Symmetry and cusum test in NARDL model

[marias]

Dear All,

I am facing two issues with a NARDL Model:

1. When I run the symmetry test I get a near singular matrix error. I don't face any issue with the estimation of the model but only with the symmetry test.
2. The cusum and cusums tests give me results for a specific period of time and not for the total sample period. I suppose that it is due a specific variable which have zero values for several months. How can I overcome it?

Can you please advise how I could resolve these problems?

Thank you in advance

[EViews Mirza]

Hi Marias.

Mind sharing the workfile with the exact NARDL specification you're using for estimation?

[marias]

Hi Mirza,

Please find attached the workfile.

Thank you

[EViews Mirza]

Hi Marias.

I have checked in a fix for the "near singular matrix error". The issue arose due to the presence of 0-lag regressors. Now, should any 0-lag regressors arise, they will produce NA values in the statistics table.

As for your concern number 2., I'm not entirely sure what you expect the test to do? The formal test looks at the equivalence (either sum or lag-specific) of asymmetric positive and negative coefficients after estimation. Since estimation is done on a certain sample, the tests are based on the coefficients that arise from said estimation. The literature (as far as I'm aware) doesn't typically deal with time-varying equivalence tests, if I understand your concern correctly, and so EViews doesn't offer such a test.

Kind regards,
Mirza

[marias]

Hi Mirza,

Thanks for fixing the 1st issue. Concerning the 2nd issue I'm afraid I wasn't clear. I run the NARDL model for the whole sample period (01/01/1996-31/12/2023) and I include a dummy that takes the value 0 and 1 before and after a specific date, e.g. T1. The plot of the cusum test does not start from the beginning of the estimation sample but from T1. I would like the cusum and cusums test to give me results for the total estimation period. Do you have any solution to propose?

Thank you

Maria

[EViews Mirza]

I see now. You were referring to the stability cusum test.

The issue there is that your regressor x3 is a categorical variables with a small slice of non-zero data toward the end of the sample. If you look at the cusum documentation https://www.eviews.com/help/helpintro.h ... 23ww183726, you'll note there's a section which describes how recursive LS estimates are derived, the residuals of which are used in the derivation of the cusum test.

When you have a regressor such as your x3 variable, LS estimates are undefined for the first block of observations where x3 is entirely zero. This means the cusum test will be undefined for the same set of observations. When LS becomes defined after x3 starts to display non-zero values (which is toward the end of your sample), the cusum test becomes defined as well, and the result is what is displayed.

You really want to avoid regressors which have a long initial chain of zeros as there's little you can do to remedy the situation, because, well math.

Hope this clarifies things.

Mirza

[marias]

Thank you very much Mirza