Forum: help
Monitor Forum | Start New ThreadRE: stability of covariance matrix [ Reply ] By: Ott Toomet on 2016-07-25 01:36 | [forum:43392] |
Short answer: it may or it may not. VC matrix is coming from the second derivative of the log-likelihood function. So think--even if the location is (rather) stable, what may make the second derivative unstable? * the function is highly curved around the optimum * you have a large number of explanatory variables and the optimum is rather flat. The numerical errors when inverting the Hessian may add random noise. * the derivative is contaminated by numerical noise. For instance: ** you use numeric second derivative ** you have a huge number of observations and the rounding errors accumulate when you add those for the log-likelihood value. ** you use analytic gradients, but those use finite-precision approximations of algebraic functions. There are more ways you can get derivatives wrong ;-) Try Arne's suggestions, or let us know what exactly are you doing and what kind of problems do you have. Cheers, Ott |
RE: stability of covariance matrix [ Reply ] By: Arne Henningsen on 2016-07-17 08:39 | [forum:43387] |
Some suggestions: - change tolerance limits / stopping conditions - change the optimisation routine - provide analytical gradients (if you haven't done this) - provide analytical Hessian (if you haven't done this) - obtain the final Hessian by the BHHH method (if available for the selected optimisation routine) |
stability of covariance matrix [ Reply ] By: elebe nwezza on 2016-07-02 00:42 | [forum:43332] |
I noticed that the values of estimators were the same with different start values using maxLik function but the values of the variance covariance matrix of the estimators were changing. Is it possible to have stable variance covariance matrix of the estimators just as with the estimators? |