In gretl 1.8.0 we made this change: "The matrices returned by the accessors $sigma and $vcv for VAR systems now have a degrees of freedom correction." However, the $sigma accessor for a VECM gives the ML estimator (although the reported standard errors use a df adjustment). I feel that our knickers are in a bit of a twist here, and I'm not sure what the best solution is. In the context of VECM output we print the cross-equation convariance matrix (accessed by $sigma) and report its log-determinant. Now, the log-determinant as it enters the likelihood calculation is obviously based on the ML estimator, so it seemed to me problematic to print a different matrix, and equally problematic to provide, via $sigma, a matrix different from the one printed. This becomes a live issue with the new $vcv accessor for VECMs. For consistency with the reported standard errors it should be df-adjusted, but for consistency with $sigma it should not be adjusted. Urgh! (Right now in CVS $vcv is df-adjusted.) I think that my (mild) preference would be to use ML values throughout for VECMs, but as I recall we changed from that for compatibility with, e.g., PcGive. Any further thoughts on this? One further point: I've now enabled the $df accessor for VARs and VECMs, so it's pretty easy for a user to multiply by $T/$df if needed -- provided, of course, that we're clear on which variant is provided by $sigma and $vcv. Allin.