[Jihad El Yaagoubi]

And should I test for Normality, Heterogeneity and Multicollinearity before presenting resultsnof the pooled OLS model ? Is it obligatory to report it in my thesis ?
many thanks
Jihad

[Carlo Lazzaro]

Jihad:
- you should check omitted variable bias (or, better non-linear relationships between your dependent variable and (usually) one of your predictors) (see -estat ovtest-);
- quasi-extreme multicollinearity (see -estat vif-), although it is debated whether it is an issue or not (see http://www.hup.harvard.edu/catalog.p...ontent=reviews, Chapter 23).

[Jihad El Yaagoubi]

I tried estat ovtest and estat vif and get these results, but didn't how to interpret
reg ROA i.CSR Risk Size ib2.Industry,vce(cluster Companyscode)

Linear regression                               Number of obs     =        139
                                                F(6, 27)          =       2.15
                                                Prob > F          =     0.0803
                                                R-squared         =     0.6649
                                                Root MSE          =     .43165

                          (Std. Err. adjusted for 28 clusters in Companyscode)
------------------------------------------------------------------------------
             |               Robust
         ROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       1.CSR |   .1158301   .0642641     1.80   0.083    -.0160289    .2476891
        Risk |   .6491344   .3023649     2.15   0.041     .0287329    1.269536
        Size |  -.1009088   .0477627    -2.11   0.044    -.1989097   -.0029079
             |
    Industry |
          1  |   -.192214   .0868874    -2.21   0.036    -.3704923   -.0139358
          3  |  -.1182151   .1022348    -1.16   0.258    -.3279836    .0915534
          4  |  -.1268181    .083513    -1.52   0.141    -.2981726    .0445364
             |
       _cons |   1.028692    .442875     2.32   0.028     .1199876    1.937396
------------------------------------------------------------------------------

. estat ovtest

Ramsey RESET test using powers of the fitted values of ROA
       Ho:  model has no omitted variables
                 F(3, 129) =    873.56
                  Prob > F =      0.0000

. estat vif

    Variable |       VIF       1/VIF  
-------------+----------------------
       1.CSR |      1.04    0.965412
        Risk |      1.04    0.964214
        Size |      2.09    0.478117
    Industry |
          1  |      2.05    0.487958
          3  |      3.72    0.268542
          4  |      1.99    0.503174
-------------+----------------------
    Mean VIF |      1.99

.

[Carlo Lazzaro]

Jihad:
as VIF is not worrying, -estat ovtest- is.
Your model lacks of a squared or cubic relationship between a given predictor and the dependent variable or od an interaction between two (in general) predictors.
At worst, your model suffers from endogeneity (residuals might be correlated with a given predictor and with the dependent variable at the same time; the latter does no harm; the first violate one of the OLS requirements: residuals should not be correlated with the vector of regressors).
Eventually, I would log the dependent variable and see whether the outcome of -estat ovtest- still remains statistically significant.
As I do not know your research field and related literature, I cannot help you out any further with the substantive interpretation of your results: maybe discussing with a classmate/teacher/supervisor can shed light on the whole matter.

[Jihad El Yaagoubi]

So the solution is to calculate log of the dependant variable and re run regression and ovtest
OR add a control variable that is related to both depvar and indepvar ? because I found that if p value of ovtest is signifcant that means that the model omits important variables
Thank you Carlo.. Your replies are VERY HELPFUL. Many thanks
jihad

[Carlo Lazzaro]

Jihad:
your interpretation of -estat ovtest- outcome is correct.
However, the statistical significance of -estat ovtest- can have different meanings:
- the model omits important predictors, which, in turn, can mean that your model is poorly specified (see the literature in your research field to have a comprehensive idea about the predictors to be included in your regression models) or that a given predictor has a non-linear relationship with the dependent variable;
- eventually, omitted variable bias may imply enodogeneity (which a relevant bias, that makes your regression resulst unreliable), if some of the omitted predictor, hidden in the residual, is correlated with both a predictor and the dependent variable.
In some cases, logging the dependent variable can make the -estat ovtest- unsignificant.
Finally, I fail to get what you mean by

...a control variable that is related to both depvar and indepvar...

as a fix for your problem.

[Jihad El Yaagoubi]

add a control variable that is related to both depvar and indepvar

I want to say that maybe I should add another variable to the model which is correlated to the dep var and indepvar. (but variables I included are sufficient in regard to literature
- You're right in the literature of my research field, there are some authors who said that the relationship isn't linear between these variables but nonlinear ==> Should I run a nonlinear regression? so I will give up on the pooled ols model?

[Carlo Lazzaro]

Jihad:
as your colleagues report, non-linearity among a given independent variable and the the dependent variable is perfectly legal under OLS.
Conversely, you cannot have squared, cubic (or higher degree) coefficients.
Thats said, you should go on with pooled OLS and add squared terms or interactions, consistently with previous research experience on the same topic.

[Jihad El Yaagoubi]

add squared terms

that means instead of

 reg ROA i.CSR Risk Size ib2.Industry,vce(cluster Companyscode)

, I type reg ROA² i.CSR Risk Size ib2.Industry,vce(cluster Companyscode) ?

[Jihad El Yaagoubi]

Conversely, you cannot have squared, cubic (or higher degree) coefficients.

Could you explain more, I didn't understand. Are you talking about the coseficients of predictors ?

[Carlo Lazzaro]

Jihad:
you canno have a regressand (ie, dependent variable) such as ROA² in OLS.
What I meant is that you can (say) have a squared terms for -Size- (provided that I do not know whether what follows makes any sense in your research field):

reg ROA i.CSR Risk c.Size##c.Size ib2.Industry,vce(cluster Companyscode)

[Jihad El Yaagoubi]

On what basis I will choose a term to square?

[Carlo Lazzaro]

Jihad:
usually, the literature is the favourite source.
For instance, have you read about any example with both linear and squared terms for (say) -Size- (of the firm, I assume)?

[Jihad El Yaagoubi]

No, I didn't read about that

[Jihad El Yaagoubi]

I'm trying to get more information about squared terms in a regression, I will run this model and share it. Thank you Carlo