Help for CS-ARDL and CS-DL

[Mohammad Qabaja]

Hello, I need help with the (PMG ARDL), (CS-ARDL) and (CS-DL) with the following Panel model specification:-

The T= 32 years
N= 11 countries
variables = 7 variables
models = 4 models

The questions are the following:-

1- I perform the command (forval i = ....) to know the suitable number of lags and it gives for all the 4 models the following number of lags (1,0,1,0) and I think it works for the ARDL method but can I apply the same number of lags for CS-DL?

2- If I want to perform the (ARDL PMG) and (CS-ARDL) command with a number of lags (1,0,1,0) how the command will be?

3- If I want to perform the (CS-DL) command with a number of lags (1,0,1,0) how the command will be? and did the CS-DL command including pooled or just the Mean group command?

I have take a look at Jan Ditzen commands but I don't find a solution for my case and you can help me by writing the command with (Y,x1,x2,x3) for more understanding.

Thank you all

[JanDitzen]

Hi Mohammad,
first of all let me point you towards the working paper which explains how to estimate the CS-DL, CS-ARDL and CS-ECM using xtdcce2. The paper is forthcoming in the next issue of The Stata Journal and can be downloaded here:https://ideas.repec.org/p/bzn/wpaper/bemps81.html. I believe you will find the answers in the paper.

Let me point out one important fact: all three estimator are based on large N, large T. While in theory this means that N,T increase to infinity, practice is a bit less clear. It mostly depends on having enough observations to estimate the model. From my experience the data you are having is at the very lower end, especially the cross-section dimension. Your N is rather small than large! Mean group estimates will be biased using an estimator with 11 units (think of how adding or removing an additional unit will change the mean group estimates). In addition the concept of cross-section dependence which the CS estimator take care of is designed for large N (and T) as well. Thus identifying and discussing if you have strong cross-sectional dependence will be hard as well.

My suggestion would be to see how the estimators compare to estimators without cross-section averages and with pooled coefficients. Hence estimating a panel ARDL and DL rather CS-ARDL and CS-DL.

In any case, if you want to use those results for a dissertation, thesis, report or publication, you have to address and discuss the issue of using a large N estimator in a small N dataset!

I hope this helps.

[Mohammad Qabaja]

Hi Mohammad,
first of all let me point you towards the working paper which explains how to estimate the CS-DL, CS-ARDL and CS-ECM using xtdcce2. The paper is forthcoming in the next issue of The Stata Journal and can be downloaded here:https://ideas.repec.org/p/bzn/wpaper/bemps81.html. I believe you will find the answers in the paper.

Let me point out one important fact: all three estimator are based on large N, large T. While in theory this means that N,T increase to infinity, practice is a bit less clear. It mostly depends on having enough observations to estimate the model. From my experience the data you are having is at the very lower end, especially the cross-section dimension. Your N is rather small than large! Mean group estimates will be biased using an estimator with 11 units (think of how adding or removing an additional unit will change the mean group estimates). In addition the concept of cross-section dependence which the CS estimator take care of is designed for large N (and T) as well. Thus identifying and discussing if you have strong cross-sectional dependence will be hard as well.

My suggestion would be to see how the estimators compare to estimators without cross-section averages and with pooled coefficients. Hence estimating a panel ARDL and DL rather CS-ARDL and CS-DL.

In any case, if you want to use those results for a dissertation, thesis, report or publication, you have to address and discuss the issue of using a large N estimator in a small N dataset!

I hope this helps.

Thank you JanDitzen for the reply.

I forgot to write about the Cross-sectional dependence test that I applied and gives significant (P=0.00) results and for that, I want to apply the CS-ARDL and CS-DL.

I understood your point and I tried to apply the ARDL with apply the Hausman test to see what I must apply the PMG or MG but the state softwate shows as following: -



chi2(3) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= -4.30 chi2<0 ==> model fitted on these
data fails to meet the asymptotic
assumptions of the Hausman test;
see suest for a generalized test

and from that I know the cross-sectional dependence is the reason for that error to show then I tried to apply other methods. I'm working on a paper for my PhD thesis and I think to apply different methods to see the difference in results with CS and without CS, that is why I'm applying different methods. So, what is your suggestion with this case or what methods I must apply with the problem of the Hausman test and Small N and T? I cant add more T because there is no available data more than I have for N (countries). Thank you for the help.

[JanDitzen]

Hi Mohammad,
I would not rely too much on the value of the CD test because of the small N. It will be severely distorted.

I think the problem is that the hausman test cannot calculate the difference in the coefficients and variance matrices because of different naming conventions of the coefficients. Can you try using the option lr_options(xtpmgnames)?

Given the size of your cross-sectional dimension I would run a pooled model without and with as test (!) cross-section averages. In your paper I would strongly recommend to discuss the issue of a small sample size. If you are really interested in the individual coefficients, I would suggest to do a separate time series regression for each cross-sectional unit.

[Ronak Parikh]

Hi,

I'm thinking of using a CS-ARDL model - when writing the command for CS-ARDL, I was confused by one thing: how do you find the value for cr_lags?

[JanDitzen]

cr_lags() is the number of lags of the cross-section averages and should be around T^(1/3), where T is the number of observations over time.

[ahmed mohamed ak]

What are the conditions and steps for applying the CS-ARDL model?