Random effects model with dummy variable as control variable

Jihad El Yaagoubi

Join Date: May 2018
Posts: 65

#46

28 May 2018, 16:06

I tried it just to see the rsults of the ovtest, it's still significant

Code:

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

Linear regression                               Number of obs     =        139
                                                F(7, 27)          =       2.85
                                                Prob > F          =     0.0231
                                                R-squared         =     0.6765
                                                Root MSE          =     .42571

                           (Std. Err. adjusted for 28 clusters in Companyscode)
-------------------------------------------------------------------------------
              |               Robust
          ROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
        1.CSR |   .0640246   .0534338     1.20   0.241    -.0456126    .1736618
         Risk |   .6607711   .2967091     2.23   0.034     .0519743    1.269568
         Size |    2.07716   1.024637     2.03   0.053    -.0252224    4.179542
              |
c.Size#c.Size |  -.1113785   .0534507    -2.08   0.047    -.2210504   -.0017067
              |
     Industry |
           1  |  -.3233247   .1391975    -2.32   0.028    -.6089344    -.037715
           3  |  -.1931089   .1427413    -1.35   0.187    -.4859899    .0997721
           4  |  -.2692938    .148752    -1.81   0.081    -.5745076      .03592
              |
        _cons |  -9.408673   4.788201    -1.96   0.060    -19.23325    .4159049
-------------------------------------------------------------------------------

. estat ovtest

Ramsey RESET test using powers of the fitted values of ROA
       Ho:  model has no omitted variables
                 F(3, 128) =    829.19
                  Prob > F =      0.0000

.

Comment

Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#47

28 May 2018, 22:51

Jihad:
what if you log the dependent variable (ROA)?

Kind regards,
Carlo
(Stata 19.0)
Comment

Jihad El Yaagoubi

Join Date: May 2018
Posts: 65

#48

29 May 2018, 08:30

All values of ROA become negative because they represent a percentage. I'm sharing an extract from my data concerning ROA. and result of regression when logging ROA.

Code:

 Company's code
Year
CSR
ROA

1
2012
0
0.019338

1
2013
0
0.267034

1
2014
0
0.154225

1
2015
0
0.180171

1
2016
0
0.041677

2
2012
0
0.013732

2
2013
0
-0.02692

2
2014
0
0.269017

2
2015
0
0.087709

2
2016
0
0.142721

Code:

reg logROA i.CSR Risk Size ib2.Industry,vce(cluster Companyscode)
Linear regression Number of obs = 135
F(6, 27) = 2.99
Prob > F = 0.0227
R-squared = 0.1279
Root MSE = .54202
(Std. Err. adjusted for 28 clusters in Companyscode)
Robust
logROA Coef. Std. Err. t P>|t| [95% Conf. Interval]
1.CSR .0652971 .1033878 0.63 0.533 -.1468372 .2774314
Risk .1659696 .0808908 2.05 0.050 -4.50e-06 .3319438
Size -.0332269 .0549726 -0.60 0.551 -.1460213 .0795675
Industry
1 -.2474763 .1242313 -1.99 0.057 -.5023779 .0074253
3 -.027201 .1548339 -0.18 0.862 -.3448939 .2904919
4 -.2775045 .144188 -1.92 0.065 -.5733539 .0183449
_cons -1.015902 .4692717 -2.16 0.039 -1.978768 -.0530359
5 . estat ovtest
Ramsey RESET test using powers of the fitted values of logROA
Ho: model has no omitted variables
F(3, 125) = 2.07
Prob > F = 0.1070

Even if the ovtest is no longer significant, but also the relationship is not significant

Comment

Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#49

29 May 2018, 09:07

Jihad:
the lack of statistical significance in not good/bad in itself: it's just a matter of fact (by the way, your sample is not that large).
Please also note that your previous "significant" relationships were, in fact, unreliable due to model misspecification.
I would report those results, along with some concisderations on the sample size and "more research is needed".
Last but by no means least, please keep on using CODE delimiters to share codes/outcome/results with interested listers. Thanks.

Kind regards,
Carlo
(Stata 19.0)
Comment
Jihad El Yaagoubi

Join Date: May 2018

Posts: 65
#50

29 May 2018, 13:08

Is it normal to have all values of depvar negative?
If I want to report descriptive statistics, I will report about log ROA or ROA ?
Comment
Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#51

30 May 2018, 00:07

Jihad:
if ROA is expressed in percentage, logging a number which<1 in its original metric gives back a negative number.
That said, I would report ROA (no ln(ROA)) in descriptive statistics.

Kind regards,
Carlo
(Stata 19.0)
Comment
Jihad El Yaagoubi

Join Date: May 2018

Posts: 65
#52

30 May 2018, 06:39

And I must say that In the regression I used the logROA not ROA to avoid OVB ?
Comment
Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#53

30 May 2018, 06:43

Correct.

Kind regards,
Carlo
(Stata 19.0)
Comment
Jihad El Yaagoubi

Join Date: May 2018

Posts: 65
#54

30 May 2018, 07:25

Please Carlo give me a reference a book or an article that give this solution. I must add it to my report to defend the results. THANK YOU for your big help
Comment
Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#55

30 May 2018, 08:18

Jihad:
there's no universal solution.
In your case, logging the dependent variable produces no evidence of a weak specification of your regression model according to -estat ovtest-: that's all.
Chapter 8 of https://www.amazon.com/Introduction-.../dp/0321278879 covers this issue (pages: 267-276).

