[Carlo Lazzaro]

Qurat:
in long T, small N panel datasets you can add the -i.panelid- and a squared term for -timevar- to investigate potential turning point, as in the following toy-example:

. use http://www.stata-press.com/data/r15/invest2

. xtgls invest market stock i.company c.time##c.time

Cross-sectional time-series FGLS regression

Coefficients:  generalized least squares
Panels:        homoskedastic
Correlation:   no autocorrelation

Estimated covariances      =         1          Number of obs     =        100
Estimated autocorrelations =         0          Number of groups  =          5
Estimated coefficients     =         9          Time periods      =         20
                                                Wald chi2(8)      =    1505.04
Log likelihood             =  -561.653          Prob > chi2       =     0.0000

-------------------------------------------------------------------------------
       invest |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
       market |   .1067342   .0154291     6.92   0.000     .0764938    .1369745
        stock |    .360213   .0326232    11.04   0.000     .2962727    .4241533
              |
      company |
           2  |   56.58426   58.67971     0.96   0.335    -58.42586    171.5944
           3  |  -160.9345   41.61103    -3.87   0.000    -242.4907   -79.37841
           4  |   28.55782   59.07646     0.48   0.629    -87.22991    144.3455
           5  |   175.1112    41.3142     4.24   0.000      94.1369    256.0856
              |
         time |     1.1721   4.913048     0.24   0.811    -8.457298     10.8015
              |
c.time#c.time |  -.0989248   .2392159    -0.41   0.679    -.5677793    .3699297
              |
        _cons |   -86.2353   70.45214    -1.22   0.221     -224.319    51.84836
-------------------------------------------------------------------------------

.

As the outcome does not support the evidence for turning point, the linear term (or -i.year-) is probably the way to go:

. xtgls invest market stock i.company i.time

Cross-sectional time-series FGLS regression

Coefficients:  generalized least squares
Panels:        homoskedastic
Correlation:   no autocorrelation

Estimated covariances      =         1          Number of obs     =        100
Estimated autocorrelations =         0          Number of groups  =          5
Estimated coefficients     =        26          Time periods      =         20
                                                Wald chi2(25)     =    1852.07
Log likelihood             = -551.8661          Prob > chi2       =     0.0000

------------------------------------------------------------------------------
      invest |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      market |   .1260306   .0199352     6.32   0.000     .0869582    .1651029
       stock |   .3617764    .030956    11.69   0.000     .3011038    .4224491
             |
     company |
          2  |   127.6598   71.85344     1.78   0.076    -13.17037      268.49
          3  |  -114.3793   49.70343    -2.30   0.021    -211.7962   -16.96232
          4  |   100.1193   72.18316     1.39   0.165    -41.35707    241.5957
          5  |   221.2348   48.80957     4.53   0.000     125.5698    316.8998
             |
        time |
          2  |  -39.55323   40.75388    -0.97   0.332    -119.4294    40.32291
          3  |  -88.14978   45.33628    -1.94   0.052    -177.0073    .7077036
          4  |  -68.78843   38.55563    -1.78   0.074    -144.3561    6.779217
          5  |  -119.1439   40.40383    -2.95   0.003    -198.3339   -39.95381
          6  |  -89.04318   41.24527    -2.16   0.031    -169.8824   -8.203932
          7  |  -23.11078   40.68394    -0.57   0.570    -102.8498    56.62828
          8  |   -13.7191   38.85216    -0.35   0.724    -89.86795    62.42974
          9  |  -54.27664   39.62219    -1.37   0.171    -131.9347    23.38144
         10  |  -56.97186   39.81186    -1.43   0.152    -135.0017    21.05794
         11  |  -73.18178   41.14196    -1.78   0.075    -153.8185    7.454967
         12  |  -31.06858   41.96616    -0.74   0.459    -113.3207    51.18358
         13  |  -30.44892   39.47631    -0.77   0.441    -107.8211    46.92323
         14  |  -35.47242   39.77744    -0.89   0.373    -113.4348    42.48992
         15  |  -85.85061   40.23306    -2.13   0.033     -164.706   -6.995251
         16  |  -81.26772   40.58741    -2.00   0.045    -160.8176   -1.717866
         17  |  -63.89134   42.50002    -1.50   0.133    -147.1899    19.40717
         18  |  -62.64179   43.57669    -1.44   0.151    -148.0505    22.76696
         19  |  -68.87581   47.70983    -1.44   0.149    -162.3853    24.63373
         20  |  -109.4707   48.44522    -2.26   0.024    -204.4215   -14.51978
             |
       _cons |  -113.0192   76.14074    -1.48   0.138    -262.2523    36.21389
------------------------------------------------------------------------------

