[Eyup TANIL]

Hi Eyup TANIL,
there is - to the best of my knowledge - no guidance on the lag length of ARDL models in panel data models with cross-sectional dependence. My advice would be to determine the lag structure for each country individually and then compare results with different lag lengths.

Given the size of N, I would strongly advice against the use of xtdcce2, respectively the CCE estimator. The estimator and the concept of strong cross-sectional dependence are only valid if N and T are large. N=4 is certainly not large!

I do not know the Youtube videos or people who record them. In general, if you estimate an ECM model, you should take the first difference of the variables.

Hope this helps.

Dear Mr. Ditzen,

I have made updates to my dataset, and currently I have N:15 T:46 panel data. I have 4 variables;1 dependent and 3 independent variables. Among these variables, the dependent variable and 2 independent variables are integrated of order 1 (I(1)), and one variable is integrated of order 0 (I(0)). There is cross-sectional dependence.

When modeling these variables using the (CS-ARDL) xtdcce2 command, I am wondering whether to use the differences or the raw data. What are the differences or advantages between these two methods?

Additionally, you mentioned "My advice would be to determine the lag structure for each country individually and then compare results with different lag lengths." How can I perform this process ?

I run a model as like below; ı run CS-ARDL (1 0 0 0) with difference of all variables. results coeffient good for the model. but what does mean CD-istatistics and prob> F. is this model biased and have problem.

[JanDitzen]

I have made updates to my dataset, and currently I have N:15 T:46 panel data. I have 4 variables;1 dependent and 5 independent variables. Among these variables, the dependent variable and 2 independent variables are integrated of order 1 (I(1)), and one variable is integrated of order 0 (I(0)). There is cross-sectional dependence.

N = 15 is much better, but it is still on the lower end. I would be very careful drawing conclusions regarding cross-sectional dependence and mean group coefficients. The small N will bias results.

When modeling these variables using the (CS-ARDL) xtdcce2 command, I am wondering whether to use the differences or the raw data. What are the differences or advantages between these two methods?

In a nutshell, the ECM assumes a equilibrium with short term deviations from it. This allows the data to be integrated and the short term deviations are the first differences. The advantage of the ARDL model is that it does not require this assumption.

Additionally, you mentioned "My advice would be to determine the lag structure for each country individually and then compare results with different lag lengths." How can I perform this process ?

I was thinking about a simple time series model. You select the lag length for each country, which is essentially a time series.

I run a model as like below; ı run CS-ARDL (1 0 0 0) with difference of all variables. results coeffient good for the model. but what does mean CD-istatistics and prob> F. is this model biased and have problem.

The CD test tells you that the null hypothesis of weak cross-sectional dependence is not rejected. The alternative is strong dependence. The F-test is the standard F test for joint significance.

[Eyup TANIL]

So, according to the F-test results, my results are insignificant. Am I concluding that the model I have created is producing meaningless results?

[Eyup TANIL]

JanDitzen can we select cr_lags(0)? ıf we found cd-statisctic p value above 0,10 . here is the my results . ıf ı take cr_lags(0) cd statistic p value and F statistic good for model.

xtdcce2 d.ef, lr(L(1/3)d.ef L(0/2).d.fdi L(0/2).d.gdper L(0/2).d.gdper2 ) lr_options(ardl) cr(d.ef d.fdi d.gdper d.gdper2) cr_lags(0) fullsample
(Dynamic) Common Correlated Effects Estimator - (CS-ARDL)

Panel Variable (i): id Number of obs = 705
Time Variable (t): year Number of groups = 15

