The RATS Software Forum

This procedure does various types of the unit root regressions with threshold breaks on the residuals from an Engle-Granger cointegrating regression (assumed to be done already---input are the residuals).


The regression run is

du = (rho1 D1 x u{1} + rho2 D2 x u{1}) + lags of du

where D1 is the dummy for threshold series < tau and D2 is 1-D1. You can either input the threshold value or request a search. If you provide a value of TAU, the regressions are done with that fixed value for tau. If you don't, the values of the threshold series (excluding the PI fraction at each end) are searched for the one that minimizes the sum of squared residuals

Enders and Siklos(2001), "Cointegration and Threshold Adjustment," JBES, vol. 19, no. 2, 166-76.

The example from the paper is at http://www.estima.com/forum/viewtopic.php?f=8&t=1110.

Hi guys, I have checked the code and this code embodies wrong commands. It gives implausible results. I suggest to use the old version @endersgranger code. I have checked it with other software and it works better.

wolly77 wrote:Hi guys, I have checked the code and this code embodies wrong commands. It gives implausible results. I suggest to use the old version @endersgranger code. I have checked it with other software and it works better.

@endersgranger is for unit root testing and @enderssiklos is for cointegration. The latter has the "attractor" estimated using the preliminary Engle-Granger regression, so the two aren't designed to be comparable. If you have an example where you think this procedure is wrong, please post it so we can check.

Hi, I send you the dataset and the code. In the file txt there are two codes: 1) enderssiklos.src; 2) another code of engle-granger 1998.
enderssiklos gives implausible results because TAR-C Model estimation with "above" and "below" coefficients is wrong. I have checked it with other software. The code 2) is better.

You said that enders-granger 1998 code cannot be used. But in the 1998 there is an example of asymmetric cointegration test. Using the residuals of long-run relation it is possible to calculate the threshold level such as Enders and Granger did. The results have to be equal because the methodology used is the same in the threshold value calculation: Chan's method 1993.

Thank you.

Hi Tom, I explain the problem I encountered in more details. I am using the Enders-Granger (1998) and Enders-Siklos (2001) RATS codes. When applied to the residuals of long-run relation, I suppose that these two codes have to give (more or less) the same results in terms of threshold and coefficients above/below threshold in the asymmetric Dickey-Fuller (FD) test. This is because both procedures use Chan's method to detect the optimal threshold. In Enders-Granger (EG) 1998 paper, the authors apply the procedure to run a cointegrating test applied to yiled curve long-run relation. For this reason EG procedure can also be used to run a cointegrating test, even if Enders-Siklos (ES) 2001 is appropriately studied for cointegration.
The problem is that I find (in some cases) very different results in terms of threshold values detected. According to Chan's procedure the best threshold has the the smallest SSR. This is always true with EG code but not in the most recent ES. In addition, in some cases the coefficients above/below the threshold calculated of the asymmetric DF test are wrong in ES. I have detected these points replicating the results with other software. If you run my database you will discover these discrepancies and incongruences. I suppose there is an error in the ES code. Thank you Tom.

As I said earlier, the two aren't estimating the same model. In your program, you are using the commented out versions, while Enders-Siklos uses the others (basically forcing the mean through zero).

* set z_plus = flag*yy{1}
* set z_minus = minus*yy{1}
set z_plus = flag*u{1}
set z_minus = minus*u{1}

Thank you Tom. I have seen many papers using Enders-Granger 1998 procedure to run asymmetric cointegrating test. The same procedure is used in Enders book at the page 481. Enders uses the long-run relation residuals to calcualte the threshold using Enders-Granger 1998 approach. Enders-Siklos is different but Enders-Granger is not wrong and can also be used. Could you be so kind to confirm this view? Thank you again.

wolly77 wrote:Thank you Tom. I have seen many papers using Enders-Granger 1998 procedure to run asymmetric cointegrating test. The same procedure is used in Enders book at the page 481. Enders uses the long-run relation residuals to calcualte the threshold using Enders-Granger 1998 approach. Enders-Siklos is different but Enders-Granger is not wrong and can also be used. Could you be so kind to confirm this view? Thank you again.

That question is probably best addressed to Prof. Enders. However, I would point out that your series are clearly by examination not cointegrated, or even close to being cointegrated, so any of these models are misspecified.