
Would the inclusion of border variable alter the dropping of trade creation and diversion variables?

My main problem is that the way I have introduced the trade creation and trade diversion (import and export) variables, the -ppml- command does not drop these while the -ppml_panel_sg- command does. What is wrong with my estimation, that different results are being generated?

The results from both are as follows:


ppml_panel_sg trade_rescale TC_ij TD_iin_jout TD_iout_jin ln_DISTANCE BORDER LANG_off COLONY, ex(exporter) im(importer) y(year) nopair

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

Number of obs = 44425
Log likelihood = -203037.42 R-squared = 0.7146
-------------------------------------------------------------------------------
trade_rescale | Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
TC_ij | 0 (omitted)
TD_iin_jout | 0 (omitted)
TD_iout_jin | 0 (omitted)
ln_DISTANCE | -.7160885 .035974 -19.91 0.000 -.7865963 -.6455808
BORDER | .8110483 .0929958 8.72 0.000 .62878 .9933167
LANG_off | .8219184 .1016943 8.08 0.000 .6226011 1.021236
COLONY | -1.12261 .1766887 -6.35 0.000 -1.468913 -.7763064
-------------------------------------------------------------------------------
Fixed Effects included: exporter-year, importer-year
Robust standard errors (default)
23300 obs. dropped because they belong to groups with all zeros or missing values

TC_ij appears to be collinear with your set of fixed effects

TD_iin_jout appears to be collinear with your set of fixed effects

TD_iout_jin appears to be collinear with your set of fixed effects



ppml trade_rescale exp_time_fe* imp_time_fe* ln_DISTANCE LANG_off COLONY TC_ij TD_iin_jout TD_iout_jin, cluster(pair_id)

Number of parameters: 2446
Number of observations: 86860
Pseudo log-likelihood: -555775.97
R-squared: .6794458
Option strict is: off
(Std. Err. adjusted for 6158 clusters in pair_id)
---------------------------------------------------------------------------------
| Robust
trade_rescale | Coef. Std. Err. z P>|z| [95% Conf. Interval]
----------------+----------------------------------------------------------------
ln_DISTANCE | -.979299 .0933645 -10.49 0.000 -1.16229 -.7963079
LANG_off | .8983399 .2905224 3.09 0.002 .3289265 1.467753
COLONY | -.6977831 .4677202 -1.49 0.136 -1.614498 .2189317
TC_ij | .5604474 .7795663 0.72 0.472 -.9674744 2.088369
TD_iin_jout | -4.057525 .8491111 -4.78 0.000 -5.721753 -2.393298
TD_iout_jin | 2.201877 1.044074 2.11 0.035 .155529 4.248225
_cons | 7.236376 .9732081 7.44 0.000 5.328923 9.143828
---------------------------------------------------------------------------------

[Joao Santos Silva]

Dear Richa,

The think that is important to compare is the coefficient of distance, and that will change if you omit the border dummy. If you estimate the models with the same variables you should get the same results for the variables whose coefficients are identified; for the others you may get different results and different variables being dropped, but that is irrelevant.

Best wishes,

Joao

[Richa Khurana]

Dear Sir,

Thank you for the response. This means that the results for identified variables will be the same but those which are not identified (like the TC TD variables) can vary or dropped.
Am I understanding this correctly Sir?

One more question, in the paper "log of gravity", the MRTs have been controlled for by introducing exporter and importer specific fixed effects in a cross-sectional setup. But if panel data were to be used will it be wrong to control for MRTs the same set of fixed effects? In that case would Exporter and importer time varying fixed effects correctly capture the MRTs?

Thank you very much

Best regards
Richa

[Joao Santos Silva]

Dear Richa,

You are right in both cases; with panels the MRT should be time varying.

Best wishes,

Joao

[Richa Khurana]

Dear Sir,

Thank you for clarifying all the doubts.

Thank you very for your time and patient guidance.

Best regards
Richa

[Anthony Macedo]

Dear forum members,

First of all, thanks to João for all the answers in this forum, they helped me to clarify some questions.

I have several questions, but let me explain what I am doing. I am estimating the determinants of exports of a specific product from one country to its main trade partners. The data is 10 years long and I have a sample of 20 destination countries that represent around 90% of all exports.

So, now i have some questions:
1- I should not use ppml command, right? I think João already suggested to someone xtpoisson or xtpqml, but are 200 observations enough to do that?

2- If I use the option fixed effects I will not be able to estimate time-invariant variables like distance or language. What can I do to estimate them? To use random effects?
I previously estimated a Random effects model and a Fixed effects model and the Hausman test pointed preference for the FE model.
Can I estimate:
Code:
xtpoisson value lgdppc ler ltar ldist lang land Prod10 emig yr*
and code:
xtpoisson value lgdppc ler ltar ldist lang land Prod10 emig yr*, fe (it will drop ldist, lang and land)
then let Hausman postestimation command determine which is the best?

Also, I know that there is the possibility of not estimating them and just assume that they do not have
additional information.

Thank you for your attention.

[Joao Santos Silva]

Dear Anthony,

Why are you using data only for 20 countries? It will be difficult to interpret the results you get with such small and non-random sample.

Best wishes,

Joao
PS: Forget about the RE estimator in this context.