Announcement

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

  • Panel Data Tests (stationarity, pooled-OLS fe, re, hausman)

    Dear members,

    I am currently working on my bachelor thesis in economics. Unfortunately I have chosen a topic I have not been prepared for during my studies. Fortunately I am learning a lot as a result.

    I am investigating the impact of internet use on economic growth in China. I have collected 8 years worth of data for 31 provinces. All my variables are in log form. My dependent variable is log(Real GDP per Capita), my independent variable is log(internet user percentage) and I control using log(gross capital formation percentage of gdp), log(urbanization percentage of gdp), log(high education attainment percentage of total population).

    The next step is regressing these. I have learned that there are various ways which in particular are useful to me. I would like to run three models: pooled-OLS, fixed effects, and random effects. In a paper before me I saw they first check for stationarity in the data. I did thise too using the Levin-Lin-Chu test, which showed all the variables except education is stationary. However I read that this test is not appropiate to use since there are more provinces then years (N>T). I then found the Harris-Tzavalis test, however now it showed that all my variables contain a unit root (p values ranging from 0.70 to 1.00).

    Ignoring stationarity I did the fixed effects and random effects model and the hausman test. Fortunately all my variables are significant. Unfortunately I got the following error for the hausman test (V_b-V_B is not positive definite):

    I have posted the results of these below:








    I am not sure what I should do now. Most other papers researching this effect does not show the appendix, so I have no idea what steps they took.

    Furthermore I have already tried researching of course on the internet, but it is all quite hard for me to understand. I just wanted to run some basic regressions, the goal of my thesis, but then realized it all much harder than that. For example someone has suggested running random effects first, then fixed effects for the hausman test. However later on I read that this is not a good way to solve the not positive definite problem.

    Furthermore theory suggests a fixed effects model is appropriate to use, however the hausman tests rejects this.

    I was hoping to gain some new insights.

    Thank you very much.

    Kind regards,

    Jens
    Last edited by Jens Jong; 07 Jun 2016, 05:56.

  • #2
    Jens:
    welcome to the list.
    This happens quite frequently.
    For more on this topic, take a look at : http://www.stata.com/statalist/archi...05/msg00723.ht...
    Whoever advised you to invert the -fe- -re- sequence before -hausman- is wrong.
    Kind regards,
    Carlo
    (StataNow 18.5)

    Comment


    • #3
      Besides Carlo's recommendations, Jens could also wish to take a look at the Mundlak test: http://blog.stata.com/2015/10/29/fix...dlak-approach/

      Best,

      Marcos
      Best regards,

      Marcos

      Comment


      • #4
        Also, for the Hausman test, I strongly recommend you add the -sigmamore - option so as to "back up" positive definite differences between variances and covariances.
        Best regards,

        Marcos

        Comment


        • #5
          Dear Carlo and Marcos,

          Thank you very much for your suggestions. When I get back to the university tomorrow, for STATA access, I will try your suggestions. I really appreciate the information.

          I have corrected your link Carlo, since it was invalid. Please see the following page for those who are interested:
          http://www.stata.com/statalist/archi.../msg00723.html

          Regarding my first problem, stationarity, do you have similar suggestions for that as well? I think I have overestimated myself when I went into this topic, to be honest I did not not that STATA could be this complex.

          Kind regards,

          Jens

          Comment


          • #6
            Jens:
            thanks for amending the link I meant to point you out to.
            I'm not expert in (mainly) time-series stuff, such as stationarity,
            Complexity is embedded in econometrics (and quantitative methods, at large): Stata has nothing to be blamed for in this respect.
            A valuable textbook on (micro)econometrics for Stata users is: http://www.stata.com/bookstore/microeconometrics-stata/.
            Kind regards,
            Carlo
            (StataNow 18.5)

            Comment


            • #7
              I'm also not an expert itn time series. But if I remember correctly, taking the first difference of your variable(s) may solve your problem with stationary.

              Comment

              Working...
              X