How can I improve the R-squared in the Random Effect Model

[Olalere Oluwaseyi]

Have got a Random Effect Model results with relatively low R-squared, i will like to know if its possible to improve the R-squared in the Model.
Thanks.
Oluwaseyi

. xtreg ROA CR1 CR2 LR OR MR SIZE GDP, re

Random-effects GLS regression                   Number of obs      =       112
Group variable: Bank                            Number of groups   =        16

R-sq:  within  = 0.1586                         Obs per group: min =         7
       between = 0.4481                                        avg =       7.0
       overall = 0.1983                                        max =         7

                                                Wald chi2(7)       =     25.72
corr(u_i, X)   = 0 (assumed)                    Prob > chi2        =    0.0006

------------------------------------------------------------------------------
         ROA |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         CR1 |  -.1700334   .0541952    -3.14   0.002    -.2762541   -.0638127
         CR2 |  -.0114985   .0321967    -0.36   0.721    -.0746029    .0516059
          LR |   -.028556   .0262699    -1.09   0.277    -.0800441     .022932
          OR |  -.0451855   .0301742    -1.50   0.134    -.1043257    .0139548
          MR |  -.2029836   .0949081    -2.14   0.032    -.3890001   -.0169671
        SIZE |   .1798149   .1131103     1.59   0.112    -.0418772    .4015071
         GDP |   .2847649   .2394238     1.19   0.234    -.1844971     .754027
       _cons |   .1831775   .0624028     2.94   0.003     .0608703    .3054847
-------------+----------------------------------------------------------------
     sigma_u |          0
     sigma_e |  .03621169
         rho |          0   (fraction of variance due to u_i)
------------------------------------------------------------------------------

[Clyde Schechter]

If you are doing a hypothesis-driven study and this model is appropriate to your hypothesis test, it would be inappropriate to change anything for the purpose of "improving" the R². The results are what they are.

If you are doing exploratory modeling, and are trying to identify relationships among these variables, then I would recommend doing some graphical exploration to see if adding more variable or interaction terms, or transforming some variables would result in a better description of the data. That said, with only 16 groups and 7 predictor variables you are already stretching the limits of your data. There is an excellent chance that the model you have is already overfit and your R² statistics are artifactually high.

[Olalere Oluwaseyi]

Thanks Clyde.

Actually, my study is hypotheses driven....however, i might still need to detect if there exist a significant effect btw the dep_var and the indep_var. But then, in estimating the R-squared in the model, am i gonna report the between r-squared or the overall?. Thanks in advc.

[Clyde Schechter]

That would depend on what your hypothesis is. If your hypothesis is about an effect between firrms/countries/whatever, then you would report the between R². If our hypothesis is about an effect within firms/countries/whatever, then you would report the within R². If your hypothesis incorporates both within- and between- entity effects, then the overall R² is appropriate.

[Phil Bromiley]

Many researchers do not pay much attention to the r-square. I generally emphasize the difference in predicted value for a given change in x which is what most people using ROA as a dv are interested in. Such predictions are easily done with the margins command.