Announcement

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

  • Time fixed effects in random effects model

    Hi everyone

    I am conducting a research whereby my dependent variable is the quality of the corporate annual report which take the values (1-60) and the focal independent variable is tenure years (1-9 years).
    the relationship comes as significant before and after controlling for other determinants and industry dummy but not when I include the year dummy.

    My command is as follows:

    Code:
    xtreg Quality tenure CONTROLS SIC, re vce(cluster SIC)
    For the sake of simplicity I have omitted the control variables which ranges from binary to continuous variables.

    Is it justified to ignore the year dummy and report the results on the grounds that my main effect is time variable (tenure)?

    Many Thanks

  • #2
    Paul:
    how did you -xtset- your data?
    Kind regards,
    Carlo
    (Stata 18.0 SE)

    Comment


    • #3
      Hello Carlo,

      I have declared my data using the menu method as shown below, where isinnum is the companies registration number and YEAR is the sample years (2005-2011)




      Click image for larger version

Name:	declare panel.png
Views:	1
Size:	13.4 KB
ID:	933495

      Comment


      • #4
        Paul:
        thanks a lot for providing more details.
        However, you woud be better off in terms of potentially helpful replies if you post not only what you typed, but also whar Stata gave you back (as per FAQ).
        For example, when you include i.year dummies, are (some of) them statistical significant, wheras other predictors stop to be so?
        Kind regards,
        Carlo
        (Stata 18.0 SE)

        Comment


        • #5
          Hi Carlo

          Thanks for your reply.

          I don't usual include the year dummy as
          Code:
          i.year
          but only includes the variable name as it is (year).
          When I do this (i.year) I only get the year 2007 to be significant but not with the other years.

          Comment


          • #6
            Paul:
            considering those futher details, I would omit year and i.year from the righ-hand side of your regression.
            Kind regards,
            Carlo
            (Stata 18.0 SE)

            Comment


            • #7
              Thanks Carlo for your comments.

              Any justification or reference for the omission of the year dummy in this case?

              Is it justified to omitted it because I have two time variables (tenure and year) so the model is better off excluding the year dummy?

              Comment


              • #8
                Paul:
                I have no reference about this issue, I have just expressed my opinion. However, if you intend to submit for publication your research, the usual advice is to take a look at what previous researchers did on the same topic (in terms of both regression approach and predictors). I might well be that, when i.dummy is included you get only 2007 being statistical significant due to a limited variation across the year for the same unit. It's difficult to say (for me, at least) whether this finding is sample-related or mirrors something substantive about the population your sample was drawn from.
                Eventually, provided this is not my research field, I would find less interesting including -year- as a predictor, as I fail to get the information you can convey via 1-unit change in this independent variable: does, in the empirical world, an 1-unit increase in time expressed in year explain change in the quality of corporate annual report when adjusted for other predictors?
                Anyway, please take a look once more at what others colleagues of yours published in the past.
                Kind regards,
                Carlo
                (Stata 18.0 SE)

                Comment


                • #9
                  Carlo,

                  Thanks for your reply.

                  The thing is that my research area isn't well established and only one study relates to my topic that is US based with a lot of differences in terms of predictors choices (not necessarily available in UK databases). They also use fixed effect model which in my case is rejected by the Hausman test.
                  Therefore, there is no benchmark to rely on.
                  Additionally, most of the research in the broader Accounting topics include the year dummy.

                  Many Thanks

                  Comment


                  • #10
                    Paul:
                    a temptative approach might be to include in your paper both the regression models (with and without i.year) and explain the different results obtained.
                    Kind regards,
                    Carlo
                    (Stata 18.0 SE)

                    Comment


                    • #11
                      Carol,

                      Thanks for your comments.

                      That sounds logical, but I finds it confusing to report two contradicting findings. It is a matter of whether there is a significant relationship between the the predictor and response.

                      One more question:
                      You kept using the expression (i.year) in your posts. Is the correct form of including the year and industry dummies?
                      I always include them as (year sic) and this seems to report different p values!

                      Comment


                      • #12
                        Paul:
                        - even a non significant relationship can be informative ("more research is needed" is often the companion statement in that instance);
                        -however, should you make a choice based on Hausman'sn specification test, go -re-;
                        - i.year would consider the year dummy only (from your post I get that this approach has been followed by previous research on the same topic);
                        - I still believe that the best way to give you more helpful replies is to take a look at the result of your Stata session, that you can easily post via code delimiters (# button among the advance editor options).
                        Kind regards,
                        Carlo
                        (Stata 18.0 SE)

                        Comment


                        • #13
                          Carlo,

                          Thank you for your constructive replies.
                          I will come back to this post with more detailed regression results.

                          Regards

                          Comment

                          Working...
                          X