3
$\begingroup$

I am dealing with what I believe to be cointegrated statistics (Johansen says cointegration is present) however, when regressing the model the error terms are not stationary.

One of the causes of this is autocorrelation in the dependent regressor. So by imposing an AR structure on the model, I do get stationarity of the error terms, but is this cooking the equation to make it stationary?

Thanks!

$\endgroup$
  • $\begingroup$ We cannot say anything about "cooking the equation", since you provide no context at all. So no one knows if the "original" equation has any merit. Please provide context: say a few things about the model, write the equation before introducing the AR term, etc. $\endgroup$ – Alecos Papadopoulos Jul 28 '15 at 13:23
1
$\begingroup$

Johansen cointegration will usually yield more than one cointegrating vectors (cointegration regression coefficients). By definition, for a cointegrating vector, residuals must be stationary. And a stationary process can take any stationary ARMA form. So, adding AR term is cheating. What you shall do is to test for stationarity of residuals obtianed from cointegration vector. I do not mean to be rude, but you must be doing something wrong. Perhaps, your original variables are not I(1), or you need to include a determinsitc or stationary trend in your stationarity test, and the like. Just for comparison purposes, how do your Johansen results compare to Engel-Granger test. This might shed some light. Best,

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.