. testparm i.time

 ( 1)  2.time = 0
 ( 2)  3.time = 0
 ( 3)  4.time = 0
 ( 4)  5.time = 0
 ( 5)  6.time = 0
 ( 6)  7.time = 0
 ( 7)  8.time = 0
 ( 8)  9.time = 0
 ( 9)  10.time = 0
 (10)  11.time = 0
 (11)  12.time = 0
 (12)  13.time = 0
 (13)  14.time = 0
 (14)  15.time = 0
 (15)  16.time = 0
 (16)  17.time = 0
 (17)  18.time = 0
 (18)  19.time = 0
 (19)  20.time = 0

           chi2( 19) =   22.09
         Prob > chi2 =    0.2796

.

-testparm- outcome does not support the evidence of the joint significance of -i.time-.

[ashar awan]

Qurat ul Ain could you please share do.file in email at aaajku @ gmail .com

[Rui Zha]

Hi. I want to jump in. If the Hausman test indicates the FE model is used, and after testing for hetero, auto, and cross sectional dependence all 3 are there. Can I use xtgls dep indep, panels(hetero) corr(ar1)? Provided I have a long panel. My T is greater than N?

Hi zahra, you got the answer?

[Saom Shawleen]

Hello Mr. Lazzaro,

I am very new to stata and doing undergrad research. My dependent variable is NPLs (Non-performing Loans) and 6 independent variables - M3 = Money Supply, ER= Exchange Rate, GDP= GDP growth rate, BUG= Govt Budget, INF =Inflastion Rate, and DEBT = Public Debt. I have chosen 8 countries and the period of 12 years. After the Hausman test when I run the fixed effects model, I did not have any cross-sectional interdependence or autocorrelation but I have only a heteroscedasticity problem in the panel data. Now can I remove heteroscedasticity by using command- xtgls NPLs M3 ER GDP BUG INF DEBT under the FGLS method? To inform you that- I have a strongly balanced dataset which is T>N.

note: I am very noobie so I run the Hausman test only. Please help me!

[Carlo Lazzaro]

Saom:
if you have a T>N panel dataset and you're interested in the -fe- estimator, I would consider -xtregar-, that, unfortunately, does not allow you to manage heteroskedasticity.
Despite probably being a bit too near to the limit, I would stick with -xtreg,fe- with robust or clustered standard errors (both option do the very same job under -xtreg-).

[Saom Shawleen]

Thank you for your response! by this time I have run -xtreg,fe- with both robust and clustered standard errors. but unfortunately, it is showing that- Prob > F is 0.086, which indicates that this is not statistically significant and one and only the P-value of GDP is significant (table below).

Regression results

NPLs	Coef.		St.Err.	t-value		p-value	[95% Conf		Interval]		Sig
M3	.0003		.165	0.00		.999	-.389		.39
ER	-.0572		.078	-0.73		.486	-.241		.127
GDP	-.7327		.251	-2.91		.023	-1.327		-.138		**
BUG	.2072		.197	1.05		.329	-.26		.674
INF	-.1173		.064	-1.84		.109	-.268		.034
DEBT	-.0335		.053	-0.63		.547	-.159		.092
Constant	19.709		6.457	3.05		.019	4.439		34.978		**
Mean dependent var		7.928			SD dependent var			5.947
R-squared		0.190			Number of obs			96
F-test		3.109			Prob > F			0.086
Akaike crit. (AIC)		570.403			Bayesian crit. (BIC)			585.789
*** p<.01, ** p<.05, * p<.1

In this circumstance, I have little bit changed the approach. Now, at first, I have run the Pooled OLS, then I used Breusch-Pagan / Cook-Weisberg test for heteroskedasticity, and again I found my dataset is heteroskedastic. After that, I created Log value for my dependent variable which is NPLs2 (Non-performing Loans). Then with that Log dependent variable, I run the Hausman test. Even though this time Hausman test suggested the Random Effect Model, but now Prob > chi2 is showing 0.000 and 4 of the independent variables are showing significance (table below). Now my question is my new approach correct?

