[Joyce Wu]

Dear Joao,
Thank you for your prompt reply！
i think the only different variable between the two estimators ,is the dependent variable(value),in FE model, the dependent variable is" ln_value", while in PPML, in order to control the zero value, i used "value "as dependent variable.
But,which was very weird, the results are the same except for constant and R-square.

I think maybe the effects are same when i controled yeardummy(7years) in FE and controled yeardummy & iddummy(over 7000id) in PPML ,but actually I don't know the reason.
I have uploaded the results( I have deleted iddum in the PPML part)
and i wonder that if the iddummy setting was wrong or not

egen id=group(reportingcountry partnercountry)
tab id,gen(iddum)

I'm sorry for long questions,and i'd appreciate your help

best regards

Joyce

[Joao Santos Silva]

Dear Joyce,

This cannot be right; please post the Stata log file, OK?

Joao

[Joyce Wu]

Dear Joao,
It seems that my data included minus values while I didn't drop them.
so that's way even I regress PPML after FE, the outreg2 save the result of FE.
now I'll drop the minus values and try again.
Anyway,Thanks for your response
Best regards

Joyce

[Gabriel Arrieta]

Dear Joao,

I just participated, previously, with others comments in this forum abut the using of ppml estimator.

I had some questions about the ppml estimator. I just have estimate my models, first my ppml estimation of the classic gravity model and my second estimation is about the augmented gravity equation estimation. In those cases I noted that my R-squared are too small. I don't understand why is this happen. How I can read the outputs in case of that R-squared level?

P.S: I attached my outputs.

Thank you very much Mr. Silva.

[Joao Santos Silva]

Hi Gabriel,

From your results, it looks as if you are not taking logs of the regressors.

All the best,

Joao

[Gabriel Arrieta]

Good afternoon Mr. Santos Silva,

So, Should I use the log of GDPs and log of distance? I just did the rescalling divided by millons or thousands. Then, my other regressors are dummies variables (1 or zero value depend on the conditions).

Do you think with these changes of logs of the regressors could better my R-squared? or I have to read my R-squared differently, because is a poisson-based model?

Thanks for your answer, Mr. Santos Silva.

Gabriel

[Gabriel Arrieta]

Good afternoon Mr. Santos Silva,

I was taking logs of the regressors in my new estimation, but the PPML continues give me a small R-squared (nearly that my previous post where I showed my results).

I had a question about the R-squared: In other pages I read that when I use a Poisson-based model, I dont have to use the R-squares, is it correct? Or I have to take in mind the R-squared as the PPML report?

Thanks you very much Mr. Santos Silva,

Gabriel

[Joao Santos Silva]

Dear Gabriel,

Sorry for missing your earlier post.

In short, your dependent variable should not be logged, but all the continuous regressors should be logged. Generally the R2 can be ignored but this type of regression generally have large values of the R2 (as computed by -ppml-) and therefore a very low value is surprising and may indicate that you are doing something incorrectly and it is worth making sure that all is well. What is m_trade?

Joao

[Gabriel Arrieta]

Good evening Mr. Santos Silva,

My variable "m_trade" is my Peru's bilateral trade level (exports and imports) in millons with other countries -I did this transformation because the estimator told me that-

I have been repeated the estimation but now with logs of my regressors; nevertheless, my R2 did not changed too much and countinues being small.

[Joao Santos Silva]

Gabriel,

Do the results change if you change the scale of the dependent variable? Are GDP and distance in logs? Their coefficients are very small.

Best of luck,

Joao

[Gabriel Arrieta]

Good evening Mr. Santos Silva,

The results don't change when I change the scale of the dependent value. The only which change is the constant.

Then these are my results when I'm using GDPs and distance in logs. In this case, I'm attaching 4 estimations with PPML:

i) Classic Gravity Model (with m_trade in millions as a dependent variable) (independent: Logs of GDPs and distance)
ii) Classic Gravity Model (with mil_trade in thousands as a dependent variable) (independent: Logs of GDPs and distance)
iii) Augmented Gravity Model (with m_trade in millions as a dependent variable) (independent: the rest of variables)
iv) Augmented Gravity Model (with mil_trade in thousands as a dependent variable) (independent: the rest of variables)

As you can see my R-squared continues being too small for each estimation, but I have all my regressors significant at 1% and all the signs are OK.

I wish you can help me with my R-squared problem. Or what can I do to interpret that R-squared? Or maybe have I take account that in Poisson-based models I don't have to use the R-squared? (http://www.ats.ucla.edu/stat/stata/o...son_output.htm)

Thank you very much for your disposition to help me, Mr. Santos Silva,

[Joao Santos Silva]

Dear Gabriel,

The results look OK to me. As you may know, the R2 is never important, so I would not worry to much about it. As I said above, the only concern I had is that the unusually low value could indicate that something was incorrect, but the results you provided suggest that all is OK.

All best wishes,

Joao

[Gabriel Arrieta]

Good evening Mr. Santos Silva,

Thank you very much for your advices and help.

I have the last question about ppml applications. Do you have any paper for reccomend me about the application of ppml on gravity model estimation apart from "The Log of Gravity"? I would like to read more papers about this topic and I would like to know your reccomendations of the applications post your paper with the presentation of ppml estimation.

Thank you very much, again, for your disposition to help us with our doubts and problems.

Have a great day,

Gabriel

[Joao Santos Silva]

Dear Gabriel,

The original paper and the information we have on our webpage dedicated to it are the sources I recommend. There are plenty of other papers using PPML, but unfortunately not everything that has been written about it is correct (although a lot is, of course).

All the best,

Joao

[Gabriel Arrieta]

Good afternoon Mr. Santos Silva,

Thank you very much for the recommendations.

All the best,

Gabriel