1
$\begingroup$

I have a question related to Granger Causality testing.

Is it okay to use a lag-length of lag=1 in my Granger-test? The optimum lag length selection in my R VARselect(data,lag=maxlag,type=trend) model says that lag=1 shows the best and most stable information criteria values according to AIC, BIC and FPE.

I have a 30-year set of quarterly data and I'm using a maxlag of 4.

$\endgroup$
2
$\begingroup$

If your data is stationary, then yes, based on the information you provided it seems OK to use $\text{lag}=1$.

If your data is nonstationary, you should follow Toda-Yamamoto procedure described very explicitly and clearly in Dave Giles' blog post. There are certain important points to pay attention to with respect to lag order selection under cointegrated data (see especially basic steps 5. and 8. of the Toda-Yamamoto procedure in the above source).

$\endgroup$
  • $\begingroup$ I see. And "lag=1" essentially also means that the effect is delayed by "1" time unit, at least when I find p<0.05? $\endgroup$ – user100531 Jan 13 '16 at 1:19
  • $\begingroup$ The answer is a little tricky. If the contemporaneous correlation matrix of model errors is diagonal, then yes. If not, you may have an alternative representation of the model, a structural VAR (SVAR), where there are also contemporaneous effects. $\endgroup$ – Richard Hardy Jan 13 '16 at 6:29

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.