Announcement

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

  • Model Selection for Balanced Micro Panel Data

    Have a good day to all contributers,
    I have a panel data set with T=3 and N=22 with 1 dependent 4 independent variables.
    To choose which model will fits best for my date i have tried the following procedure:
    1- Run the regression for Fixed Effect Model >> xtreg y x1 x2 x3 x4, fe
    2- Run the regression for Random effext model >> xtreg y x1 x2 x3 x4, re
    3- Run Hausman Test to compare FE and RE >> hausman fixed random, sigmamore // Prob>chi2=0.0563 then i go for RE model.
    4- Run LM Test to compare RE and POLS >> xttest0 // Prob > chibar2 = 0.0000 then i go for RE model.
    5- I have checked for heteroscedasticity for RE model wiht the following commands:
    5.1- xtreg Consumption Price Rainy_Day Total_Rain Total_Evaporation, re
    5.2- predict eps,e
    5.3- robvar eps, by(groups) >> W0: Pr > F =0.0007, W50: Pr > F = 0.85, W10: Pr > F = 0.0007 since W0 and W10 < 0.05 Reject the Null and conclude that there is heteroscedasticity
    6- I have checked for auto-correlation >> xtregar y x1 x2 x3 x4, re lbi >> DW=2.43 Baltagi-WU LBI=2.95. Since both test statistics are greater than 2 then there is no auto-correlation.
    7- I have checked for cross-section dependency using Pesaran and Friedman Test
    7.1 xtreg y x1 x2 x3 x4, re
    7.2 xtcsd, pesaran >> Pr= 0.58
    7.3 xtcsd, friedman >> Pr= 1.00 both values are greater than 0.05 then do not reject H0 and i have concluded that there is no cross-section dependency.
    8- At last, i have run the following robust RE model estimation
    xtreg y x1 x2 x3 x4, re vce(robust)

    My questions:
    a-) My final model at the 8th step has Prob>chi2 = 0.0991 and Walt chi2(4) = 7.80.
    With these results i concluded that my model is not appropriate for my data. Am i right?
    b-) If my perception on section a is right, how can i fix this problem or improve my model? Any suggestion would be very appreciated.
    Kind and best regards.
    Last edited by Emin Bildirici; 10 Nov 2023, 05:06.

  • #2
    Any comment or feedback would be very appreciated.

    Comment


    • #3
      Emin: You shouldn’t be computing any of those tests. The default is fixed effects with time dummies, and with N = 24 you shouldn’t be relying on asymptotic tests (which is every test you report). Random effects is a feasible GLS estimator justified as N gets large. At best, you can cluster your standard errors by unit, and N = 24 is pushing it. If nothing is significant using two-way fixed effects then you should see if you collect more data.

      Note per FAQ that we ask you to show your commands and Stata output. Also, I notice you opted for RE because the nonrobust Hausman test had a p-value just above .05. You don’t want to take the arbitrary cutoffs seriously.

      Comment


      • #4
        Originally posted by Jeff Wooldridge View Post
        Emin: You shouldn’t be computing any of those tests. The default is fixed effects with time dummies, and with N = 24 you shouldn’t be relying on asymptotic tests (which is every test you report). Random effects is a feasible GLS estimator justified as N gets large. At best, you can cluster your standard errors by unit, and N = 24 is pushing it. If nothing is significant using two-way fixed effects then you should see if you collect more data.

        Note per FAQ that we ask you to show your commands and Stata output. Also, I notice you opted for RE because the nonrobust Hausman test had a p-value just above .05. You don’t want to take the arbitrary cutoffs seriously.
        Hi Mr. Wooldridge, thank you for your enlighting answer. I almost lost my hope that someone is gonna answer my question .
        I have changed my model after lots of digging on the web. Since i have very limited observation in time axis, now i gave up to investigate time effect now only interested with the cross-section data.
        I have run the OLS with cross-section data and now i have a statistically significant model and coefficients.
        My current pathway and Stata commands as below.
        Thanks again for your time and answer.
        Best and kind regards.
        //Regression
        regress y x1 x2 x3 x4
        y Coef. St.Err. t-value p-value [95% Conf Interval] Sig
        x1 .451 .225 2.01 .049 .002 .9 **
        x2 -1.937 .411 -4.72 0 -2.758 -1.116 ***
        x3 .055 .275 0.20 .842 -.495 .605
        x4 -.031 .313 -0.10 .92 -.657 .594
        Constant 11.567 1.523 7.60 0 8.523 14.612 ***
        Mean dependent var 8.725 SD dependent var 0.394
        R-squared 0.372 Number of obs 66
        F-test 9.033 Prob > F 0.000
        Akaike crit. (AIC) 42.553 Bayesian crit. (BIC) 53.501
        *** p<.01, ** p<.05, * p<.1
        //Multi Colinearity Check
        lmcol y x1 x2 x3 x4 >>Suggests that there is no multi-collinearity

        //Heteroscedasticity Test
        regress y x1 x2 x3 x4
        estat hettest >>Suggests that there is heteroscedasticity

        //Final Model
        regress Consumption Price Rainy_Day Total_Rain Total_Evaporation, vce(robust)
        Linear regression
        y Coef. St.Err. t-value p-value [95% Conf Interval] Sig
        x1 .451 .204 2.21 .031 .043 .858 **
        x2 -1.937 .477 -4.06 0 -2.891 -.983 ***
        x3 .055 .299 0.18 .854 -.543 .653
        x4 -.031 .316 -0.10 .921 -.663 .6
        Constant 11.567 1.606 7.20 0 8.356 14.778 ***
        Mean dependent var 8.725 SD dependent var 0.394
        R-squared 0.372 Number of obs 66
        F-test 6.101 Prob > F 0.000
        Akaike crit. (AIC) 42.553 Bayesian crit. (BIC) 53.501
        *** p<.01, ** p<.05, * p<.1

        Comment

        Working...
        X