Announcement

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

  • doubts on the Hausman test

    Hi to everyone! I have some doubts about the Hausman test. I used it before to choose between fixed and random effects. Now I have the same situation, panel data and fixed effects to estimate, but I noticed some different answer after the test.

    1 First of all sometimes Stata tells me the matrix is not positive-define, how I can solve this problem? I read someting about the options sigmamore but the use is not pretty clear.
    2 Sometimes test failed since the answer is "model fitted on these data fails to meet the asymptotic assumptions of the Hausman test; see suest for a generalized test". I have to use this different test? How it works? I read also about xtoverid test, it could be an alternative?
    3 I use robust clustered stantard error but the Hausman test cannot be performed with them. Could be a problem in the results?
    4 When I use the model in the whole sample I have no problem with the test but I'm not sure about its interpratation. I have a model that estimate the relation between globalization and growth, with different kinds of controls. If I run the simplest model I have fixed effect, If I add controls about financial development the model gives me fixed effect, if I add also government controls I have again random effect. It's just an example since I have different regressors but the question is about the alternance of the fixed and random effects, it makes sense?

    Thanks you all

  • #2
    Enza:
    1. re-run -hausman- with -sigmamore- option;
    2. and 3.: try -xtoverid-;
    4. you're talking about different regression models; no wonder that -hausman- points you out to different specifications.

    As an aside, please post what you typed and what Stata gave you back (via CODE delimiters, please) to increase your chances of getting helpful replies (as reminded by FAQ). Thanks.
    Kind regards,
    Carlo
    (Stata 19.0)

    Comment


    • #3
      Thank you Carlo, in the case that I explained before, it's better use hausman with sigmamore or xtoverid? Considering the fact that with hausman I cannot consider robust std error, it could be a problem for my model? I just want to understand very clearly your advice, I always used hasuman without problems and if it can be a good solution I prefer to use a test that I know and I've used before, but obviously if it give me wrong answer i cannot use it.

      Comment


      • #4
        Enza:
        if you have robustified/clustered standard errors (SEs) in panel data regression models, you should go -xtoverid-, as -hausman- supports default SEs only.
        Last edited by Carlo Lazzaro; 01 Mar 2017, 04:25.
        Kind regards,
        Carlo
        (Stata 19.0)

        Comment


        • #5
          Ok, thank you Carlo

          Comment


          • #6
            You might want to consider a "robust" regression-based Hausman test. See the following discussion in the Stata Blog:
            Fixed effects or random effects: The Mundlak approach
            https://www.kripfganz.de/stata/

            Comment


            • #7
              Perhaps a reflection on "Not positive definite matrices: causes and cures" would help.
              Best regards,

              Marcos

              Comment


              • #8
                Thank you Sebastian, can I use the Mundlak's approach with more than two regressors? And the controls? I have to consider them in the regression?

                Comment


                • #9
                  Enza:
                  yes, you can use Mundlak's approach with more than two regressors.
                  I would also keep the controls in.
                  Kind regards,
                  Carlo
                  (Stata 19.0)

                  Comment


                  • #10
                    Hi! If I run the xtoverid test sometimes it contradicts the result of the Hausman and I think that it can happen since I consider clustered SE. But sometimes the results of xtoverid is an error that tells me "saved RE estimates are degenerate (sigma_u=0) and equivalent to pooled OLS", so what I can do in this case? I have to work with the pooled model or I have to run an Hausman test and follow its results? Thanks!!!

                    Comment


                    • #11
                      Enza:
                      - as you noted -hausman- does not support any standard error that differs from default;
                      - as far as your second question is concerned, -xtoverid- outcome tells you that POLs outperforms -xtreg, re-.
                      I assume that -xtreg, fe- is out of debate in your case.
                      Kind regards,
                      Carlo
                      (Stata 19.0)

                      Comment


                      • #12
                        Ok, so the POLS performs the best results in this case and there is no variance in the panel-specific intercepts, they are all the same intercept. How can I justify this situation?
                        Last edited by Enza Testa; 14 Mar 2017, 04:11.

                        Comment


                        • #13
                          Enza:
                          it's difficult to say, as -xtoverid- outcome is data-sependent.
                          However, I would not say that POLS performs the best results; rather, there's no evidence that -xtreg, re- is more informative than POLS (it happens and should be of no concern).
                          Kind regards,
                          Carlo
                          (Stata 19.0)

                          Comment


                          • #14
                            Ok, thank you!!

                            Comment

                            Working...
                            X