Announcement

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

  • Fixed effects

    Hi,

    I'm trying to model the following image
    Click image for larger version

Name:	model.PNG
Views:	1
Size:	79.4 KB
ID:	1497837



    Using the hausman test gives me "model fitted on these data fails to meet the asymptotic assumptions of the Hausman test;see suest for a generalized test. I then used the the mundlak test and that gives a value of 0, suggesting fixed effects. However, using fixed effects will eliminate the point of the model including education and experience dummies since these are constant over time. I am modelling constant Education-Experience groups and the impact of the immigrant shock variable on wages. please could you help resolve what the best possible strategy is. Thank you.
    Last edited by Arjun Panchalingam; 11 May 2019, 02:14.

  • #2
    Arjun:
    probably Authors did not use something like -xtrreg,fe-, but an OLS.
    Actually, -xtreg- does not allow interactions with ui and other predictors.
    As an aside, the excerpt of the manuscript that you reported does not seem that clear, as those reported as fixed effectcs are, in my opinion, categorical precictors that were interacted each other.
    Eventually, it would be interesting to see what you typed and what Stata gave you back (as recommended by the FAQ). Thanks.
    Kind regards,
    Carlo
    (Stata 19.0)

    Comment


    • #3
      Carlo provided excellent advice. We do need more information to clarify the message as well.

      Just hazarding a guess, I’m wondering whether this issue can be (to some extent) tackled with the option - sigmamore - for the Hausman test.
      Best regards,

      Marcos

      Comment


      • #4
        Carlo and Marcos, thank you both for the help.
        The stata output when using fixed effects was:
        Click image for larger version

Name:	stata fixed effects.PNG
Views:	1
Size:	35.4 KB
ID:	1497860

        Then when clustering by group:
        Click image for larger version

Name:	stata fixed effects with cluster.PNG
Views:	1
Size:	30.1 KB
ID:	1497861

        Yes, that part i have interpreted as categorical variables that interact with each other. However, on the same paper, they have also said:"is fixed effects correct?"
        Click image for larger version

Name:	model2.PNG
Views:	1
Size:	94.6 KB
ID:	1497862

        Click image for larger version

Name:	model3.PNG
Views:	1
Size:	55.9 KB
ID:	1497863


        Reading papers which use this methodology i was getting the impression the fixed effects meant the categorical variables and the model is OLS but they never explicitly said. Due to this second paragraph suggesting it is a fixed effects model (from one of the papers, which used this methodology for a different country) , i started doubting what the methodology is. The interactions of the categorical variables (education and experience) are eliminated though in fixed effects and therefore i don't see the point of them in the model if it is fixed effects.

        To understand the study better, i have categorised observations for each year into (Education,Experience) groups , calculated the percentage of immigrants in each group and aiming to find the impact of this on the wage.

        I have also done the regression using OLS:
        Click image for larger version

Name:	OLS.PNG
Views:	1
Size:	35.1 KB
ID:	1497864



        Thank you again for all the help. It is very much appreciated. If you any other info, please let me know.

        Comment


        • #5
          Arjun:
          some comments on your laudably provided Stata codes and outcome,.
          1) -fe- machinery wipes out time-invariant predictors (and clustering the standare errors has no effect on it);
          2) what you should have categorized first in your OLS is the -panelid- (-c-id- in your case). As far as predictors are concerne, please note that catagorizing a continuous variable (such as age and/or experience) is, in general, unrecommended (https://www.ncbi.nlm.nih.gov/pubmed/16217841). In addition, some of those predictors may have a non-linear relationship with the regressand, that is worth exploring (at least to check whether some endogeneity issue due to higher level polynomial predictors could ever come alive): see -help fvvarlist- for that.
          Eventually, please do not post screenshots, but (as per FAQ) use CODE delimiters to share what you typed and what Stata gave you back. Thanks.
          Last edited by Carlo Lazzaro; 11 May 2019, 06:56.
          Kind regards,
          Carlo
          (Stata 19.0)

          Comment


          • #6
            Hi Carlo,

            Thank you for the swift reply. Sorry, could you clarify the panelid part. What should i type in? I tried "panelc_id" but that was unrecognised. Thank you, i'll look into the non-linear relationship the predictors may have. Judging by what the paper says, do you think they have used an OLS regression?

            Thank you!

            Comment


            • #7
              Also, if it is an OLS, does that explain the " model fitted on these data fails to meet the asymptotic assumptions of the Hausman test;see suest for a generalized test" in the Hausman test? Is it still appropriate to the sigmamore - for the Hausman test.as Marcos suggested?

              Thank you.

              Comment


              • #8
                Arjun:
                if Authors actually performed an OLS (by the way: is there any clue about that in the Methods section of the article you quoted?), -hausman. is not appropriate.
                Kind regards,
                Carlo
                (Stata 19.0)

                Comment


                • #9
                  Just a few comments, now that some output was shared. It seems the sample size is rather small for a panel data design with so many aspects.

                  Additionally, the (huge) R-squared for the OLS model, plus the (tiny) residuals, let alone the (widespread) collinearity under xtreg, well, I feel there are clues which point out to potential overffitting.
                  Best regards,

                  Marcos

                  Comment


                  • #10
                    Thank you Carlo and Marcos!

                    Carlo:
                    Unfortunately.,there aren't any obvious clue i can see but reading the different papers i did get the feeling it was OLS except the "is the model specification correct?" part in one of the papers(image above). Carlo, could just clarify what you meant i had to do for this part "what you should have categorized first in your OLS is the -panelid- (-c-id- in your case)." Thanks

                    Marcos:
                    Thank you for the feedback. Given the sample size, i will aim to provide stages to the regression in the results, adding each interaction term one by one. I am trying to replicate the study as well as i can but I do understand it is a limitation of the analysis.

                    Comment


                    • #11
                      Arjun:
                      I meant that you should add -i.c_id- in your set of predictors.
                      Model speciifcation refers to investigating possible non-linearity between predictor(s) and regessand and/or the absence of necessary interactions or predictors.
                      Kind regards,
                      Carlo
                      (Stata 19.0)

                      Comment

                      Working...
                      X