c**********e 发帖数: 2007 | 1 Should the stationary test be applied to the full error term or the
corrected error term?
The AR(p) model has two predictions. One is only the structured term, i.e.
XB.
The other one is Yule-Walker estimate, which includes the error corrections.
Correspondingly, there are full error, and corrected error.
Should we run stationary test on which error term? | t****g 发帖数: 120 | 2 I guess you are talking about AR Error Model:
y_t = X_t B + u_t
u_t = rho u_{t-1} + e_t
You question is to apply unit root tests on u_t or e_t.
First of all, if you have X, you may talk about cointegration test, instead
of unit root test. For example, ADF test becomes Engle and Granger
cointegration test, and KPSS test becomes Shin cointegration test, if there
is X; the statistics may follow different distribution. Note also that X
must be I(1). If X is I(0), that is, X is stationary, I think you'd better
directly test y, because X has no effect on whether y is non-stationary, but
may have effect on the distribution of the statistics.
Second, to my knowledge, a specific kind of unit root test or cointegration
test can only apply on y or (y X) directly with its given regression form;
that is, you cannot apply the test on neither u nor e! Check the original
paper on the test you'd like to apply to see if there is such a limit. |
|