Which model is best for my panel data?

[Nicholas Akam]

Dear Statalist members,

I am using an unbalanced panel of 144 countries from 1970-2015 (3300 obs) to measure the effect of globalisation on inequality in developing countries. I am using the KOF measure of globalisation, avg years of education data and log gdp per capita as independent variables and Gini coefficients (from Frederick Solt's SWIID) as the dependent.

I am struggling to understand whether to use random or fixed effects in my model, and when one should use time and entity FE or simply entity.

The FE and RE outputs from STATA are:

. xtreg gini_disp kofgi educ  ln_gdppc ln_gdppc2, fe rob

Fixed-effects (within) regression               Number of obs     =      3,299
Group variable: countrycode                     Number of groups  =        144

R-sq:                                           Obs per group:
     within  = 0.0405                                         min =          1
     between = 0.0887                                         avg =       22.9
     overall = 0.0870                                         max =         46

                                                F(4,143)          =       1.56
corr(u_i, Xb)  = 0.1466                         Prob > F          =     0.1873

                          (Std. Err. adjusted for 144 clusters in countrycode)
------------------------------------------------------------------------------
             |               Robust
   gini_disp |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       kofgi |   .0711179   .0365415     1.95   0.054    -.0011134    .1433492
        educ |  -.5249671   .3325533    -1.58   0.117    -1.182323    .1323885
    ln_gdppc |   2.322239   1.694989     1.37   0.173    -1.028233    5.672712
   ln_gdppc2 |  -.1558884   .1132115    -1.38   0.171    -.3796727    .0678958
       _cons |    32.9037    6.38667     5.15   0.000     20.27922    45.52818
-------------+----------------------------------------------------------------
     sigma_u |  7.3203003
     sigma_e |  1.6181866
         rho |  .95341139   (fraction of variance due to u_i)
------------------------------------------------------------------------------

. xtreg gini_disp kofgi educ  ln_gdppc ln_gdppc2, re rob

Random-effects GLS regression                   Number of obs     =      3,299
Group variable: countrycode                     Number of groups  =        144

R-sq:                                           Obs per group:
     within  = 0.0404                                         min =          1
     between = 0.0900                                         avg =       22.9
     overall = 0.0882                                         max =         46

                                                Wald chi2(4)      =       6.74
corr(u_i, X)   = 0 (assumed)                    Prob > chi2       =     0.1505

                          (Std. Err. adjusted for 144 clusters in countrycode)
------------------------------------------------------------------------------
             |               Robust
   gini_disp |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       kofgi |   .0727446   .0352125     2.07   0.039     .0037294    .1417597
        educ |  -.5576193   .3121589    -1.79   0.074    -1.169439    .0542009
    ln_gdppc |    2.33882   1.694603     1.38   0.168    -.9825407     5.66018
   ln_gdppc2 |  -.1549905   .1130468    -1.37   0.170    -.3765582    .0665771
       _cons |   32.15156   6.363721     5.05   0.000      19.6789    44.62423
-------------+----------------------------------------------------------------
     sigma_u |  7.0202524
     sigma_e |  1.6181866
         rho |  .94954907   (fraction of variance due to u_i)
------------------------------------------------------------------------------

Another concern is that my controls do not have any significance, despite theoretical models stating their importance. Is this something I should be worried about?

From the literature on the subject I have seen some papers with similar models that use a GMM model but I am not familiar with this estimator, if you could explain the merits of the models and when they should applied I would be enormously grateful!

Thanks for reading,

Nick Akam
Undergraduate Liberal Arts & Sciences, University of Birmingham, UK

P.s Apologies if this is not a well phrased/structured topic, it is my first time using this forum.

[Carlo Lazzaro]

Nicholas:
set aside the GMM issue (that woud deserve a full teaching note; you'd better off with taking a look at -gmm- entry in Stata .pdf manual and related references), as you imposed non-default standard errors, -hausman- cannot help you out, whereas the user-written command -xtoverid- can do the trick:

xtreg gini_disp kofgi educ  ln_gdppc ln_gdppc2, re rob
xtoverid

If -xtoverid- rejects the null, you should go -fe-.

[Nicholas Akam]

Carlo,

Thanks very much for your helpful advice! This helped sort my problem.