Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Appropriate commands in Gravity Model with Fixed Effects

    Hello to everyone,
    I have started to use Stata this month and I am completing my undergraduate thesis about the "resilience" of gravity models when it comes to make predictions starting from ex-post evaluations.
    The panel dataset I am working on consists of 20 countries (each country appears twice, first as an exporter and then as an importer) in a time-span of 9 years (2008-2016). The main variables I am working on are bilateral exports as dependent and as independent the following: bilateral tariffs, bilateral real exchange rates in the currency of the exporter (computed from the nominal and the Consumer Price Indexes rearranged), the GDP of the exporter and the GDP of the importer, the population of both, bilateral distance, and a series of dummies (colonial relationship, common language, contiguity, landlocked exporter and landlocked importer). Moreover, as I am studying a North-South integration which may happen in the future, I have created 3 dummies for any North-South FTA already in place to control for trade creation inside, trade diversion and trade creation outside.

    My dataset is almost ultimated and I am going to use the following FEs: country-pair fixed effects and time fixed effects. Given that, I have some doubts about what comes next:

    1) As I am not inserting country-time fixed effects, should I construct an index of "remoteness" to control for the famous multilateral resistance terms? In this case, should I use the world GDP to compute such index, or the sum of the GDP of my countries (as the world was made up of only such countries?

    2) I would like to run a OSL regression and a PPML one as proposed by most of the literature. I would like to insert the already mentioned FEs. Should I insert them manually, and then insert all the dummies in the regression command and add the option FE too? In particular what I was doing was:
    *For the OSL*
    Code:
    quietly tab pairid, gen(pair_)
    quietly tab year, gen(year_)
    xtreg logY logX1 logX2 year_* pair_*, fe
    *For the PPML*
    Code:
    xtpoisson Y logX1 logX2 year_*, fe i(pairid) vce(robust)
    For the second one, I am not sure about the command efficiency, in particular for the use of FE.
    Are such commands correct? Is there any option I am missing for providing for better results?

    3)At the end, I would like to run the regression with the PPML but in this case, I know that such command doesn't support the FEs. In this case I could investigate the contribution of even such variables which were collinear with the FE in the previous two regressions. But how specifically does this command do?
    Code:
    ppml Y logX1, cluster(pairid)
    In the end, my questions can be summarized in: which commands should I use to run a OSL regression and, after that a PPML regression with FEs and which options should be included in the command? How does PPML command work with respect to FEs and how should I interpret the different results coming from the PPML regression with xtpoisson, fe and ppml?

    Thank you for the attention.

  • #2
    Dear Andrea,

    The multilateral resistance terms are time-varying origin and destination dummies. In PPML you can easily include them using the command ppml_panel_sg. If you really want to use OLS, you will need to include the dummies yourself, but I do not really see the point of using OLS in this context.

    Best wishes,

    Joao

    Comment


    • #3
      Dear Joao thank you for the nice reply,
      Actually I have run my regression with the ppml_panel_sg command, but obviously most of my variables dropped, except for the tariff one. As a consequence, I was not able to investigate trade diversion/creation in North-South FTAs. Indeed, I was thinking about running a regression with xtpoisson with only country-pair and time fixed effects and, after that, another one with ppml_panel_sg to investigate the role of tariffs when all the other variables are controlled for by the fixed effects.

      When it comes to xtpoisson should I insert the dummies manually?

      Code:
       xtpoisson Y logX1 logX2 year_* pair_*, fe vce(robust)
      Is this command correct?

      Regarding the OLS, my point was demonstrating (as in one of your paper, which is my strongest reference) that the PPML regression is really more robust.

      Kind regards,
      Andrea

      Comment


      • #4
        Dear Andrea,

        With xtpoisson you do not need to include the pair FE.

        Best wishes,

        Joao

        Comment


        • #5
          Dear Joao, thank you for answering. I have one last question about the data.
          With panel data and PPML estimation, is it better to include GDP in nominal terms, GDP PPP or GDP per capita and why?
          Accordingly, for the exchange rate. I have used the Consumer Price Index (base year 2010) and the Nominal Exchange rate, to derive the Bilateral Real Exchange Rate. Could I use such index in my regression? Thank you

          Comment


          • #6
            Dear Andrea,

            I am afraid I do not know enough to help with that.

            Best wishes,

            Joao

            Comment

            Working...
            X