Announcement

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

  • Fix heteroskedasticity and autocorrelation with RE model in panel data

    Hello everyone,
    I am a newbie in using Stata, so hope you guys will explain everything clearly and in plain language.
    I have read a lot of previous posts before but I'm so confused because almost posts with a similar subject were about fe model.
    My database has N = 24 and T=12 => 288 observations. I'll present the steps of running my model as below:
    1. Running OLS with command of -reg depvar indepvar-.
    2. Running Correlation matrix and VIF to check multicollinearity.
    3. Running fe model (-xtreg depvar indepvar, fe) and re model (-xtreg depvar indepvar, re)
    4. Conducting the Breusch and Pagan Lagrangian multiplier test (-xttest0-) and Hausman test (-hausman-) to select the appropriate model, which is REM method.
    5. Performing 3 tests for RE model as Pesaran's test of cross- sectional independence (-xtcsd, pesaran abs-), Breusch and Pagan Lagrangian multiplier test for heteroskedasticity (-xttest0-)
    Wooldridge test for autocorrelation (-xtserial depvar indepvar-). The results show my RE model has heteroskedasticity and autocorrelation.

    I have some questions:
    1. Is my modeling process good enough?
    2. Could I use command of -xtreg depvar indepvar, re robust/vce(cluster)/cluster(N) (N stand for the numbers of firms) in order to fix heteroskedasticity and autocorrelation like FE model? And then I will use these RE robust results for my finale conclusions?
    3. Or I should run -xtreg depvar indepvar, re robust- and -xtreg depvar indepvar, fe robust-, then choose appropriate model by robust hausman with command of -xtoverid-? I also don't know how to use -xtoverid- command.

    Thank you so so much for your help. I've been stuck with all these thoughts for months.
    Best regards,
    Linh Bui


  • #2
    Then major problem you have is that neither your cross sectional, nor your time series dimension are large. 24 cross sectional observations is not particularly big sample.

    All the tests that you perform have asymptotic justification, so they are all suspect with 24 observations.

    Yes, you can and you probably should calculate robust/clustered by N standard errors post random effects models.

    You cannot test overidentifying restrictions with -xtoverid- because without additional instruments your model is exactly identified.

    Comment


    • #3
      I appreciate your answer, Mr. Kolev
      I don't understand why my steps of running model are "asymptotic justification". Did you mean that because my database has too little observations? And I should add N and T to increase my observations? What can I do to increase the liability of the model?
      Best regards,
      Linh Bui

      Comment

      Working...
      X