Announcement

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

  • FGLS regression with weights

    Hello everyone,

    I want to replicate a study with a given data-set. The regression should be a feasible generalized least-square estimation with White heteroscedasticity robust standard errors (corrected for heteroscedasticity in a variable called "iid"). Now, if I run the STATA code:

    reg depvar var1 var2 ... [weight=iid], robust

    I dont get the same results as in the original study. Do i use the wrong command or what I am missing?

    Thanks for your help, if you need additional info, please comment as well.

    Cheers

    Kurt

  • #2
    Your command nominally assumes that the variance is proportional to 1/iid. Could it be the author assumed the variance was proportion to iid, in which case the weight should be 1/iid? You haven't provided enough details.

    Comment


    • #3
      Okay, thank you very much first for your quick answer and sorry for the lack of details.

      What I want to estimate is a gravity equation in the form of ln(y) = ln(var1) +.... + some dummies. y is the trade between two countries in 9 different years. So iid is the country pair and is itself a nominal variable (it starts with 1 and adds 9 for each new country pair i.e. 1 10 19 28 and so on untill 11017). The regression is pooled over all time periods and country pairs, but I want to control for heteroscedasticity in the country pairs, iid. I assume, that weights then dont make sense, because the variance in iid is given by the definition and so should be arbitrary. So I should use a different command?

      Again thank you and I hope it is quite understandable what I wrote.

      Best

      Comment


      • #4
        I wondered, given its name, whether iid was an identifier. As you said, it makes no sense to use it in weights for weighted least squares. You have to be careful with phrases such as "control for heteroscedasticity." You can either adjust your inference or you can try to model the specific form of heteroskedasticity. If you want to allow it when computing the OLS standard errors then you just use the "robust" option -- no weights. The only way to model it is to allow it to be a function of the x variables for each country pair. I doubt this will be very natural. It sounds like you're not really replicating a study with the original data set, but applying a method from a paper to a new data set.

        It's easy to use OLS and allow the standard errors to be robust to arbitrary heteroskedasticity using the robust option. More of an issue is allowing for correlation across country pairs. Colin Cameron and Doug Miller have recent unpublished research on that.I think you can find it at Colin's web page at U.C. Davis.

        Comment


        • #5
          Okay. Thank you again!

          I just asked the authors for the data-set and now work with the one they sent me. If I just use reg and the robust option, I get coefficients which are very very close to the original ones. Just in the study they write "pooled OLS estimates with White, heteroscedasticity-robust standard errors of the baseline regression (Standard errors are corrected for heteroscedasticity among country pairs)". And about the baseline regression " We estimate our baseline regression, which includes times dummies, by pooling the data and using feasible generalized least-squares...".

          From these quotes I assumed the robust option with the reg command but was confused by the information in parenthesis and how to deal with it. Since my estimates only differ in one case more than 0.001 from the original estimates, I could be that the data-set changed after the study was published...Number of obs. and R² is also the same in my estimations and theirs.

          Thank you also for the hint to the article, I will have a look at it soon and it will help me further for sure.

          best

          Comment

          Working...
          X