I'm trying the effect of different tolerance criteria and derivstep values to obtain global minima (least squares). Generally, lower tolerance criteria give at least as good a fit as expected. Very occasionally setting to a lower tolerance gives a slightly worse fit. In these cases a lower nlsf.derivstep value helps.
My question is what range of values is reasonable and what exactly are the implications of lowering the nlsf.derivstep value? The help gives 'A smaller value gives higher precision but may cause numeric round-off errors. The default value is 0.005.' Where does these round-off errors occur? What does the higher precision refer to?
Thanks for reply. I'm still trying to assess the practical limits but I think I go along with the default value of 0.005 - just occasionally a lower value helps but I think the more important feature is the tolerance criteria which typically I'll set to 0.001 for dose response type studies.
nlsf derivstep and tolerance parameters
Posted - 10/24/2003 : 10:37:30 AM
