yes, it is okay, just give value 1 for the unusual values and 0 otherwise. this will be a series on the independent side. if the dummy variable is significant then it implies there is structural break otherwise NO.

Thanks for your response! Kripfganz, Sebastian, who developed the ARDL test for Stata says otherwise however. This he wrote in an email to me: "With such structural breaks, the notion of a single long-run relationship becomes questionable as it might also depend on the break. Also, the Pesaran et al. (2001) bounds testing procedure for the existence of a long-run relationship becomes invalid in the presence of structural breaks. The ardl command is just not designed for such cases." Would you argue against this? Can you direct me to a paper?

Kripfganz is correct and all that he is saying is that one should take that into account. Without taking such instability into account, the obtained long run relationship is invalid

One odd thing I have discovered is that the ARDL command in Eviews includes the dummies into the long-run equation, and they get somewhat different coefficients compared to what they get through the ARDL command in Stata, where they enter the short-run. The coefficients on the other variables are about the same however, as are diagnostics. Did the Eviews or the Stata programmers make a mistake, or are both ways valid?

Please share your results of both soft wares

......

Using Dummy variable as independent var as like as other variables may be incorrect. An ARDL model needs some lags for variables in short-run. Therefore, dummy's lags will be taken into account and the risk of multi-colinearity increase. I think you could exclude dummy variable from lag mechanism.

http://link.springer.com/article/10.1007/s11069-015-2050-x

http://link.springer.com/article/10.1007/s11069-015-2050-x

According to Sebastian Kripfganz lags of dummies should not be included. Also, David Giles does not include lags of dummies: http://davegiles.blogspot.se/2015/01/ardl-modelling-in-eviews-9.html

Btw, for David Giles it seems perfectly legit to include dummies (such as breaks) in the long-run equation...

Lars is correct! Put dum at lag zero.

You can include ndum inside the coint space or exclude it from the coint space. But both dum must be a lag zero.Then for the first assuming you have y x v and dum & c s1 s2 s3 then set you lag length 4 because yi 

for the second, you have y x z & c s1 s2 s3  ---use lag three

ARDL equation is a dynamic specification and there is no problem if u use dummy on the right hand side. I used it in some of my projects to capture structural break. Gour

If you have z (dependent) x and y variables, ARDL model with one dummy could be formulated as follow (in attachment)

That is what I have seen elsewhere: the dummies goes into the short-run equation. But in Eviews this doesn't seem to be the case? As a matter of fact, the dummies goes in as first differences (impulse dummes) in the short-run and as step dummies in the long-run. Does that make sense?

Your question is very sophisticated and very specialist.

I checked Eviews, dummy variables could be set by "list of fixed regressors" field. Therefore, due to the results the length selection was done by excluding fixed variables.

At the modeling view, the breaks that were found by break point tests, suggest many short-run and unexpected breaks. The main model is devoid of break variable and the tests found an abnormal break through the model that is not permanent.

This is what the output looks like in EViews.

There is two method for cointegrat,on. The first method that known as ECM, claim two different mechanism. Long-run mechanism that I sent as first forumla and short-run mechanism with a deflater for long-run residuals (eq no 2).  The long-run cointegration can be determined by t-statistics by Banerjee et al. (1998) .

The second method is Bound test approach (eq 3). it needs Pesaran (2001) bound values.

Eviews estimates both methods and your outputs belong to ECM method, then all shorrt-run variables (including dummy) should be differentiated. 

It seems we can agree that (a) Stata allows dummies in the short-run whereas, for Eviews 9, dummies enter the long-run as well as the short-run, (b) the current programs as written, may lead to different results when dummies are forced in and syntax errors often occur. We have not discussed whether they allow a know shift-point tests of stability to be conducted such as Chow test. I see many uses CUSUM and CUSUMSQ which are not powerful tests. They are general tests and focus on gradual shifts. Bear in mind, that inclusion of just a dummy at the ith point means the exclusion of the data at the ith point. It does not tell us about slope shifts.  