Kind regards,
Carlo
(Stata 19.0)
Comment

Jihad El Yaagoubi

Join Date: May 2018
Posts: 65

#56

30 May 2018, 08:47

I tried ovtest and linktest and have contradictory results

Code:

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

Linear regression                               Number of obs     =        162
                                                F(7, 27)          =       2.21
                                                Prob > F          =     0.0654
                                                R-squared         =     0.0372
                                                Root MSE          =     .56907

                          (Std. Err. adjusted for 28 clusters in Companyscode)
------------------------------------------------------------------------------
             |               Robust
      LOGROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       1.CSR |   .1160269    .093839     1.24   0.227    -.0765148    .3085686
        Size |  -.0381379    .071132    -0.54   0.596    -.1840888     .107813
        Risk |    .026861   .0184761     1.45   0.158    -.0110489    .0647709
         Age |  -.0002112   .0020855    -0.10   0.920    -.0044903     .004068
             |
    Industry |
          1  |  -.2277112   .1562495    -1.46   0.157    -.5483087    .0928862
          3  |  -.1319128    .217863    -0.61   0.550    -.5789309    .3151052
          4  |  -.1931123   .1679689    -1.15   0.260     -.537756    .1515314
             |
       _cons |  -.9144854   .6376664    -1.43   0.163    -2.222869     .393898
------------------------------------------------------------------------------

. estat ovtest

Ramsey RESET test using powers of the fitted values of LOGROA
       Ho:  model has no omitted variables
                 F(3, 151) =      2.86
                  Prob > F =      0.0387

. linktest

      Source |       SS           df       MS      Number of obs   =       162
-------------+----------------------------------   F(2, 159)       =      4.07
       Model |  2.52183304         2  1.26091652   Prob > F        =    0.0189
    Residual |  49.2791343       159  .309931662   R-squared       =    0.0487
-------------+----------------------------------   Adj R-squared   =    0.0367
       Total |  51.8009673       161  .321745139   Root MSE        =    .55672

------------------------------------------------------------------------------
      LOGROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        _hat |   10.01324   6.529626     1.53   0.127    -2.882749    22.90922
      _hatsq |   3.428235   2.478896     1.38   0.169    -1.467575    8.324045
       _cons |   5.868018   4.279266     1.37   0.172    -2.583516    14.31955
------------------------------------------------------------------------------

.

So what should I conclude ?

Comment

Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#57

30 May 2018, 09:04

Jihad:
the -linktest- usually tests the (mis)specification of the dependent variable: there's no evidence of that in your case;
- estat ovtest- check whether a misspecification exists among the predictors (as it would seem in you case).
You have increased your sample size, but your pooled OLS (see the F-test-related p-value) does not seem to differ from

Code:

mean LOGROA

You can try a different model specification, repeat -estat ovtest- and see what happens.

Kind regards,
Carlo
(Stata 19.0)
Comment

Jihad El Yaagoubi

Join Date: May 2018
Posts: 65

#58

30 May 2018, 09:24

Code:

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

Linear regression                               Number of obs     =        162
                                                F(7, 27)          =       2.21
                                                Prob > F          =     0.0654
                                                R-squared         =     0.0372
                                                Root MSE          =     .56907

                          (Std. Err. adjusted for 28 clusters in Companyscode)
------------------------------------------------------------------------------
             |               Robust
  MeanlogROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       1.CSR |   .1160269    .093839     1.24   0.227    -.0765148    .3085686
        Size |  -.0381379    .071132    -0.54   0.596    -.1840888     .107813
        Risk |    .026861   .0184761     1.45   0.158    -.0110489    .0647709
         Age |  -.0002112   .0020855    -0.10   0.920    -.0044903     .004068
             |
    Industry |
          1  |  -.2277112   .1562495    -1.46   0.157    -.5483087    .0928862
          3  |  -.1319128    .217863    -0.61   0.550    -.5789309    .3151052
          4  |  -.1931123   .1679689    -1.15   0.260     -.537756    .1515314
             |
       _cons |  -.9144854   .6376664    -1.43   0.163    -2.222869     .393898
------------------------------------------------------------------------------

. estat ovtest

Ramsey RESET test using powers of the fitted values of MeanlogROA
       Ho:  model has no omitted variables
                 F(3, 151) =      2.86
                  Prob > F =      0.0387

it's still significant, apparently, there is no model that fits my data

Comment

Carlo Lazzaro

Join Date: Apr 2014

Posts: 17673
#59

30 May 2018, 10:00

Jihad:
what I meant in my previous reply was exactly:

Code:

mean LOGROA

not the regression code you wrote!

Kind regards,
Carlo
(Stata 19.0)
Comment

Jihad El Yaagoubi

Join Date: May 2018
Posts: 65

#60

30 May 2018, 10:03

I'm sorry for the misunderstanding.

Here what I get

Code:

 mean LOGROA

Mean estimation                   Number of obs   =        163

--------------------------------------------------------------
             |       Mean   Std. Err.     [95% Conf. Interval]
-------------+------------------------------------------------
      LOGROA |  -1.377567   .0444684      -1.46538   -1.289755
--------------------------------------------------------------

Announcement

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment