log likelihood info pooled EGLS(cross section SUR)

[carmomi]

Dear all,

I'm estimating panel data models using the pool object because it returns the covariance matrix and the correlation matrix. However, I notice that the output does not included log likelihood when the estimation is done using cross section sur weights (only with pooled least squares).
Is the model estimated as a panel exactly the same (selecting the same options) so I can get the covar and correl matrixs from the pool and the log likelihood from the panel?

P.S I need the log likelihoods of different specifications to test covariance hypotheses..

Many thansk in advance,
Miguel

[EViews Glenn]

If the coefficient estimates are the same it's estimating the same model. My guess it's that the panel isn't reporting the likelihood since the covariance estimate isn't iterated to convergence so it's not ML.

[carmomi]

Problem solved. Many thanks Glenn. I did not select the iterate box...
Miguel

[carmomi]

Dear Glenn,

Still one related issue for which I would much appreciate any help that you can provide.

I got the log likelihood values of 2 models and used them to compare the models' hypothesis by means of a likelihood ratio test. The result of the ratio is negative, which is fine (-1381.81).

Knowing that the distribution of the likelihood ratio is asymptotically Chi-Square I have tried to calculate lr_pval=1-@cchisq(-1381.81,1) but I got an error message "positive or non-negative argument to funtion expected".

Any hint on what I'm doing wrong?

Many thanks,
Miguel

[startz]

You've probably switched the order of the log likelihoods in computing the statistic.

[carmomi]

I´ve checked with another pair of nested models and the problem persists.
model 1 (more restricted) with log likelihood 247,5779
model 2 (less restricted) with log likelihood -234.3651

Thus apllying the likelihood ratio test: 2X(log likelihood(m2)-log likelihood(m1))= -963,886
Knowing that the likelihood ratio is asymptotically Chi-Square it would be useful to compute the critical value and p values (e.g P value=1-@cchisq(-963,886,10) but with a negative result I have no idea on what can I do.

Any further help/guidance would be very much appreciated.
Many Thanks
Miguel

[startz]

I´ve checked with another pair of nested models and the problem persists.
model 1 (more restricted) with log likelihood 247,5779
model 2 (less restricted) with log likelihood -234.3651


Any further help/guidance would be very much appreciated.
Many Thanks
Miguel

Something's wrong, although it's hard to say what it is. If you're doing maximum likelihood, the less restricted likelihood has to be higher than the restricted likelihood. Check to be sure that both estimates are using identical data as a start. Past that, you might post your results and see if anyone notices what's wrong.

[carmomi]

The data is the same.

less restrictive model results:
Dependent Variable: LOG(LONG_RUN_COST?)
Method: Pooled EGLS (Cross-section weights)
Date: 04/01/12 Time: 18:03
Sample: 2014 2028
Included observations: 15
Cross-sections included: 11
Total pool (balanced) observations: 165
Iterate weights to convergence
Convergence achieved after 1 weight iterations

Variable Coefficient Std. Error t-Statistic Prob.

C -3.757966 1.735350 -2.165538 0.0320
1--LOG(TT1) 0.650797 0.064424 10.10185 0.0000
2--LOG(TT2) 0.150792 0.061744 2.442240 0.0158
3--LOG(TT3) 0.575433 0.045810 12.56130 0.0000
4--LOG(TT4) 4.669496 0.684239 6.824365 0.0000
5--LOG(TT5) 0.445847 0.121654 3.664885 0.0003
6--LOG(TT6) 0.881914 0.010192 86.53370 0.0000
7--LOG(TT7) 1.960630 0.485182 4.041017 0.0001
8--LOG(TT8) 0.277291 0.264580 1.048042 0.2964
9--LOG(TT9) 0.146530 0.081256 1.803316 0.0734
10--LOG(TT10) 0.793821 0.175719 4.517555 0.0000
11--LOG(TT11) 0.933341 0.115100 8.108977 0.0000
Fixed Effects (Cross)
1--C 8.783593
2--C 19.12414
3--C 9.795011
4--C -76.61697
5--C 11.80479
6--C 3.798947
7--C -18.70791
8--C 15.40867
9--C 18.18730
10--C 6.071297
11--C 2.351135

Effects Specification

Cross-section fixed (dummy variables)

Weighted Statistics

R-squared 0.999407 Mean dependent var 60.44414
Adjusted R-squared 0.999319 S.D. dependent var 58.46600
S.E. of regression 0.057968 Akaike info criterion 3.107456
Sum squared resid 0.480525 Schwarz criterion 3.521582
Log likelihood -234.3651 Hannan-Quinn criter. 3.275564
F-statistic 11468.35 Durbin-Watson stat 1.344303
Prob(F-statistic) 0.000000

Unweighted Statistics

R-squared 0.993634 Mean dependent var 17.80794
Sum squared resid 0.480525 Durbin-Watson stat 1.050816


more restrictive model results:
Dependent Variable: LOG(LONG_RUN_COST?)
Method: Pooled Least Squares
Date: 04/01/12 Time: 18:06
Sample: 2014 2028
Included observations: 15
Cross-sections included: 11
Total pool (balanced) observations: 165

Variable Coefficient Std. Error t-Statistic Prob.

C -3.757966 2.876193 -1.306577 0.1935
1--LOG(TT1) 0.650797 0.147418 4.414623 0.0000
2--LOG(TT2) 0.150792 0.233140 0.646791 0.5188
3--LOG(TT3) 0.575433 0.330343 1.741927 0.0837
4--LOG(TT4) 4.669496 1.292161 3.613711 0.0004
5--LOG(TT5) 0.445847 0.135498 3.290425 0.0013
6--LOG(TT6) 0.881914 0.108532 8.125866 0.0000
7--LOG(TT7) 1.960630 0.189612 10.34020 0.0000
8--LOG(TT8) 0.277291 0.289077 0.959230 0.3391
9--LOG(TT9) 0.146530 0.546599 0.268075 0.7890
10--LOG(TT10) 0.793821 0.168287 4.717068 0.0000
11--LOG(TT11) 0.933341 0.108432 8.607581 0.0000
Fixed Effects (Cross)
1--C 8.783593
2--C 19.12414
3--C 9.795011
4--C -76.61697
5--C 11.80479
6--C 3.798947
7--C -18.70791
8--C 15.40867
9--C 18.18730
10--C 6.071297
11--C 2.351135

Effects Specification

Cross-section fixed (dummy variables)

R-squared 0.993634 Mean dependent var 17.80794
Adjusted R-squared 0.992699 S.D. dependent var 0.678427
S.E. of regression 0.057968 Akaike info criterion -2.734277
Sum squared resid 0.480525 Schwarz criterion -2.320151
Log likelihood 247.5779 Hannan-Quinn criter. -2.566169
F-statistic 1062.863 Durbin-Watson stat 1.050816
Prob(F-statistic) 0.000000

Is it possible that the "no weights" model is less restrictive than the "cross section weights" (in terms of the restrictions imposed in the covariance matrix), contrary to what I have assumed?

Many thanks

[startz]

First, you appear to have done everything correctly.

Second, the results are weird. The coefficient estimates are identical. The sum of squared residuals is identical. But the F-stat is very different.

Third, it looks like there is a sign error in the log likelihood in the restricted model...but it's hard to imagine how that would come about.

So...grasping at straws...what's your version number and build date of EViews?

[carmomi]

Many thanks again.

It's Eviews version 7.1
Standard edition Nov 3 November 2010 build

I have no idea where the problem may lie...

[EViews Gareth]

You should update to the latest version. No guarantee it will work, but worth doing anyway.

[carmomi]

I have updated and the problem persists.

I have compared the regression outputs of the pool with the panel and I suspect that there is a bug in the estimation of the log likelihood of the pool with cross section weights and fixed effects. Below you can find the comparison the results of the pool and panel results with 3 alternative specifications - the negative sign of the log likelihood of the pool with cross section weights and fixed effects should be wrong.


Dependent Variable: LOG(LONG_RUN_COST)
Method: Panel Least Squares
Date: 04/01/12 Time: 19:41
Sample: 2014 2028
Periods included: 15
Cross-sections included: 11
Total panel (balanced) observations: 165

Variable Coefficient Std. Error t-Statistic Prob.

C 1.321578 1.233373 1.071515 0.2856
LOG(TT) 0.810813 0.060658 13.36701 0.0000

Effects Specification

Cross-section fixed (dummy variables)
R-squared 0.990551 Mean dependent var 17.80794
Adjusted R-squared 0.989872 S.D. dependent var 0.678427
S.E. of regression 0.068275 Akaike info criterion -2.460592
Sum squared resid 0.713211 Schwarz criterion -2.234705
Log likelihood 214.9989 Hannan-Quinn criter. -2.368897
F-statistic 1458.169 Durbin-Watson stat 0.722826
Prob(F-statistic) 0.000000

Dependent Variable: LOG(LONG_RUN_COST?)
Method: Pooled Least Squares
Date: 04/01/12 Time: 19:42
Sample: 2014 2028
Included observations: 15
Cross-sections included: 11
Total pool (balanced) observations: 165

Variable Coefficient Std. Error t-Statistic Prob.

C 1.321578 1.233373 1.071515 0.2856
LOG(TT?) 0.810813 0.060658 13.36701 0.0000
Fixed Effects (Cross)
PPP1--C 0.317840
PPP10--C 0.651481
PPP11--C -0.271145
PPP2--C 0.117268
PPP3--C -0.132104
PPP4--C -0.860633
PPP5--C -0.293797
PPP6--C 0.197051
PPP7--C 0.152002
PPP8--C -0.092458
PPP9--C 0.214494

Effects Specification

Cross-section fixed (dummy variables)

R-squared 0.990551 Mean dependent var 17.80794
Adjusted R-squared 0.989872 S.D. dependent var 0.678427
S.E. of regression 0.068275 Akaike info criterion -2.460592
Sum squared resid 0.713211 Schwarz criterion -2.234705
Log likelihood 214.9989 Hannan-Quinn criter. -2.368897
F-statistic 1458.169 Durbin-Watson stat 0.722826
Prob(F-statistic) 0.000000

Dependent Variable: LOG(LONG_RUN_COST)
Method: Panel EGLS (Cross-section weights)
Date: 04/01/12 Time: 19:43
Sample: 2014 2028
Periods included: 15
Cross-sections included: 11
Total panel (balanced) observations: 165
Iterate weights to convergence
Convergence achieved after 12 weight iterations

Variable Coefficient Std. Error t-Statistic Prob.

C 0.125637 0.206062 0.609706 0.5430
LOG(TT) 0.869631 0.010134 85.81242 0.0000

Effects Specification

Cross-section fixed (dummy variables)

Weighted Statistics

R-squared 0.998003 Mean dependent var 47.87619
Adjusted R-squared 0.997860 S.D. dependent var 59.80508
S.E. of regression 0.068485 Akaike info criterion -3.515031
Sum squared resid 0.717594 Schwarz criterion -3.289144
Log likelihood 301.9901 Hannan-Quinn criter. -3.423336
F-statistic 6952.374 Durbin-Watson stat 0.775115
Prob(F-statistic) 0.000000

Unweighted Statistics

R-squared 0.990493 Mean dependent var 17.80794
Sum squared resid 0.717594 Durbin-Watson stat 0.719515

Dependent Variable: LOG(LONG_RUN_COST?)
Method: Pooled EGLS (Cross-section weights)
Date: 04/01/12 Time: 19:43
Sample: 2014 2028
Included observations: 15
Cross-sections included: 11
Total pool (balanced) observations: 165
Iterate weights to convergence
Convergence achieved after 12 weight iterations

Variable Coefficient Std. Error t-Statistic Prob.

C 0.125637 0.206062 0.609706 0.5430
LOG(TT?) 0.869631 0.010134 85.81242 0.0000
Fixed Effects (Cross)
PPP1--C 0.269111
PPP10--C 0.669585
PPP11--C -0.254775
PPP2--C 0.072083
PPP3--C -0.147483
PPP4--C -0.896863
PPP5--C -0.229033
PPP6--C 0.170621
PPP7--C 0.123350
PPP8--C -0.045429
PPP9--C 0.268832

Effects Specification

Cross-section fixed (dummy variables)

Weighted Statistics

R-squared 0.998003 Mean dependent var 47.87619
Adjusted R-squared 0.997860 S.D. dependent var 59.80508
S.E. of regression 0.068485 Akaike info criterion 2.987681
Sum squared resid 0.717594 Schwarz criterion 3.213568
Log likelihood -234.4837 Hannan-Quinn criter. 3.079376
F-statistic 6952.374 Durbin-Watson stat 0.775115
Prob(F-statistic) 0.000000

Unweighted Statistics

R-squared 0.990493 Mean dependent var 17.80794
Sum squared resid 0.717594 Durbin-Watson stat 0.719515

Dependent Variable: LOG(LONG_RUN_COST)
Method: Panel EGLS (Cross-section SUR)
Date: 04/01/12 Time: 19:45
Sample: 2014 2028
Periods included: 15
Cross-sections included: 11
Total panel (balanced) observations: 165
Iterate weights to convergence
Convergence not achieved after 500 weight iterations

Variable Coefficient Std. Error t-Statistic Prob.

C 0.134876 0.062327 2.164014 0.0320
LOG(TT) 0.869177 0.003065 283.5642 0.0000

Effects Specification

Cross-section fixed (dummy variables)

Weighted Statistics
R-squared 0.999966 Mean dependent var 27.94184
Adjusted R-squared 0.999963 S.D. dependent var 2850.367
S.E. of regression 1.038475 Akaike info criterion -4.987272
Sum squared resid 165.0000 Schwarz criterion -4.761385
Log likelihood 423.4500 Hannan-Quinn criter. -4.895577
F-statistic 406743.5 Durbin-Watson stat 2.075511
Prob(F-statistic) 0.000000

Unweighted Statistics

R-squared 0.990494 Mean dependent var 17.80794
Sum squared resid 0.717526 Durbin-Watson stat 0.719571

Dependent Variable: LOG(LONG_RUN_COST?)
Method: Pooled EGLS (Cross-section SUR)
Date: 04/01/12 Time: 19:45
Sample: 2014 2028
Included observations: 15
Cross-sections included: 11
Total pool (balanced) observations: 165
Iterate weights to convergence
Convergence not achieved after 500 weight iterations

Variable Coefficient Std. Error t-Statistic Prob.
C 0.134876 0.062327 2.164014 0.0320
LOG(TT?) 0.869177 0.003065 283.5642 0.0000
Fixed Effects (Cross)
PPP1--C 0.269487
PPP10--C 0.669445
PPP11--C -0.254901
PPP2--C 0.072432
PPP3--C -0.147364
PPP4--C -0.896583
PPP5--C -0.229533
PPP6--C 0.170826
PPP7--C 0.123572
PPP8--C -0.045792
PPP9--C 0.268412

Effects Specification

Cross-section fixed (dummy variables)

Weighted Statistics

R-squared 0.999966 Mean dependent var 466.3324
Adjusted R-squared 0.999963 S.D. dependent var 2811.865
S.E. of regression 1.038475 Akaike info criterion -4.987272
Sum squared resid 165.0000 Schwarz criterion -4.761385
Log likelihood 423.4500 Hannan-Quinn criter. -4.895577
F-statistic 406743.5 Durbin-Watson stat 2.075511
Prob(F-statistic) 0.000000

Unweighted Statistics

R-squared 0.990494 Mean dependent var 17.80794
Sum squared resid 0.717526 Durbin-Watson stat 0.719571

[EViews Glenn]

Can you please post your workfile.

[EViews Glenn]

Bug in the pool computation when you iterate to convergence. Sorry for the inconvenience and thanks for your help in finding this. Fixed and will be posted shortly.