[Mario Scintille]

Dear Joao,

thanks a lot for your answer and the very useful reference. If you have time, I would like to ask you one additional questions. The post above says that using negative binomial for non-negative values is not standard, I guess you mean "for continuous variables" , correct?

[Joao Santos Silva]

Indeed, I meant for general non-negative data; it is standard for the particular case of counts.

Best wishes,

Joao

[Borworn Tanrattanaphong]

Hi Joao,

I'm estimating the gravity model. My data is panel data. My dependent variable is trade volume. After I have read a number of literature review, especially papers of Prof Santos Silva, J.M.C. So, I decide to use PPML. I use the STATA command - xtpoisson - and add the option - fe -. And, the result is quite good since many independent variables are significant. However, I still confuse some issues as follows:

1. Since the gravity model is generally presented in a nonlinear form and can be presented in a linear form by taking logarithm. So, do I need to check the multicollinearity problem?

2. Since my data is panel data, contained 194 countries and 19 years, do I need to check the autocorrelation problem?

3. I try to add the option vce(robust) in the STATA command -xtpoisson dependent var independent var, fe -. However, I found that the standard error is larger. As far as I know, after correcting SE, the SE should be smaller. So, should I add the option - vce(robust) after -xtpoisson dependent var independent var, fe ?

4. I try to use RESET test following some guidance presented at http://personal.lse.ac.uk/tenreyro/lgw.html. However, if I don't correct SE, the result always shows that the model is not appropriate (Reject H0). And, if I correct SE, the result shows that the model is appropriate. However, as I mention in the number 3, SE become larger after using the option vce(robust). So, do I need to test for model specification by using RESET test. And are there any test that can I use instead of RESET test?

5. If I want to compare the result performance of PPML and Negative Binomial PML, how can I do that? Can I use "hpc test"? Or, are there any test can do that?

Many thanks in advance

Regards
BT

[Mario Scintille]

Dear Joao, let me join the many that have already express their gratitude for your help, I really appreciated your time and advices. Muy obrigado.

Mario

[Joao Santos Silva]

Dear Borworn Tanrattanaphong ,

Here are some answers to your questions:

1 - No, there is no need.
2 - No, there is no need.
3 - You should use robust (clustered) standard errors; these are often larger that the default standard errors.
4 - The RESET is a good test and it looks as if your model passes it when you use the appropriate standard errors.
5 - The NB model is clearly not suitable in this context, so no need to test.

Best wishes,

Joao

[Borworn Tanrattanaphong]

Dear Joao Santos Silva

Many thanks for your reply. I have spent a lot of time on literature review and received some suggestion from you, and found that up to now PPML is the best method to deal with the gravity model in the international trade field. So, I will go ahead with PPML and distribute this notion to others as well. Thank you so much again.

Have a nice day
BT

[Andrew Peterson]

Joao Santos Silva
Hi Joao (and others)

Thank you so much for your input on these forums. Joao, you have previously mentioned the following (25 March 2014):

Dear Tereza,

I would say that the standard in the trade literature is to estimate (by ppml) a model with importer and exporter dummies interacted with time dummies. That is, to include a full set of importer and exporter dummies for each year, but not country-pair dummies. Richard Baldwin has written about this.

All the best,

Joao

Has this standard changed in the last four years or so, or do you consider that this is still the most suitable method for a model that is used to consider trade creation and trade diversion effects of trade agreements?

Regards
A

[Joao Santos Silva]

Dear Andrew,

I believe this is still the standard, but note that I am note a trade economist.

Best wishes,

Joao

[Said Jafar]

Dear Andrew,

I would like to add a bit to what Mr Joao said.

Estimating the model with only importer and exporter fixed effects produce theoretically consistent results ala AvW (2003). Estimating the model with importer-time and exporter-time fixed effects will produce theoretically consistent results and will also account for time specific effects in the countries (like economic turmoil and so).

But these fixed effects will not account for the endogeneity of the FTAs, which can be accounted for by including pair fixed effects in the model. As you may know, this will drop all time invariant variables (like distance). See the paper cited below for more details.

Baier, S. L. & Bergstrand, J. H. 2007. Do free trade agreements actually increase members' international trade? Journal of International Economics 71(1): 72-95.


Hope this helps,
Said

[Andrew Peterson]

Thank you Joao Santos Silva and Said Jafar

[Jaime Giacomozzi]

Hi Joao Santos Silva

I'm working on my masters thesis using gravity and panel data on dairy trade flows. I wanted to ask you a couple of questions.

1. Would it be valid to estimate my model using ONLY IMPORTS from just one particular country as dep variable? i.e. country X imports from country Y in year 1,2,3..... country X imports from country Z in year 1,2,3... country X imports from country K in year 1,2,3.... and so on.
2. Can I still use PPML in that context?

Thank you!!!

Jaime

[Joao Santos Silva]

Dear Jaime,

1 - You can do that, but whether that is sensible or not depends on what is the population of interest. In particular.
2 - Yes!

Best wishes,

Joao

[Jaime Giacomozzi]

Joao, thanks for the prompt response!
Sorry I'm not sure what do you mean in point 1... I'm trying to study what determines Chile's dairy products importing trade patterns from 20 countries for the last 25 years.... does that sound reasonable to you?

Many thanks!!

Jaime

[Joao Santos Silva]

Sorry, it appears that I deleted part of a sentence! If you just want to model trade with those 20 countries, it is fine to just use data for them. You can then do one model for each country but you will only have 25 observations for each model right? Why don't you do a model for all the countries? Also, why just focus on those 20 partners?

Best wishes,

Joao

[Jaime Giacomozzi]

Hi Joao,

Regarding your first question, I was thinking to do just one model for the 20 countries using panel data (so the destination country is just one (Chile) and the origins are 20 countries, over the period 1993-2017), so it will be just one model. Hope doesn't sound too weird.

Regarding your second question, if I understand you correctly do you mean Chile vs All the countries? OR all the countries vs all the countries?... I think it is because my main interest is just what affects Chile's import patterns, so I decided to model just Chile's imports, but I'm not sure if that is a valid approach from a methodological point of view.

Regarding your third question, I'm focusing on those 20 countries because they are the main dairy trade partners during that period of time.

Thank you for your time!

Jaime