r squared for xtpoisson

[Rob Baker]

I am trying to assess the trade effects of an RTA. I am using PPML to estimate a gravity model. It's a panel dataset so I am trying to use dyadic and time fixed effects as per much of the literature.

I run the following commands with the output provided in the attached png. file:

xtset countrypairnumber year
xtpoisson tradevalueus lnGDPIM lnGDPEX lnpopim lnpopex ceta1 ceta2 ceta3, fe


My questions are:
1) Are these commands correct for my methodology?
2) I am required to produce an r squared value in my published results, how do I ascertain an r squared value for the xtpoisson command?
3) Literature employing apparently similar methods include a constant in their published results, can I / should I ascertain a constant value for the xtpoisson command?


Thanks,

Rob

[Joao Santos Silva]

Dear Rob Baker,

1) You should use robust standard errors. I suggest you use ppmlhdfe instead of xtpoisson
2&3) forget about the R2 and the constant

Best wishes,

Joao

[Joro Kolev]

Regarding the R-squared, you can always generate a pseudo-R-squared, by calculating:

1 - [Likelihood of the full model] / [likelihood of the empty model].

see my post #4 on this thread here: https://www.statalist.org/forums/for...nential-models

[Rob Baker]

Dear Joao Santos Silva & Joro Kolev,
Thank you both for taking the time to get back to me. Your assistance was massively appreciated. If I may, I have a couple of further questions...

For my first model (time and dyadic fixed effects) I ran ppmlhdfe as suggested via the commands below:

egen imp=group(importer)
egen exp=group(exporter)
ppmlhdfe tradevalueus lnGDPIM lnGDPEX lnpopim lnpopex ceta1 ceta2 ceta3, a(year imp#exp) nolog

This gave me the output shown in the first attached png. file.

For my second model, I would like to use country-pair, exporter-and-year and importer-and-year fixed effects (as in Yang & Martinez Zarzoso 2014 & others). In this instance, lnGDPIM lnGDPEX lnpopim lnpopex are effectively covered so I drop them as independent variables. I run the following commands:

ppmlhdfe tradevalueus ceta1 ceta2 ceta3, a(imp#year exp#year imp#exp) nolog

This gave me the output shown in the second attached png. file with CETA3 omitted because of collinearity. CETA3 is a dummy variable for import trade diversion so I understand why this may be omitted with the inclusion of importer-and-year fixed effects. Curiously though, this doesn't occur with CETA2, dummy variable for export trade diversion, even after I change the order of the syntax.

As a prospective solution, I generated CETA4, a dummy variable to cover both import and export trade diversion. I then ran the following commands:

ppmlhdfe tradevalueus ceta1 ceta4, a(imp#year exp#year imp#exp) nolog

This gave me the output in the third attached png. file, which appears fine.

My questions are:

1) Are the commands for my first model correct?
2) Is there any way of running my second model without CETA3 being omitted? (Yang & Martinez Zarzoso have used identical dummy variables within the same model successfully, I'm just not sure how!)
3) If this is impossible, are the commands for my final model (with CETA4) correct?

Best,

Rob

[Joao Santos Silva]

Sorry, Rob Baker, I cannot really comment on this without knowing much more about your data.

Best wishes,

Joao

[Rob Baker]

Joao Santos Silva of course - I understand. Thank you again for your help though!

Best,

Rob