NPLs2	Coef.		St.Err.	t-value		p-value	[95% Conf		Interval]		Sig
M3	-.02		.009	-2.26		.024	-.036		-.003		**
ER	-.003		.004	-0.62		.538	-.011		.006
GDP	-.105		.017	-6.07		0	-.139		-.071		***
BUG	.039		.023	1.70		.09	-.006		.084		*
INF	-.036		.013	-2.79		.005	-.061		-.011		***
DEBT	-.002		.006	-0.40		.689	-.014		.009
Constant	4.282		.627	6.83		0	3.053		5.51		***
Mean dependent var		1.825			SD dependent var			0.757
Overall r-squared		0.220			Number of obs			96
Chi-square		43.712			Prob > chi2			0.000
R-squared within		0.342			R-squared between			0.135
*** p<.01, ** p<.05, * p<.1

[Carlo Lazzaro]

Saom:
did you run the last codes you provided with Stata?

[Saom Shawleen]

Sorry..I could not understand your question. Can you please elaborate on the code that you are asking about? actually, I have very little idea about this research software. It would be a great help if you walk me through it. In the last answer, I tried to explain that I have run OLS model first and then created a Log variable for my dependent variable (NPLs2), and instead of NPLs, I used log dependent variable (NPLs2) to conduct Hausman Test and then Hausman Test suggested me the RE model and I posted the RE model table in my last reply.

[Carlo Lazzaro]

Saom:
1) Regression results: these are the results from an OLS (not from -xtreg,fe-).Moreover, this is not the usual outcome table that Stata gives back. That said, if you have panel data and you consider the observations as independent, your estimates are unreliable, as you actually fail to consider that panel structure of your dataset.
2) the second table (which, again, does not resmple the usual outcome table that Stata gives back) refers to the outcome of a (fixed effect?) panel data regression.
As a general advice, panel data commands are the first choice when you deal with panel datasets.

[Saom Shawleen]

I declared the dataset as panel data first and then I did all the tests. here, please go through attached the pictures. I am sending all the pictures and due to the picture limit issue I am posting another reply which contains the first table that I mentioned in my reply.

1) Regression results (table-1): FE robustness

[Saom Shawleen]

This is the robustness table about which I was talking. and after that, I went for a new approach. I am sending the pictures of the new approach as well

[Saom Shawleen]

The second table: please refer to the reply (In this circumstance, I have little bit changed the approach. Now, at first, I have run the Pooled OLS, then I used Breusch-Pagan / Cook-Weisberg test for heteroskedasticity, and again I found my dataset is heteroskedastic. After that, I created Log value for my dependent variable which is NPLs2 (Non-performing Loans). Then with that Log dependent variable, I run the Hausman test. Even though this time Hausman test suggested the Random Effect Model, but now Prob > chi2 is showing 0.000 and 4 of the independent variables are showing significance (table below). Now my question is my new approach correct?)

[Saom Shawleen]

based on the suggestion of the Hausman test, I again run that Random effect model and found significance in the P value.

[Carlo Lazzaro]

Saom:
two issues here:
1) if you have panel data with a continuous regressand, you should go -xtreg- first;
2) if you detect heteroskedsticity and/or autocorrelation, you should impose -robust- or -vce(cluster clusterid)- standard errors, that -hausman- does not support.
Hence you should compare -fe- vs. -re- via the community-contributed module -xtoverid- (just type -search xtoverid- to spot and install it).
As an aside, please do not post screenshots, but use CODE delimiters insteas (as per FAQ). Thanks.

[Saom Shawleen]

Saom:
two issues here:
1) if you have panel data with a continuous regressand, you should go -xtreg- first;
2) if you detect heteroskedsticity and/or autocorrelation, you should impose -robust- or -vce(cluster clusterid)- standard errors, that -hausman- does not support.
Hence you should compare -fe- vs. -re- via the community-contributed module -xtoverid- (just type -search xtoverid- to spot and install it).
As an aside, please do not post screenshots, but use CODE delimiters insteas (as per FAQ). Thanks.

At first, I am extremely apologized for sending the screenshots.

I am dealing with a panel dataset which is T>N. I run go -xtreg- first; Then I tried my best to use -xtoverid- command but I found severe difficulties while dealing with -xtoverid-. It requires creating some dummy variables which I was unable to do. Hence, I run Hausman test and Hausman test suggested me to follow fixed effect model. Based on fixed-effect model, I found there is no autocorrelation and cross-sectional dependence but only the heteroscedasticity problem is there. I run both -robust- and -vce(cluster clusterid)- standard errors but this time the result shows that Prob>F is .0818 which is not statistically significant.

Please, suggest me what should I do now. Thank you.