Announcement

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

  • Fixed effects model

    Hello everyone,

    For my thesis, I am looking at the effect of environmental controversies on the profitability of Chinese and European firms. I am using panel data (2013-2018).
    As I want to compare both groups (China vs. Europe), I split the dataset and ran 2 regressions.

    After computing the Hausman test for the European dataset, it showed that random effects model is the superior model. Not problems here.
    However, for the Chinese dataset it would be beter to use a fixed effects model (according to Hausman).

    With panel variable id (company) and time (year) (strongly balanced), i conducted the following fixed effects model:
    xtreg ROA EnvCont1_lag EnvCont2_lag CEP1_lag CEP2_lag GUOState OwnCon RD FirmSize Leverage i.Industry i.Year, fe robust
    (OwnCon, RD, FirmSize, Leverage and Industry are my control variables)

    I included i.Industry and i.Year as I would like to add industry and time fixed effects as well. I think this is the correct way?
    However, when I run the regression, the variables OwnCon (time-invariant) and i.Industry are omitted (because of collinearity), and there is barely any significance.
    Click image for larger version

Name:	Screenshot 2020-04-03 at 19.55.03.png
Views:	1
Size:	86.6 KB
ID:	1544561





    Also, GUOState (dummy for state ownership) is omitted, but this is an important variable for one of my hypotheses.

    My questions:
    - Am I correct in splitting the dataset for China and Europe? (I did this as I was unsure if a dummy variable would be enough to compare them?)
    - Is this the right way to conduct a fixed effects model (with firm, industry and time fixed effects)?
    - Is it correct to use firm fixed effects, industry fixed effects AND time fixed effects? Or how do I know what has to be 'fixed'? (I have not been able to find answers to do this online)

    This is also my first time using panel data, so any suggestions on how to move forward would be great.
    Thank you in advance!!
    Last edited by Hanne Zwertvaegher; 03 Apr 2020, 12:09.

  • #2
    Depending on how you do it, you should get almost identical results with the split sample and a dummy for region interacting with all the rhs variables. It is easier with the dummy and clustered robust standard errors.

    You can't include both firm and industry because firms don't change industries - the firm effect fully controls for industry. Many automatically include time effects. See what is done in your area.

    Comment


    • #3
      Thank you, Phil! Your explanation as to why I shouldn't include industry effects here makes perfect sense. One more question, how do I interpret the omitted OwnCon (ownership concentration - control variable) and GUOState (dummy for state ownership - important independent variable)? For example, when I want to answer the question whether these variables influence the profitability? Can I simply not interpret them?

      Comment


      • #4
        Hanne:
        as they are collinear with the fixed effect or time-invariant, there's simply no way that you can estimate their coefficients via -fe-.
        Kind regards,
        Carlo
        (Stata 18.0 SE)

        Comment


        • #5
          Hello Carlo! Thank you for your response. In my case, do you think I could use OLS with industry and year dummies instead of FEM with year and firm fixed effects in order to be able to interpret the coefficients of the time-invariant variables (eventhough Hausman signals otherwise)? Or is this simply wrong?

          Last edited by Hanne Zwertvaegher; 07 Apr 2020, 09:32.

          Comment


          • #6
            Hanne:
            if you're interested in estimating the coefficient of time-invariant predictors, you should switch to the community-contributed programme -xthybrid- and read Jeff Wooldridge ' s replies on this topic before starting your analysis.
            Kind regards,
            Carlo
            (Stata 18.0 SE)

            Comment

            Working...
            X