[Gabriel Arrieta]

Good evening Mr. Santos Silva,

I have another question about ppml: What kind of robustness test or check can I use in a PPML estimation? Is there any command in Stata to use for this kind of test or isn't necessary in this estimator?

Thanks very much for your comment Mr. Santos Silva.

Greetings,

Gabriel

[Joao Santos Silva]

Dear Gabriel,

I usually do the RESET test. There is no Stata command for that, but we describe how to do it in our page.

All the best,

Joao

[Gabriel Arrieta]

Good morning Mr. Santos Silva,

Thank you very much.

[Gabriel Arrieta]

Good afternoon Mr. Santos Silva,

I would like to ask you some question about your paper "The Log of Gravity". In the page 651, footnote number 27 says: "The formula to compute this effect is (e^bi -1) x 100%, where bi is the estimated coefficient"

So, Is this the process that I have to do with all my estimated coefficients (including those that are in logs and my trade creation and diversion coefficients) to obtain the effects, or is it only for contiguity variable?

And also I would like to know on what paper or methodology is based on or taken that formula to construe the effect? Because, in all the papers that I read this formula, never says where was taken that.

Thank you very much, Mr. Santos Silva.

[Joao Santos Silva]

Gabriel,

That is only for discrete variables such a dummies. You can workout the formula by yourself if you compute the percentage change in the mean of y resulting from a unit change in x.

All the best,

Joao

[Gabriel Arrieta]

Good evening Mr. Santos Silva,

Thanks for your helpful answer.

Greetings

[Gianluigi Giustiziero]

Hi Joao,

In a Stata Journal Paper, you and Silvana argue that the Stata poisson command at times fails to provide accurate estimates when there is perfect collinearity for the subsample with positive observations of the dependent variable. Does the same problem exist with the xtpoisson command?

Thanks for your time and consideration!

Gianluigi

[Joao Santos Silva]

Dear Gianluigi,

I am afraid that is the case. If you have problems with that, you can try to run -ppml- including the fixed effects as dummies (which is valid in the Poisson case, but not in other non-linear models).

Best wishes,

Joao

[Anja]

Dear Joao,

I would need some clarification when estimating ppml (or xtpqml) with country-pair fixed effects and country-specific effects as well as with country-year effects.
I estimate a gravity model of bilateral exports with 8300 observations, where u_ij unequals u_ji.

1. I yield identical results when running the following regressions. Hence, I feel that there is something wrong and that I cannot simply add country-pair and country-specific effects in ppml?

xi: ppml depvar indepvars i.ID i.country_i i.country_j i.Year, cluster(ID)
xi: xtpqml depvar indepvars i.Year, fe

A few observations are dropped because of all zero outcomes but this is not my concern. I rather do not fully understand why I get identical results.

2. Is is appropriate to add country-varying effects and, at the same time, country-pair fixed effects to the ppml command? I have no convergence problems but I am simply afraid that I miss something and should not this.

Thank you for your help.

Best,
Anja

[Joao Santos Silva]

Dear Anja,

1 - Unless I am missing something, the 2 regressions you mention should not generally produce the same results. If you want me to have a look into this, please post your data (or email it directly to me).

2 - Computationally, it is possible to do what you say, but I would not say it is appropriate. First of all, as far as I am aware standard trade models do not include country-pair effects. Additionally, there may be an incidental parameter problem when you include all those fixed effects.

Best wishes,

Joao

[Tom Zylkin]

Hi Anja,

Just to add to what Joao has already said, I think I can shed some light on why these two specifications produce the same result. The reason is because, if ID is a pair specific indicator, then it will absorb your country-specific dummy variables and cause them to drop from the regression. When you only have

xi: ppml depvar indepvars i.ID i.Year, cluster(ID)

this should always be the same as xi: xtpqml depvar indepvars i.Year, fe.


As for country-pair fixed effects, I would give a more nuanced answer than Joao here. If you are specifically interested in identifying, for example, the average treatment effect of a regional trade agreement on trade and if some pairs are more likely to select into RTAs than others (a common problem in that literature) then not including country-pair fixed effects would not be appropriate. Joao is correct though that, while ppml with country-pair fixed effects has become popular for this type of problem, there are still potential issues with incidental parameters that need to be examined further.

If you want to use ppml with country-by-time fixed effects as well as pair fixed effects, I have written a command specifically for this type of specification called ppml_panel_sg, which is avialable via SSC. I would recommend checking the documentation there for more information. Alternatively, if you are concerned about incidental parameters bias, you can instead estimate an OLS model.

[Anja]

Dear both,

thanks a lot for your answers.

On the more general concern you raised, the incidental parameter problem, I cautiously argue that I do not face an incidental parameter problem when including importer- and exporter-specific effects as each country acts as an importing and an exporting nation. Since I have N(N-1) observations I do not estimate specific effects for each unit of observation.

@Joao: Tom confirmed my initial guess that the dummy variables are absorved. Still, I might get back to you with the output via email.
@Tom: Thanks a lot for the clarification on both questions and the command. It yields identical results and will be very helpful in case of computational problems.

Best,
Anja

[Joao Santos Silva]

Dear Anja,

Just to clarify, there is no incidental parameter problem if you include origin and destination fixed effects, even if these vary with time. Also, there is no problem if you include country-pair fixed effects. However, there may be a problem if you include both; I am not aware of any research on that topic.

Best wishes,

Joao

[Kalpana Tokas]

Dear Joao,

I am estimating a gravity model which consists of a panel with 7 years and 12810 observations in total to estimate the effect of FTA on trade flows. I include all three effects in my specification : exporter-time, importer-time, exporter-importer ( country-pair) :

reg lntrade fta i.it i.jt i.ij, cluster(ijt)

Since the trade flows consists if many zeroes, I tried ppml using these two specifications :

xi : ppml lntrade fta i.it i.jt lndist contiguous com_language same_country common_colony, cluster (ijt)

and

xi ; ppml lntrade fta i.it i.jt i.ij, cluster (ijt)

In both cases I get a warning :variance matrix is nonsymmetric or highly singular. There are no standard errors or z values displayed in the result.

Can you suggest where am I making an error and what can be done to rectify it?


Thanks & Regards
Kalpana

[Joao Santos Silva]

Dear Kalpana,

First of all, note that the dependent variable should be trade, not log of trade.

I suggest you try:

xtset ij
xtpoisson trade fta i.it i.jt , fe cluster (ijt)

Best regards,

Joao