The speed of adjustment is negative and less than 1. This is one checking.

Check for the following tests, normality, serial correlation, heteroskedasticity and functional tests, including CUSUM and CUSUMQ stability tests.

I hope this is helpful.

Ahamed Lebbe Mohamed Aslam Agreed with the above two answers. i can suggest sign biased test and ARCH test also. But all these tests are applied to check model fit. By robustness, what you really want to convey, is bit lacking sir. More general suitability of ECM is that it should be negative and less than 1.

The CointEq (ECM) must be negative and less-than one (1) but you must check if the probability value is less-than the chosen level of significance of either 1%, 5% or 10%. If the CointEq (ECM) conforms to the stated conditions, then it is robust.

All Steps of ARDL Model:

• Check Normality by using jarque-bera test.
• Check Serial Correlation by using Breusch-Godfrey serial correlation LM test.
• Check Heteroskedasticity by using Breusch-Pagan-Godfrey test. In the last stage Check the model stability by using CUSUM & CUSUMSQ techniques.

I hope these steps are helpful.

ROBUSTNESS: (i) reliability, and (ii) validity. Reliability is the measure is consistency. The proposed model must be able to produce consistent result in repeated measurement. Validity is the measure f precision; the proposed model must be accurate within an acceptable level of precision, i.e. between forecast and actual outcome.

REPEATED MEASUREMENTS: With large sample size, you need only minimum

sample size to construct the model. The remaining sample may be used for repeated measurements to check for reliability and validity.

SIMULATION: Outside of the sample at hand, you may want to construct a simulation data to test the proposed model whether it passes (i) and (ii). A proposed model that fails reliability or validity test or both cannot be

considered robust.

LINKS:

https://eml.berkeley.edu/~mcfadden/e240b_f01/ch7.pdf

http://www.lse.ac.uk/website-archive/GeographyAndEnvironment/neumayer/pdf/Robustness%20Tests%20and%20Statistical%20Inference%20.pdf

Dear Ahamed, I want to assume that you already know the relevant test statistics required for a robustness check as have been outlined by others researchers above. It also appears you're using eviews version 10, which has the ability to perform ARDL bound cointegration test, as well as summarizes the longrun and shortrun (ECM) estimates of the model. The ECM results of the ARDL model usually comes in its parsimonious form based on automatic adjustment by the chosen information criterion. The relevant tests for robustness check are as follows, Breusch-Godfrey LM test, specification error test using Ramsey reset, cusum & cusumq tests, etc. If you're familiar with eviews 10, you can do these tests by clicking on view on the result interface, but for the ECM result, you may have to copy the variables shown in the parsimonious ECM and click on quick, then click on make equation. This would allow you to reproduce the parsimonious ARDL-based ECM result from where you can now do the robustness check particularly for the ECM. To check for robustness, click on view, then residual tests, or stability tests, etc. I hope this helps? Thanks

Agreed with Aqib Mujtaba.

Agreed with Aqib Mujtaba but as per my knowledge in ARDL unit root test is not mandatory and second the error correction coeff should not be lesser then (-1).

I agree with most of answers except for the one relating to the Error Correction Term Coefficient. It should be negative (to ensure convergence), however, it's absolute value can be between 0 and 2. If the absolute value is higher than 1 and less than 2 this means that the speed of adjustment is oscillatory. In other words, full adjustment to equilibrium takes less than one year.

Thanks so much Israa, I got an ECM coefficient of -0.9 and I thought it was too high speed of adjustment.

What if cusum square has breaking point? Till there all diagnostics are well behaved…