Model robustness

From: Martin Fransson Date: August 27, 2007 technical Source: mail-archive.com
Dear all, I have a question on model robustness; if different sets of initial values of thetas change the standard errors a lot, should the model be considered to be too sensitive? I obtain the same estimates for both thetas, omegas and epsilons regardless of initial values, but the standard errors seems to increase if one starts with initial values too far away from the minimum. (Too close seems not to be ideal either.) Are there any standard routines that one could apply for this kind of behavior? Could one, for instance, use the mean of the sets of standard errors from different runs? Or should the model be considered to be unstable as long as different initial values produce different standard errors? Is there an acceptable limit; such as when an increase by 25% on all initial values for theta implies a 20% increase in standard errors but all estimates remain the same? All comments are welcome. I have tried to find a similar topic in the archive but without success. Best regards, Martin =============================== Martin Fransson Dept. of Computer and Information Science Linköping University 581 83 LINKÖPING, SWEDEN <mailto:[EMAIL PROTECTED]> [EMAIL PROTECTED] +46 13 281467 ===============================