bmuthen posted on Friday, October 12, 2001 - 6:29 pm
Sometimes, regressing the latent class variable c on covariates x's produces a singular information matrix so that a full solution is not obtained. This is typically due to small class sizes in combination with the use of binary (or polytomous) x variables - if in a given class there is no or only one individual in an x category, slope parameters relating c to x are not identifiable. This typically shows up as a large positive or negative slope for a given class on the x variable in question (this can be seen in the provisional estimates printed in the failed run). For instance, with an x scored 0/1, a large negative value suggests that there is no (or perhaps only one) individual with x=1 in this class. The solution to this is to simply fix the coefficient at the large value (such as a logit of +-10). For example, -10 implies that changing from x=0 to x=1 status makes the probability zero of being in the class. Fixing the value will not alter the log likelihood value that was obtained in the failed run (shown at the end of Tech8) and gives estimates and s.e.'s for the remaining parameters.