Announcement

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

  • Interpreting fixed effect model

    Hello everyone!

    I am looking at the differences in Taxable income between self-employers and wage workers.
    I conducted an OLS and now I did a fixed effect (after doing the Hausman test).
    My problem is that I have never learned what exactly fixed effect does and how I can interpret it and therefore what I can tell my thesis supervisor about it.
    My question is: Can someone explain me what fixed effect is and give me some examples about how to interpret it?
    The fixed effect results are attached.

    Thanks in forward!

    P. Suiker
    Attached Files

  • #2
    P(aul?):
    there's plenty of literature on this topic.
    You may want to start to take a look at -xt- and -xtreg- entries in Stata 13.1 .pdf manual.
    You may also want to read: Allison PD. Fixed effects regression models. New York, NY: SAGE Publications, 2009.
    Kind regards,
    Carlo
    (Stata 19.0)

    Comment


    • #3
      P. Suiker,

      The interpretations of fixed-effects models vary with the analysis and can get really complicated really fast. Multiple classroom lectures can be devoted to just the interpretations themselves. If I can put it as simply as possible, the coefficient estimate for your variable of interest (Employment?), after running a FE model, will show the average effect of the type of employment on taxable income after controlling for year fixed-effects, civil status fixed-effects, education fixed-effects, and origin. This means that the effect of employment type given your data is as refined as possible instead of being influenced by all of the aforementioned factors.

      I'm assuming you already compared your FE estimates to your normal OLS results as well as those from a Random-Effects model. I would also consider adding in a time trend variable to control for any natural progression in your taxable income variable.

      -Thomas

      Comment


      • #4
        Carlo and Thomas,

        Thank you very much for your answer. Carlo I will definitely take a look at those readings.

        Thomas, in response to your comment, I have three questions if that is not a problem.

        1. You told me that the average effect of the type of employment on taxable income is controlled for year fe, civil status fe, education fe and origin fe. According to that, I have a small question: how is the average effect controlled for the following variables? Or it just isn't?
        Age
        Agesq
        Work_experience
        Work_experienceSQ
        Early_retirement
        Household_head
        Age
        Agesq
        Work_experience
        Work_experienceSQ
        Early_retirement
        Household_head
        2. I did not yet compare the results for the FE and the OLS. When I am looking at it now, the results are very divergent which gives me the feeling that the results are wrong or just very weird.
        It can also be because I did the OLS for wage workers and self-employers seperately instead of putting the employment variable in. (For your information, I made an extra dummy variable 0=wage worker, 1=self-employer from the two dummy variables wage worker 0=no 1=yes and self-employer 0=no 1=yes).
        My question is, are they comparable or do you see something wrong? (find my OLS results attached)

        3. I don't have a Random-Effects model because the Hausman test shown me that I have to use fixed effect. My question: Is it valuable to compute a Random-Effects model and compare it with the fixed one?

        Thank you so much for your time!

        Best regards,

        Patrick Suiker
        Attached Files

        Comment


        • #5
          Patrick,

          1. All of those covariates you've mentioned, plus the fixed effects from earlier, are being controlled for by being included in the regression.

          2. The results between OLS and FE models could indeed be very different. Especially if the fixed effects are statistically significant, meaning that their omission from the OLS model could have been biasing your coefficient estimates. As such, just because your results are different doesn't mean that they are wrong. I've been taught to run an F-test on the joint significance of your fixed effect variables to see whether an OLS or FE model is more appropriate. Depending on the commands you're using, the F-statistic may actually be included in the regression output without you needing to run separate tests.

          3. In my experience, I've run both an FE and an RE model first, then I've used a Hausman test to compare their estimates. If there is a way to do it without computing the RE model then by all means go for it. If not, it may be valuable to look into. Unfortunately, I'm not as knowledgeable regarding Hausman tests.

          I hope I've been able to help,

          Thomas
          Last edited by Thomas Beatty; 18 Jun 2015, 14:32.

          Comment


          • #6
            Thanks for your answer, Very helpful!

            Another question that I would like to ask:
            I did my OLS and quantile regression seperately for wage workers and self-employers.
            I did the fixed effect for the wage workers and self-employers together (those are combined in the variable Employment where 0=wage worker and 1=self-employer)

            Is it possible to compare this fixed effect for the Employment variable with the regressions that are ran seperately? Or did I make here a mistake taking the Employment variable?
            Again, all relevant results are attached.

            Thanks in forward for your answer!
            Attached Files

            Comment


            • #7
              Based on your description, I would imagine that you would want to run the OLS regression with a binary variable (just like you did with the FE model) before comparing the estimates. It's hard to say though without an intimate understanding of the data that are being used.

              For future reference, I believe that it's considered proper etiquette to post Stata output in a CODE box. Many forum users don't feel safe opening up attachments. If you take these steps, you're likely to get a greater response from the community.

              -Thomas

              Comment


              • #8
                Patrick:
                -as an aside to Thomas' remark, you can increase your chance of getting helpful replies if your attachments are in Stata format only. This approach will allow others on the list to repeat your analysis in the easiest way and commenting on the resulst accordingly;
                - you were probably right in avoiding any comparison between -xtreg, fe- and OLS, as, according to what you reported, they seem quite different models;
                - in general it is not useful to perform a random effect model when -hausman- results support the opposite (i.e., fixed effects). However, the main issue is whether -u_i- is correlated or not with the other regressors in the real world (something that the hausman test can't tell you; the literature in your research field can probably support you in this respect).
                Kind regards,
                Carlo
                (Stata 19.0)

                Comment

                Working...
                X