Announcement

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

  • #16
    Well, unless I'm missing something, your concept of bias is purely metaphysical and cannot be pinned down in the real world. In the equation
    BtP(t)= intercept+ firm fixed effects + time fixed effects+ a* lags of change in stock prices + error(t)
    the intercept, firm fixed effects, and time fixed effects are not identified. You can add anything you want to the intercept, and then subtract two somethings that add up to that anything from the firm and time fixed effects, respectively, and you have the same model: it makes the same outcome predictions in all cases. So this equation alone cannot serve as a real world model of anything. It requires some kind of constraints to identify the terms. Various constraints can be considered, but I have no sense of what would be appropriate to your situation. Your calculation of bias depends on the specific way in which you constrain the model to identify these components, so, basically, you can make it come out to be anything you want it to be by choosing the corresponding identifying constraints.

    There is no such thing as the "true" fixed effects or the "true" constant from a statistical point of view. From an accounting point of view there may be, and if so, that would govern how you constrain the model. Perhaps you can search the accounting literature to learn what the convention is in this regard.

    Comment


    • #17
      I have a similar problem--I truly want to estimate a panel data model with no intercept based on a theoretical model without intercept. I.e I would like a panel estimator for y[i,t]=beta*x[i,t]+epsilon[i,t]. In all of the panel model presentations there is an intercept that 'washes out' in a fixed effect or first difference approach, i.e. the model is
      y[i,t]=alpha[i] +beta*x[i,t]+epsilon[i,t]. I guess since my data have some measurement error there is a non-zero alpha for one or more of my panels and I would like that to be set to 0 either through a 'noconstant' option or through a constrained xtreg type process. So while I understand the linear dependence of the (overall) constant on the set of panel-specific constants, I'm not sure how that addresses the question of estimation without a constant? Thanks for any clarifications!

      Comment


      • #18
        If I understand what you want correctly, the reasoning would be this:

        y[i,t]=alpha[i] +beta*x[i,t]+epsilon[i,t], with the constraint that alpha[i] = 0 for all i is the same as:

        y[i,t]=beta*x[i,t]+epsilon[i,t]

        Now, notice that in this equation, not only is there no alpha term, there are no terms at all that depend solely on i--i.e. no panel level effects at all. So this can be done with simple regression:

        Code:
        regress y x, nocons

        Comment


        • #19
          note that if there are panel level effects, you can also use a "no constant" option for either -xtreg- or -mixed-; see the help files

          Comment


          • #20
            Hi all,

            I'm using "Piece wise linear Regression" to test the linearity of return's t-1 impact over funds's inflow in t.

            My data base is in Panel and i'm testing POLS, FE and RE models.

            My problem is that when I run

            " xtreg Captacaot Return_first_quartil Return_second_quartil Return_third_quartil Return_fourth_quartil first_quartil second_quartil third_quartil fourth_quartil Desvio_Padraot1 Taxa_Admt1 LN_PLt Alpha_Jensent1 CobraTaxadePerformance "

            the "Fourth_quartil" is dropped because of collinearity. Given that "first_quartil second_quartil third_quartil fourth_quartil " are the intercept of each equation, is there any other way to get the fourth_quartil impact?

            Thanks

            Comment

            Working...
            X