[Joao Santos Silva]

Dear Hashim,

Do not transform the dependent variable. If the results with trade as the dependent variable are not sensible, that is probably because your moder does not have the right regressors or because your sample is not representative (too few importers?).

Best wishes,

Joao

[Cesar Huaman]

Deas Joao,

I posted before in this forum and I would like your advise one again. I am trying to measure the impact of the TLC between Perú and USA in extensive and intensive margins of the exports of Perú. I have data from the 50 mains partners of Perú. I read here that use "xtreg, fe" and "xtreg, re" are not correct instead of that you suggest to use "xtpoisson" or "xtpqml". My question is if I use this two last estimator I have to use Randon Effect and Fixed Effect options or just one of both. Besides, if I have to use both options then I have to use Hausman Test to choose which is better? or there ir another test in this case?

Thank you very much for your help!

Best regards,

César

[Joao Santos Silva]

Dear César,

I would not recommend using RE with Poisson regressions. So, you can just use the commands you mentioned with the FE option, or -ppml- with all the dummies; the results should be the same.

Using just data for the 50 largest partners may not be a good idea. So, either you get more data or you need to be very careful with the interpretation of your results.

Best wishes,

Joao

[Cesar Huaman]

Dear Joao,

First of all, thanks for your answer. I have a doubt whit that. Why RE with poisson regressions are not recommended if the results are similars? In the other hand, about my data, those countries represents about 98% of the exports of Perú. Despite of that would you recommend more countries? The reason its we need the 100% countries to be more realistics results?

Thank you very much for your help!

Best regards,

César

[Joao Santos Silva]

César

The RE and FE results can be very different. What I said is that you have 3 different ways to estimate the model with FE.

If you do not use all the countries you can only make inference about exorts to those countries and cannot use your model to predict trade with counties out of the sample. That is because you do not have a random sample of the population, but a very selected sample.

Best wishes,

Joao

[Hashim Akin-Olagunju]

Dear Joao,

Thanks for the observation. I will look into the data.

Best Regards,

Hashim

[Cesar Huaman]

Dear Joao,

thanks for clarifying my doubts. What I would like to know is, in general, why in trade literature (ppml) is recommended estimate just with FA and not with RE. Another consult, which tests I could apply in my estimation.

Thank you very much for your help!

Best regards,

César

[Joao Santos Silva]

Dear Cesar,

In the PPML RE estimator does not have the robustness properties of the standard PPML or PPML FE estimators, and therefore it is not attractive. The test I would recommend is the RESET.

Best wishes,

Joao

[Cesar Huaman]

Dear Joao,

Thanks for all your help.

Best wishes,

César

[Mulei Leonard]

Hi Joao,
Thanks for your helpful comments in this forum. How can I estimate a gravity model when the independent variables are integrated at different levels {i.e I(0) & I(1)}? When I use the following command:-
. ppml lnExp lnY_it lnY_jt lnDist lnREER_jt d.lnfdi_it d.lnfdi_jt top_jt d.lntpi_it d.lntpi_jt d.pol_i d.pol_j EAC

I am getting this error message:-

"factor variables and time-series operators not allowed"
r(101);

Please advise on the way forward.
Thanks

[Joao Santos Silva]

Hi Mulei,

First of all, note that the dependent variable for PPML should not be in logs; that is the point of using PPML.

Also, the -ppml- command does not allow the use of factor variables, so you should first create all the dummies and then include them in the model.

Finally, unless you have a a panel with a very large time dimension, you do not need to worry about orders of integration.

Best wishes,

Joao

[Dilshat Obul]

Dear Joao,

I am wondering that is that necessary to add pair fixed effect dummies to the structural gravity model when we using panel data with intervals?
for my case, I am doing my research on interprovincial trade in china using gravity model and industry level panel data with intervals. key variable is time varant distance. there are 31 province and each province has 42 industry, so this a 31*42*31*42 trade matrix data. so the exporter is r province's i industry ( there are 31*42 exporter) and the same number of importer. am i right?

Best regards,
Dilshat

[Joao Santos Silva]

Dear Dilshat,

I am not an expert on trade, but in the literature you find gravity equations with and without pair fixed effects.

