Announcement

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

  • Panel Logit Regression with Fixed Effects, Interpretation of a Dummy Independent Variable(xtlogit,fe vs. aextlogit vs. xtlogit,re)

    Dear all,

    first I will thank you for any respond and help. I am not a experienced econometrician and no English native, so I apologize if my question might be not very precise, but I did my best

    I want to estimate a panel logit model, accounting for fixed effects using a unbalanced panel (T=4 N=56,398). The aim is to investigate the effect of heart diseases (heart=1 if yes, 0 otherwise) on labor force participation (inlf=1 if working, 0 otherwise).

    - My first idea was to use the xtlogit command with the option ,fe resulting to my code:

    Code:
     xtlogit inlf heart $controlls, fe vce(oim)
     margins, dydx(heart)
    ==> As pointed out in http://repec.org/usug2016/santos_uksug16.pdf the estimated marginal effects are conditional on the assumption that the fixed effects are 0, which leads to meaningless results if the assumption does not hold. To assume the fixed effects to be 0 is totally random? and thus the marginal effects are meaningless.

    -Then I read about the user written aextlogit command, which solves this problem and transforms the estimated coefficients to "average (semi-) elasticities of Pr(y=1|x,u) with respect to the regressors, and the corresponding standard errors and t-statistics". So I was going with the command


    Code:
     aextlogit inlf heart $controlls
    ==> The problem here is that my main variable of interest (heart) is a dummy variable and thus there are only discrete changes from 0 to 1 and a interpretation of a elasticity is not meaningful. Also using ln(heart) to get the semi elasticity is not possible, because of the binary nature of heart.

    - So both models are not suitable for my investigation, am I right?

    -My idea is now to run a random effects logit model with clustered standard errors of the form of

    Code:
     xtlogit inlf heart $controlls, re vce(cluster personid)
    .

    which unfortunately does not account for time constant heterogeneity, but in my opinion is the only possibility to use a non-linear model.

    -The other alternative would be to use a linear probability model


    Code:
     xtreg inlf heart $controlls, fe cluster (personid)
    which is not suitable to predict probabilities (because of constant partial effects, unbounded predicted probabilities and heteroscedasticity) as I want to do.

    So my questions are:

    - Is it appropriate to first estimate a linear probability model to get a first approximation of the effect and then go to the panel logit regression with random effects?
    -Is there any other possibility to control for the fixed effects?
    -And can anyone share some citable literature I can use for this argumentation?

    I really appriciate any help and comments and thank you in advance.

    Best,
    ​​​​​​​Claudio Schiener








  • #2
    Has nobody any idea how to solve my problem?

    I would appreciate any help or hints.

    Comment


    • #3
      Dear Claudio,

      I would not use RE. From the FE logit you can get the sign and significance of your variable of interest. If this is not enough for you, the linear probability model you suggested may give you an idea of the effect size.

      Best wishes,

      Joao

      Comment


      • #4
        Steve Samuels makes various suggestions at

        https://www.statalist.org/forums/for...-after-xtlogit

        Someday I am going to try them out!

        I'm not a big fan of odds ratios, but they may be better than nothing if you want to make the results a little more intuitive.
        -------------------------------------------
        Richard Williams, Notre Dame Dept of Sociology
        Stata Version: 17.0 MP (2 processor)

        EMAIL: [email protected]
        WWW: https://www3.nd.edu/~rwilliam

        Comment

        Working...
        X