[Megan Ward]

Dear Megan Ward,

1 - The R2 is irrelevant and it is always high in models with all the fixed effects.
2 - Those dummies are collinear with the fixed effects and will always drop out. One possibility is to include the sum of the 2 dummies instead of the separate dummies, but you need to be careful with the interpretation (this imposes the restriction that the coefficients on the 2 dummies are the same).
3 - Estimation will take a long time; you need to wait.
4 - I am not sure what you mean by this, but some observations that are perfectly predicted will be dropped and that will reduce the sample size, but not dramatically.

Best wishes,

Joao

Hi Joao,
I have managed to find some nice results I can use for my thesis, thank you for your previous help
But I have decided that it would actually be useful to calculate the trade diversion effects, as the dummy variable Dsafta (denoting 1 if either country is part of SAFTA) this variable gets absorbed by the country-time fixed effects. So I have done as you suggested, summing the two trade creation and trade diversion dummies,and yes the coefficients on the 2 dummies are the same. However I am unsure in how to interpret this.

diversion	0.674***
	(0.143)

Any advice?
Kind regards,
Megan

[Joao Santos Silva]

Dear Megan Ward,

I am not sure if I understand what you did but you may have misinterpreted my advice (sorry if it was not clear). My suggestion was for you to add the 2 trade diversion dummies (importer in and exporter in). Adding the creation and diversion dummies does not sound sensible.

Best wishes,

Joao

[Megan Ward]

Ah ok, sorry I misunderstood,
So summing the two dummies variables, produces a variable that contains a 2 for both countries being in the RTA, 1 if only one country is in the RTA, and 0 for neither.
However this still gets excluded from the regression....

. ppml_panel_sg foodimports sum_diverge colony comcur ln_dist ln_gdp1 ln_gdp2 comlang_off safta na
> fta ec contig , ex(iso3_o) im(iso3_d) y(year) nopair
Initializing...
Checking for possible non-existence issues...
note: sum_diverge omitted because of collinearity over lhs>0 (creates possible existence issue)
note: ln_gdp1 omitted because of collinearity over lhs>0 (creates possible existence issue)
note: ln_gdp2 omitted because of collinearity over lhs>0 (creates possible existence issue)
Iterating...

---------------------------
Variable | active
-------------+-------------
colony | .23320922
comcur | .44982782
ln_dist | -1.0802436
comlang_off | .191912
safta | .75907003
nafta | .49756911
ec | .67865646
contig | .18781558
---------------------------
iterations: 1272
tolerance: 1.000e-12
Computing standard errors

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

Number of obs = 16,607
Log likelihood = -2.768e+08 R-squared = 0.9413
------------------------------------------------------------------------------
foodimports | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
colony | .2332092 .0448348 5.20 0.000 .1453345 .3210839
comcur | .4498278 .0546193 8.24 0.000 .3427759 .5568797
ln_dist | -1.080244 .0209198 -51.64 0.000 -1.121246 -1.039241
comlang_off | .191912 .036084 5.32 0.000 .1211887 .2626353
safta | .75907 .1598937 4.75 0.000 .4456842 1.072456
nafta | .4975691 .0791022 6.29 0.000 .3425316 .6526067
ec | .6786565 .0628692 10.79 0.000 .555435 .8018779
contig | .1878156 .0296074 6.34 0.000 .1297861 .245845
------------------------------------------------------------------------------
Fixed Effects included: iso3_o-year, iso3_d-year
Robust standard errors (default)

[Joao Santos Silva]

Dear Megan Ward,

Maybe I am missing something but this may be caused by the particular data and model that you are considering. You have data on many importers and many exporters, right?

Best wishes,

Joao

[Megan Ward]

Hi Joao,
For now I am working with a smaller dataset, containing 38 countries, covering 15 years.

kind regards,
Megan

[Joao Santos Silva]

That may be the problem...

Best wishes,

Joao

[Megan Ward]

Running the regression for the larger dataset tells me the variable is dropped even before the results are available....

ppml_panel_sg foodimports sum_diverge colony comcur ln_dist ln_gdp1 ln_gdp2 comlang_off Safta nafta ec contig , ex(iso3_o) im(iso3_d) y(year) nopair
Initializing...
Checking for possible non-existence issues...
note: sum_diverge omitted because of collinearity over lhs>0 (creates possible existence issue)
Iterating...

perhaps i have created this variable incorrectly?

summarize sum_diverge

Variable | Obs Mean Std. Dev. Min Max
-------------+---------------------------------------------------------
sum_diverge | 512,163 .0392063 .1978517 0 2

[Joao Santos Silva]

Sorry, Megan Ward, I have run out of ideas... I'll let you know if something comes to mind.

Best wishes,

Joao

[Tom Zylkin]

Hi Megan,
About the "initial values not feasible" error, I don't see that very often so it is hard to say what is causing it. But there is a "guessols" option that may help with initializing the values for the iteration. I suggest trying that. You should feel free to email me directly about that if you are still having problems.
Regards,
Tom

[Tom Zylkin]

Megan Ward , Regarding trade diversion, it is important to keep in mind that it is not possible to estimate what you are calling a "trade diversion effect" when you have exporter-time and importer-time fixed effects. Basically, if I subtract your "safta" variable from your "sum_diverge" variable I get a variable that is always 0 for non-safta members and always 2 for safta members. Therefore there is no residual variation associated with sum_diverge that is pair-time specific (as opposed to being country-time specific). Also note you would get the exact same result if you used OLS with this specification (it is matter of perfect collinearity between regressors rather than a PPML-specific issue.)

