I am running an analysis for a dual response CBC data with only 2 choices per set and one prohibition among the 31 levels included (5 in 5 dimensions and 6 in 1 dimension). I included an interaction between the dimensions that contained these 2 levels as proven to be significant by the interaction search. (Can this be the source of the problem??)
The estimation of the 2 prohibited levels is somewhat problematic.
In the lighthouse program:
1- one level had an unrealistic estimation with a negative sign (expected to be positive and better than the one next to it).
2- both of them did not report any standard errors nor t-ration
In the stand-alone Latent Class analysis (1 group) : (results are disutilities as the signs are flipped)
1- disutility signs are all ok for one level but not for the other (opposite situation from the above case)
2- Standard Error is reported 0 for all 31 levels
3- t-ratios are a huge number (6 or more digit number)
I get a warning in both cases that says:
"The data is ill-conditioned. A ridge adjustment was made."
How can I solve all these issues? Thanks for your help.