On Wed, February 8, 2006 19:22, john w wrote: > > Gretl gives output when testing for cointegration and this is perfect. > My question is: to determine the optimal number of lags (which remooves > autocorrelation) what are your guidelines? > On what basis do you say that "this will be my optimal number of lags" in > cointegration test?
In my view, data analysis is more akin to a craft than to science, so "recipes" are something I generally dislike. However, a general strategy for estimating VECMs that I would personally consider sensible could follow these steps: 1) A cointegrated VAR (ie a VECM) is nothing but a VAR with nonlinear restrictions; therefore, an unrestricted VAR is a perfectly valid model. The only thing you miss is using the additional information coming from the restrictions on the rank of A(1), ie cointegration rank and long-run equilibrium conditions. Clearly, these are often the things you're most interestd in, but still the general VAR setup (lags, deterministic components) can be decided on the basis of the unrestricted VAR result and a pinch of economic reasoning. 2) Use Johansen's test for establishing the cointegrating rank. Again, don't be a martyr of the p-value religion. If in doubt, aim for a sensible economic interpretation whenever possible. 3) Estimate the VECM and do all the usual checks. If necessary, do some testing on the cointegration vectors too (ie zero restrictions). Unfortunately, this is an area where gretl is still a bit behind specilised packages: it shouldn't be difficult do add Wald tests on the betas, but for Johansen-like LR general tests you'll have to wait a bit more. No doubts someone will disagree with the above, with very good reasons. I make no pretence of "holding the truth", so I am open to criticism. -- Riccardo "Jack" Lucchetti Dipartimento di Economia Facoltà di Economia "G. Fuà" Ancona
