"Mike " <michael.gerstREMOVE@yale.edu> wrote in message <email@example.com>... > Bruno and Eric, thank you for you replies. So if my > variables and coefficients are badly scaled, then can this > cause the Hessian to be singular or close to singular > (because of rounding errors)? Is this then what the warning > messages could be referring to? Thanks a bunch for your help.
Yes. Usually the Hessian is computed respect to the euclidian norm of the parameters you want to estimate, all of them mixed together. If the parameters is "randomly" scaled, the Hessian is also randomly calculated, and I'm not surprised part of the ill-conditioning is due to a bad scaling.