andreas.rosenblad(a)ltv.se schrieb: I'm not opposed to user choices, but I think one has to be careful to avoid cluttering the GUI. Therefore in this case I am inclined to think that it's not worth an explicit option. Apart from that I agree that some sort of small sample correction is desirable. I could look up what PcGive does in the documentation tomorrow if you like. /sven

Allin Cottrell schrieb: I could look up what PcGive does in the Ok, so it seems that PcGive 10 uses the root of T-c in the denominator, where c is the "average number of estimated parameters per equation, rounded towards zero". Let's see if we can verify that for our specific case: Comparing standard errors between gretl and PcGive for the same beta coefficient gives me a value of: c=9 And we have: 4*4=16 short-run coefficients (lagged differences), 4 alpha coefficients, 5 beta coefficients, of which 1 is just a normalization, so actually 4, and 3*4=12 coefficients for seasonal dummies. ------ 36 / 4 = 9 coefficients per equations. Looks good! Also looks good, but since the date there is August, I guess the problems with initial values for restricted estimation in your earlier email are still there, right? Browsing the PcGive documentation, I also saw that they do not impose the normalization restrictions in the maximization stage. They actually remove them before estimation, then estimate, and in the end reimpose the normalization, saying that it's more robust. -sven

Allin Cottrell schrieb: To add to that, in terms of the formulation vec(\beta) = H * \phi + h vec(\alpha') = G * \theta only if: (1) alpha has only exclusion restrictions (I'd say that would be only zero rows or distinct unit vector rows in G) (2) a non-zero entry in h gives a fixed beta element (i.e. a corresponding zero row in H, meaning no further contribution from phi; which they call scale-homogeneous) do they actually do the scale removal. So I guess a first shot at the algorithm would be: 1. find the smallest (or largest?) fixed element in a cointegration relation (beta column) 2. replace this non-zero entry in h by 0 3. insert a new unit vector row/column "cross" in H, which replaces the respective zero row but increases the column count by 1 (if you know what I mean here) 4. do the switching 5. divide each beta column by the estimate in the position of the to-be-normalized element times the desired normalization value 6. adjust the covariance matrix accordingly (details are left as an exercise to the reader ;-) If we're lucky this would make the choice of starting values a little less important, right? -sven