Hi. I am doing discrete LCA with freq. wt for 3 binary indicator variables specified. The freq wts are quite large, in the thousands. For no of classes >2, I am getting either the error: ' standard errors may not be trusted since got non-positive definite 1st derivative matrix, prob with param #'with a cond #, or ' avoided singular info matrix by setting params #, #. Are you using Newton_Raphson to maximize the total likelihood? Can I trust the value of params & their p-values if condition # is above a certain threshold? How do I treat the params that were fixed? The first error is more serious isn't it? Thanks for your help.
I am getting these warnings & errors mentioned above for 3 class models. I want to pick the most stable solution, taking entropy into consideration. How low does the condition # get before you consider it ill-conditioned? Can I trust the values of the parameters that have been fixed? Thanks.