Announcement

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

  • Dynamic Panel Data

    I am working on Dynamic Panel Data. I am not certain on how to interpret the Sargan overidentification test (below).

    estat sargan
    Sargan test of overidentifying restrictions
    H0: overidentifying restrictions are valid

    chi2(451) = 24181.51
    Prob > chi2 = 0.0000

    Thanks Kidaya

  • #2
    In general, if H0 is one thing, and you get a significant test, then you've rejected that thing. So your data are not consistent with the over identifying restrictions being valid.

    Comment


    • #3
      As Phil said, you reject the null behind the Sargan test of overidentifying restrictions, which means that your model is misspecified. I would be trying to remove one or more of the instrumental variable you are currently using and/or try to replace them with other, more valid, instruments.

      Comment


      • #4
        Many thanks to both. I might get back to you guys.

        Comment


        • #5
          As Anat and Phil said, your model is mispecified. Said differently, your set of instruments is invalid, i.e. E(z,u) is different from zero (z is the set of instruments and u is the error term). Moreover, if you use a system-gmm is appropriate to have a look at the Hansen-in-difference test, that tests the validity of the additional moment conditions, implied by that estimator with respect to the difference-gmm.

          Comment


          • #6
            I had tried IV estimation see below. My data consist of 35 groups by id. Is this one way of reducing the numbers pf instruments?

            xtivreg D.lnundf D.lnpol D.lngdp D.lnprimedu (DL. lnundf=L2D.lnundf), fe

            Fixed-effects (within) IV regression Number of obs = 1,111
            Group variable: id Number of groups = 35

            R-sq: Obs per group:
            within = 0.8638 min = 7
            between = 0.8175 avg = 31.7
            overall = 0.8850 max = 43

            Wald chi2(4) = 12800.35
            corr(u_i, Xb) = 0.0737 Prob > chi2 = 0.0000

            ------------------------------------------------------------------------------
            D.lnundf | Coef. Std. Err. z P>|z| [95% Conf. Interval]
            -------------+----------------------------------------------------------------
            lnundf |
            LD. | .8731837 .0127488 68.49 0.000 .8481965 .8981709
            |
            lnpol |
            D1. | .0005476 .0010524 0.52 0.603 -.001515 .0026102
            |
            lngdp |
            D1. | -.0125202 .0058485 -2.14 0.032 -.023983 -.0010574
            |
            lnprimedu |
            D1. | -.013772 .0048872 -2.82 0.005 -.0233507 -.0041932
            |
            _cons | -.0021097 .0004726 -4.46 0.000 -.0030361 -.0011834
            -------------+----------------------------------------------------------------
            sigma_u | .00461233
            sigma_e | .00952727
            rho | .18987077 (fraction of variance due to u_i)
            ------------------------------------------------------------------------------
            F test that all u_i=0: F(34,1072) = 0.70 Prob > F = 0.9013
            ------------------------------------------------------------------------------
            Instrumented: LD.lnundf
            Instruments: D.lnpol D.lngdp D.lnprimedu L2D.lnundf
            ------------------------------------------------------------------------------

            Comment


            • #7
              You are combining first differences with the fe within transformation. That is probably not what you want to do. You might want to use the fd option of xtivreg instead (without manually differencing your variables).
              https://www.kripfganz.de/stata/

              Comment


              • #8
                Dear Sebastian thanks for replying to my quest. I am trying to avoid an over identification issue. I had used the xtabond (Arellano-Bond) Arellano-Bond dynamic panel-data estimation but I end up with multiple instruments. I have a panel Data of 35 countries (by id) year from 1960-2014. I don’t know what to do in order to reduce the number of instruments? See Below

                xtabond lnundf lnpol lngdp lnprimedu

                Arellano-Bond dynamic panel-data estimation Number of obs = 1,112
                Group variable: id Number of groups = 35
                Time variable: year
                Obs per group:
                min = 7
                avg = 31.77143
                max = 43

                Number of instruments = 876 Wald chi2(4) = 786995.27
                Prob > chi2 = 0.0000
                One-step results
                ------------------------------------------------------------------------------
                lnundf | Coef. Std. Err. z P>|z| [95% Conf. Interval]
                -------------+----------------------------------------------------------------
                lnundf |
                L1. | .967467 .0021936 441.03 0.000 .9631676 .9717665
                |
                lnpol | -.0024815 .0005635 -4.40 0.000 -.0035859 -.001377
                lngdp | -.0214006 .0022025 -9.72 0.000 -.0257173 -.0170838
                lnprimedu | -.0113142 .0011716 -9.66 0.000 -.0136105 -.0090179
                _cons | .4311402 .0320666 13.45 0.000 .3682908 .4939895
                ------------------------------------------------------------------------------
                Instruments for differenced equation
                GMM-type: L(2/.).lnundf
                Standard: D.lnpol D.lngdp D.lnprimedu
                Instruments for level equation
                Standard: _cons





                Comment


                • #9
                  Your time dimension is relatively large compared to the cross-sectional dimension. The Arellano-Bond and related estimators are only suitable for small-T frameworks because the number of instruments increases quadratically with T.

                  If you still want to use GMM estimators, have a look at the user-written xtabond2 command and make sure to restrict the lag depth and to "collapse" the instruments. See the help file for details and the article How to do xtabond2 ... for further information.

                  Alternatively, you might be interested in mean-group (MG) / pooled mean-group (PMG) estimators.
                  https://www.kripfganz.de/stata/

                  Comment


                  • #10
                    Thanks Sebastian, following your suggestion I tried xtdpd instaed of (xtabond2), I am using stata14. See below,


                    xtdpd lnlif l.lnlif lnpol lngdp, dgmmiv(lnlif) div(lnpol lngdp)

                    Dynamic panel-data estimation Number of obs = 1,646
                    Group variable: id Number of groups = 35
                    Time variable: year
                    Obs per group:
                    min = 22
                    avg = 47.02857
                    max = 54

                    Number of instruments = 455 Wald chi2(3) = 1.29e+06
                    Prob > chi2 = 0.0000
                    One-step results
                    ------------------------------------------------------------------------------
                    lnlif | Coef. Std. Err. z P>|z| [95% Conf. Interval]
                    -------------+----------------------------------------------------------------
                    lnlif |
                    L1. | .9614014 .0009847 976.33 0.000 .9594714 .9633314
                    |
                    lnpol | .0032994 .0001426 23.15 0.000 .00302 .0035788
                    lngdp | .0087551 .0005725 15.29 0.000 .0076331 .0098772
                    _cons | .0957177 .0039062 24.50 0.000 .0880616 .1033738
                    ------------------------------------------------------------------------------
                    Instruments for differenced equation
                    GMM-type: L(2/.).lnlif
                    Standard: D.lnpol D.lngdp
                    Instruments for level equation
                    Standard: _cons

                    Comment

                    Working...
                    X