Best wishes,

Joao

[Mulei Leonard]

Dear Joao,
THanks for your observation on the use of logs on the dependent variable. additionally, I dont have factor variables in my data set, I have converted them into dummy variables. However, when I use the XTPMG command.... I get the following
. xtpmg exports d.lnY_i d.lnY_j d.lngdppcdif d.reer_jit d.top_j d.lntpi_i d.lntpi_j d.pol_i d.pol_j EAC contig comlan
> g_off,lr( exports lnY_i lnY_j lngdppcdif lnDist reer_jit top_j lntpi_i lntpi_j pol_i pol_j) ec(ec) full pmg

Iteration 0: log likelihood = -5369.9315 (not concave)
Iteration 1: log likelihood = -5368.0443 (not concave)
Iteration 2: log likelihood = -5364.9182 (not concave)
numerical derivatives are approximate
flat or discontinuous region encountered



this continues forever, What could be the cause for this and how can I address this issue?!


On the use of Xtpoisson command, I get results which don't agree with the gravity model theory. e.g. the GDP of the importer is Negative. Could this be issues with my data (i.e collinearity)? How can I move out of this problem? See below the output of results window. Please advise.


. xtpoisson exports lnY_i lnY_j lngdppcdif lnDist reer_jit lnfdi_i lnfdi_j top_j lntpi_i pol_i pol_j EAC contig comlan
> g_off
note: you are responsible for interpretation of non-count dep. variable

Fitting Poisson model:

Iteration 0: log likelihood = -3.661e+08
Iteration 1: log likelihood = -2.476e+08
Iteration 2: log likelihood = -2.291e+08
Iteration 3: log likelihood = -2.289e+08
Iteration 4: log likelihood = -2.289e+08
Iteration 5: log likelihood = -2.289e+08

Fitting full model:

Iteration 0: log likelihood = -1.580e+08
Iteration 1: log likelihood = -1.374e+08
Iteration 2: log likelihood = -1.370e+08
Iteration 3: log likelihood = -1.369e+08
Iteration 4: log likelihood = -1.369e+08
Iteration 5: log likelihood = -1.369e+08
Iteration 6: log likelihood = -1.369e+08
Iteration 7: log likelihood = -1.369e+08
Iteration 8: log likelihood = -1.369e+08
Iteration 9: log likelihood = -1.369e+08
Iteration 10: log likelihood = -1.369e+08
Iteration 11: log likelihood = -1.369e+08

Random-effects Poisson regression Number of obs = 440
Group variable: panel Number of groups = 20

Random effects u_i ~ Gamma Obs per group:
min = 22
avg = 22.0
max = 22

Wald chi2(14) = 2.53e+08
Log likelihood = -1.369e+08 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
exports | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lnY_i | 4.768366 .0009079 5252.27 0.000 4.766587 4.770145
lnY_j | -1.616649 .0006971 -2319.25 0.000 -1.618015 -1.615283
lngdppcdif | -.2917383 .0002653 -1099.68 0.000 -.2922583 -.2912184
lnDist | -1.304272 1.482922 -0.88 0.379 -4.210745 1.602201
reer_jit | -.4596659 .0001874 -2453.08 0.000 -.4600332 -.4592986
lnfdi_i | -.5606175 .0001661 -3374.73 0.000 -.5609431 -.560292
lnfdi_j | .1874802 .0001696 1105.43 0.000 .1871478 .1878126
top_j | 1.62184 .000765 2119.93 0.000 1.620341 1.62334
lntpi_i | .0817481 .000461 177.32 0.000 .0808446 .0826517
pol_i | .0678443 .0003072 220.81 0.000 .0672421 .0684465
pol_j | .9233298 .0002576 3583.72 0.000 .9228248 .9238348
EAC | -.1860944 .0002725 -683.01 0.000 -.1866285 -.1855604
contig | -3.245779 1.420987 -2.28 0.022 -6.030863 -.4606961
comlang_off | -2.683274 1.521648 -1.76 0.078 -5.665648 .2991002
_cons | -31.45454 10.86362 -2.90 0.004 -52.74685 -10.16223
-------------+----------------------------------------------------------------
/lnalpha | 1.848907 .2389509 1.380572 2.317242
-------------+----------------------------------------------------------------
alpha | 6.352874 1.518025 3.977177 10.14765
------------------------------------------------------------------------------
LR test of alpha=0: chibar2(01) = 1.8e+08 Prob >= chibar2 = 0.000


Thanks in advance.

[Joao Santos Silva]

Dear Mulei,

I do not think -xtpmg- is appropriate in this context, so just ignore that. As for -xtpoisson- you need the fe option.

Best wishes,

Joao