Announcement

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

  • Fixed effects by group and with Likert scales

    Hi everyone, I'm going a bit crazy with the specification of a regression I need to do and I'd appreciate any advice on how to proceed.

    I have a panel dataset of 500 individuals who make 5 binary decisions. I randomly allocated half of these individuals to make decisions in a different order (to control for ordering effects). For each individual I also have 4 variables that measure their personality on a Likert scale 1-10. I'd like to see if the ordering changes any of their binary decisions, and if their personality also matters.

    My model is as follows:

    *Define panel data by unique id*

    encode id, gen(idno)
    xtset idno

    *FE Logit Regression"

    xtlogit share i.treat_n i.group_no x1 x2 x3 x4, fe nolog


    where:

    "share" is the binary variable (0 or 1)
    "treat_n" is the categorical variable, with values to 1 to 5, indicating the 5 scenarios where each individual either shares or not shares
    "group_no" is the binary variable (0 or 1) that indicates if they belonged to the group that saw the decisions in a different order
    x1 x2 x3 x4 are Likert scale values (1 to 10) based on self-reported preferences (e.g. willingness to take risks, willingness to share, etc.)

    My problems are:

    1) the variable "group_no" gets omitted, also if I run xtreg instead of xtlogit. How do I overcome this?
    2) Is that the best way to include Likert scale variables?
    3) Is there a completely different way than doing xtlogit, fe or re that I might be forgetting?


    Thank you!

  • #2
    Will
    - firrst of all, -fe- specification under -xtlogit- produces conditional fixed effect, which differ from the -fe- under -xtreg-;
    - you do not report (as you should have, following FAQ recommendations), what Stata gave you back. However, -i.group- was probably omitted due to collinearity (and you cannot do anything about that but change your model specification);
    -far from being my field, but I would check whether others used -pca- to summarize Likert scale results;
    -the usual question, that is difficult to answer, especially with non-linear panel data regressions, is why going -fe- instead of -re-.As usual, a theoretical support may come from the literature in your research field and/or an informed discussion with a more experienced colleague/teacher/professor/mentor.
    Kind regards,
    Carlo
    (Stata 18.0 SE)

    Comment


    • #3
      Just a side note, after Carlo's insightful reply. Maybe a gsem approach would be helpful.
      Best regards,

      Marcos

      Comment


      • #4
        Originally posted by Carlo Lazzaro View Post
        Will
        - firrst of all, -fe- specification under -xtlogit- produces conditional fixed effect, which differ from the -fe- under -xtreg-;
        - you do not report (as you should have, following FAQ recommendations), what Stata gave you back. However, -i.group- was probably omitted due to collinearity (and you cannot do anything about that but change your model specification);
        -far from being my field, but I would check whether others used -pca- to summarize Likert scale results;
        -the usual question, that is difficult to answer, especially with non-linear panel data regressions, is why going -fe- instead of -re-.As usual, a theoretical support may come from the literature in your research field and/or an informed discussion with a more experienced colleague/teacher/professor/mentor.
        Thank you very much for your reply Carlo.

        Indeed, in all specifications I used [ xtlogit, fe; xtreg, fe; xtreg, re; etc.] all omit the group variable due to collinearity. Using gsem as Marcos suggests basically crashes my laptop as it runs hundreds of interactions until it gets stuck (as my sample is 500 individuals making 5 choices each, I doubt it should be so cumbersome on a good laptop).

        I'm still left in the dark on this and don't quite know how to proceed...

        Comment


        • #5
          Will:
          is there any advice available from the literature in your research field?
          Kind regards,
          Carlo
          (Stata 18.0 SE)

          Comment


          • #6
            You may start with: a) Fewer or no interactions; b) smaller sample.
            Best regards,

            Marcos

            Comment


            • #7
              You may as well fiddle with the options concerning the integration method.
              Best regards,

              Marcos

              Comment

              Working...
              X