I(1) variables will always dominate I(0)) so the dummy is  I(0) but plays the role of helping to minimize the squared errors. Nevertheless, dummy's lags can create collinearity problems. In practice, the error-correction term gets biased downwards (i.e., smaller in absolute term) in the face of dummy variables. Then, for  interpretation purpose, the speed of adjustment is lessened. This is a shortcoming for the t-value of ECT(-1) . PSS(2001)  F-test is affected by the number of first differenced terms included. The t-value on the ECT(-1) is much stronger given that dummies do not affect it significantly. 

@@

Hüseyin Karamelikli

Your formulas are okay but the reader cannot use it as it because you have not shown how to get Theta and Eta. Since you opened that door, completion is necessary! Maybe explaining how the ECM term is obtained under this format may be helpful to the readers.

@@Chuck A Arize 

You are right. The expanded version of file is attached here. I hope that every points of  modeling of ARDL and its bases has been described.

What is your advice to me on EViews ARDL cointegration? What tests should I employ and report? Can you guide me through it, step-by-step?

@@Lars Ahnland Nordfors

I thinks ECM based ARDL in Eviews (COEFFICIENT DIAGNOSTICS; BOUNDS TEST) has some bugs. So, I recommend you follow following link.  I am going to report Eviews bug to IHS.

http://davegiles.blogspot.com.tr/2013/06/ardl-models-part-ii-bounds-tests.html

Was it my discovery of treatment of dummy variables that made you think it's bugged, or that post? Eviews have been updated since that was posted.

I have my own addin to estimate ARDL, after your post, I compared both outputs and I saw some differences between them.  Thanks you for motivation to compare. 

Can you explain more please? What is the difference?

Nice discussion above. Is it  mandatory to use dummies for each structural break, I have an equation in which almost all variables have structural breaks, that could cause degrees of freedom problem. Also there are cases where there is an obvious structural break but the Dummy for it turns out non-significant.

Respected Muhammad Azam!

I have applied ARDL approach without dummy variables(for structural breaks), now I am employing Time Varying Parametric Approach(Kalman Filter) instead of Dummy vaiable approach. How can I justify it? What are the disadvantages of dummy variable approach? or What are the limitations of F test in dummy variable approach? Please give references if possible.

Regards

How can I use the ADL(3,3) using Chow's method and QLR method of structural break in Eviews 8 and R with dummy and interacted dummy of lags of the regressors?

If you are using Eviews 9 you will have an option to include list of fixed regressors which will not be lagged by the software.You need to list your dummies there.

Interesting debate. Majority of the issues on ARDL modelling are extensively discussed in Eviews guide especially version 10.

Whatever you will read in E-view is already discussed here. One is tru interested would read read all Thank you!

Yes, you can but only as an exogenous variable (just like you treat seasonal dummies).

Tha nks best

Thanks Chuck A Arize for suggestions and Lars Ahnland for higlights this question.

I am using eviews 10 and treat dummies as a fix regressor (as suggested by eviews manual). Compared with the previous version, I notice that, in eviews 10, the result shows that the any dummies (e.g structural break, crisis dummies) only enter in the short run equation, not in the long run.

Unless, i include dummy variable in the dynamic regressor, so it will appear in the LR equation. In what case should we treat it as dynamic regressor?

Thank you Dr. Ghazali

kindly, could anyone help me using a dummy variable "with no lags , as suggested" in ARDL model using Eviews?

infinite thanks

Bahy Yassin I insert dummy in the 'fix regressor' column as suggested in the manual eviews 10, thats what i did..

siti Ghazali , thanks alot

This article may provide you with the solution you require.

Emeka Nkoro & Aham Kelvin Uko, 2016. "Autoregressive Distributed Lag (ARDL) cointegration technique: application and interpretation," Journal of Statistical and Econometric Methods, SCIENPRESS Ltd, vol. 5(4), pages 1-3.

Osman Damba

https://nomanarshed.wordpress.com/2014/11/16/a-manual-for-ardl-approach-to-cointegration/comment-page-3/

Thank you so much for your kind help 🙏🙏