This is actually a common point of confusion when it comes to gravity regressions with FTAs as the variable of interest. It is something of a subtle point that it is not possible to separately identify trade diversion and trade creation effects (at least not without internal trade flows in addition to international trade flows). I have an earlier post on this forum in response to a similar question that may be helpful. One thing you can do for example is create an interaction between "1 in SAFTA" and log_distance (or other similar variables) to see if SAFTA caused trade diversion from natural non-SAFTA trade partners.

[Megan Ward]

Hi Tom,

Thank you very much for your clarification, I was thinking it must be possible because i have read about it in the literature, but now i realise the papers i have been reading include internal trade flows and use fixed effects. Or just dont estimate with fixed effects. I would like to include fixed effects, as i do not want to include omitted variable bias in my results.

With your suggestion of an interaction term, should i still include the safta variable?

Would you recommend I attempt to estimate a fixed effects with random intercept model to avoid the variables from being dropped, such as this paper states https://www.researchgate.net/publica...mum_Likelihood . if so, is this possible to do with your ppml_panel_sg command?

The important thing with this analysis is that i can see the impact of trade for the south asia countries, so to conclude that there has been an increase or decrease in trade from these countries due to the trade agreement.
How i understood it is, If i include only the safta dummy variable, this tells me how much global trade has changed due to the agreement.

Any other suggestions on how to achieve this would be greatly appreciated,

And thank you Joao and Tom again for your help.
Regards,
Megan

[Tom Zylkin]

Hi Megan,

Hi Tom,

Thank you very much for your clarification, I was thinking it must be possible because i have read about it in the literature, but now i realise the papers i have been reading include internal trade flows and use fixed effects. Or just dont estimate with fixed effects. I would like to include fixed effects, as i do not want to include omitted variable bias in my results.

With your suggestion of an interaction term, should i still include the safta variable?

Yes.

Would you recommend I attempt to estimate a fixed effects with random intercept model to avoid the variables from being dropped, such as this paper states https://www.researchgate.net/publica...mum_Likelihood . if so, is this possible to do with your ppml_panel_sg command?

No it is not possible with ppml_panel_sg. And no I would not recommend it.

The important thing with this analysis is that i can see the impact of trade for the south asia countries, so to conclude that there has been an increase or decrease in trade from these countries due to the trade agreement.
How i understood it is, If i include only the safta dummy variable, this tells me how much global trade has changed due to the agreement.

How I would explain it is, the safta dummy = 1 only for safta members. Plus, if you have exporter-time and importer-time FEs, these FEs absorb all determinants of trade that are not bilateral in nature (this includes global determinants of trade). Therefore, the safta dummy should only be picking up relative differences in trade between safta members relative to non-safta members.

However, you are also indicating you want to determine if there was an increase or decrease in trade following the agreeement. For this, you will also need to include exporter-importer (or "country-pair") fixed effects. Without these fixed effects, your safta coefficient will also be picking up whether safta countries are more likely to trade with one another rather than solely isolating the change in trade that happens after the agreement.

Any other suggestions on how to achieve this would be greatly appreciated,

And thank you Joao and Tom again for your help.
Regards,
Megan

One other suggestion is to compute a simple general equilbrium model of how other countries are affected by the safta agreement. This is something of an advanced technique but you can refer to the recent "Advanced Guide to Trade Policy Analysis" put out by the WTO and UNCTAD for some very accessible discussion and for a method that can be implemented in Stata.

Regards,
Tom

[Megan Ward]

Hi Tom,
Thanks for your response!

the safta dummy = 1 only for safta members

Just to clarify, so the safta dummy takes the variable 1 if either pair is in the agreement then? for example it would be 1 for India and USA. and 1 for India and Nepal.
Or it takes 1 if both countries are in the agreement, for example it would be 1 for India and Nepal, 0 for India and USA.

and again just to clarify the interpretation of this coefficient,when using both country-pair and country-year fixed effects, i could conclude, for example if the coefficient on the safta variable is 0.2 , ((e ^ 0.2) - 1) * 100 = 22.14%. that trade is 22% higher amongst safta countries relative to global trade, compared to if there had been no trade agreement.

And thanks for the suggestion, I will look into the general equilibrium model also!

[Tom Zylkin]

Hi Megan,

Hi Tom,
Thanks for your response!


Just to clarify, so the safta dummy takes the variable 1 if either pair is in the agreement then? for example it would be 1 for India and USA. and 1 for India and Nepal.
Or it takes 1 if both countries are in the agreement, for example it would be 1 for India and Nepal, 0 for India and USA.

It should be equal to 1 only when both members are in the agreement if you want this variable to reflect the effects of SAFTA on bilateral trade. Otherwise it would be picking up the effects of safta on a country's trade in general (ie. with all other countries). As such it would be absorbed by the country-time fixed effects.

and again just to clarify the interpretation of this coefficient,when using both country-pair and country-year fixed effects, i could conclude, for example if the coefficient on the safta variable is 0.2 , ((e ^ 0.2) - 1) * 100 = 22.14%. that trade is 22% higher amongst safta countries relative to global trade, compared to if there had been no trade agreement.

Yes I think that's mostly correct. How I would word it is that trade is 22% higher among safta members relative to their trade with non-safta countries compared to if there had been no trade agreement.

Regards,
Tom

[Megan Ward]

Thank you Tom for all your help!

One last question, and I think I am ready to analyze my results...
say there was a 22% increase in trade among safta members relative to their trade with non-safta countries, compared to if there had been no trade agreement, this percentage change could either be caused by 1) an increase in trade amongst safta members, or 2) a decrease in trade amongst non-trade countries? perhaps a combination of both ?

Kind regards,
Megan