Degrees of freedom per group: Obs per group (T) = 47
without cross-sectional averages = 34
with cross-sectional averages = 30
Number of F(255, 450) = 1.22
cross-sectional lags = 0 Prob > F = 0.03
variables in mean group regression = 180 R-squared = 0.59
variables partialled out = 75 R-squared (MG) = 0.49
Root MSE = 0.29
CD Statistic = -1.37
p-value = 0.1700
-------------------------------------------------------------------------------
D.ef| Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------+---------------------------------------------------------------
Short Run Est.|
---------------+---------------------------------------------------------------
Mean Group: |
LD.ef| -.2239098 .0471513 -4.75 0.000 -.3163246 -.1314949
L2D.ef| -.0793347 .0372199 -2.13 0.033 -.1522843 -.0063851
L3D.ef| -.0936635 .0333912 -2.81 0.005 -.1591092 -.0282179
D.fdi| -.0148872 .0087699 -1.70 0.090 -.0320759 .0023015
D.gdper| .0002672 .0001431 1.87 0.062 -.0000133 .0005477
D.gdper2| -2.50e-09 2.03e-09 -1.23 0.218 -6.47e-09 1.47e-09
LD.fdi| -.0156117 .0113154 -1.38 0.168 -.0377896 .0065661
L2D.fdi| -.0275879 .0139804 -1.97 0.048 -.054989 -.0001868
LD.gdper| -.0000142 .0001173 -0.12 0.904 -.0002442 .0002158
L2D.gdper| .0000767 .0001004 0.76 0.445 -.0001201 .0002735
LD.gdper2| 1.52e-09 1.61e-09 0.95 0.344 -1.63e-09 4.68e-09
L2D.gdper2| -8.08e-10 1.98e-09 -0.41 0.684 -4.69e-09 3.08e-09
---------------+---------------------------------------------------------------
Adjust. Term |
---------------+---------------------------------------------------------------
Mean Group: |
lr_ef| -1.396908 .0792967 -17.62 0.000 -1.552327 -1.241489
---------------+---------------------------------------------------------------
Long Run Est. |
---------------+---------------------------------------------------------------
Mean Group: |
lr_fdi| -.0373162 .018993 -1.96 0.049 -.0745418 -.0000906
lr_gdper| .0002351 .0000953 2.47 0.014 .0000483 .0004218
lr_gdper2| -1.55e-09 1.81e-09 -0.86 0.392 -5.10e-09 2.00e-09
-------------------------------------------------------------------------------
Mean Group Variables: LD.ef L2D.ef L3D.ef D.fdi D.gdper D.gdper2 LD.fdi L2D.fdi LD.gdper L2D.gdper LD.gdper2 L2D.gdper2 lr_fdi lr_gdper lr_gdper2
Cross Sectional Averaged Variables: D.ef D.fdi D.gdper D.gdper2
Long Run Variables: lr_fdi lr_gdper lr_gdper2
Adjustment variable(s): lr_ef (LD.ef L2D.ef L3D.ef)
Heterogenous constant partialled out.

But I select cr_lags(3) as pesaran said. ( my data set T:51 [T^(1/3)]=3,66 ). is my model prob>F 1.00 is not good, and then CD sitatistic p value 0.03 ; What can I do this stiuation ?

And my last questions , how can ı see aic information critera for CS_ARDL results?

[JanDitzen]

You can use cr_lags(0), though following theory you should add lags. I would compare how the CD test statistic and the coefficients change if you add further lags.

I would also investigate why some of the coefficients are very small. Is this because the individual coefficients are equally small or is it because the individual coefficients are distributed around essentially zero.

[Eyup TANIL]

Firstly , thank JanDitzen for feedback, How can ı determine my varieblas lag lenth? can we compare aic infermation critera for lag selection. but ı dont know how to calculate aic info cretirea in CS-ARLD model? are you have any code for that?


I understand from your response that Ican select "cr_lag(0)" and if the probability value of the CD statistic is greater than 0.10 and F is less than 0.10, then I can use this model for my article. Am I understanding correctly?

[JanDitzen]

To the best of my knowledge there are no lag selection criteria for such a setting. One can of course use AIC/BIC, but I am unsure if they are valid in the presence of cross-section dependence.

Regarding your second question: yes, but I would present results with higher order lags as well.