|
|
 |
| LCA with ordered categorical variables |
 
|
| Message/Author |
|
|
| ruthjlee posted on Wednesday, April 23, 2014 - 10:33 am
|
|
|
Hello,
I am new to LCAs and have several questions:
1. My variables are all ordered categorical, but with a mixture of response formats and scales. Do I need to make the response categories as consistent as possible across the variables by merging response categories?
2. Even though I plan to focus on BIC as an index of model fit, I assume that I should still pay attention to warnings that the best loglikelihood value was not replicated? - I assume that is giving me information that the model did not converge, overall, rendering indices other than loglikelihood unreliable?
3. Apologies for this one, as it may just be a function of the shallowness of my stats knowledge in general. For some models, I get the message: 'ONE OR MORE PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE INFORMATION MATRIX.' - as many as 12 parameters in some cases. I generated tech1, but am not sure what to do with the information that it gives me. I found some references on this forum to constraining the parameters and re-running the model, and have consulted the user manual, but am still unsure what kind of constraint to apply. Would this come from probabilities from theory, or from something that tech1 can tell me?
Many thanks in advance for any pointers.
Ruth |
|
|
|
|
1. No.
2. The best loglikelihood must be replicated for the results to be interpretable. You have hit a local solution if it is not replicated.
3. If thresholds are fixed, it is fine. This can help define the classes. Large negative is probability of zero. Large positive is probability of one. If other parameters are fixed and you want an opinion, send the output and your license number to support@statmodel.com. |
|
|
| Back to top |
|
|
