Dear all,
I am currently attempting to run a latent profile analysis on a set of data.
Unfortunately, I only have 270 participants with 230 having complete data for indicator variables and the other 40 having partial data only. I have been struggling to understand if this may be sufficient for the LPA to be robust. Some papers have recommended needing at least 300 participants but others seem to say for sample sizes more is better but it depends on the number of indicator variables. I have ended up only using three indicator variables. Would anyone have any guidance on if this is sufficient?
Secondly, I have utilised the recommendations in this thread: https://www.statalist.org/forums/for...-profile-model to calculate BIC using different variances of the model, namely lcinvariant(none) and covstructure(e._OEn, unstructured). However, the models converged fine from 2 to 6 classes in the simple model, with 3 classes model having the lowest BIC. The models with lcinvariant(none) only converged up to 3 classes and the covstructure only converged on the 2 class model. In this instance, from a statistical threshold standpoint, should I just compare the BIC across the different models and see which one is the lowest? Alternatively, should I go with the 2 class model as it allowed for different error variances across classes and allowed for the indicators to be correlated?
Thank you for your help in advance!
I am currently attempting to run a latent profile analysis on a set of data.
Unfortunately, I only have 270 participants with 230 having complete data for indicator variables and the other 40 having partial data only. I have been struggling to understand if this may be sufficient for the LPA to be robust. Some papers have recommended needing at least 300 participants but others seem to say for sample sizes more is better but it depends on the number of indicator variables. I have ended up only using three indicator variables. Would anyone have any guidance on if this is sufficient?
Secondly, I have utilised the recommendations in this thread: https://www.statalist.org/forums/for...-profile-model to calculate BIC using different variances of the model, namely lcinvariant(none) and covstructure(e._OEn, unstructured). However, the models converged fine from 2 to 6 classes in the simple model, with 3 classes model having the lowest BIC. The models with lcinvariant(none) only converged up to 3 classes and the covstructure only converged on the 2 class model. In this instance, from a statistical threshold standpoint, should I just compare the BIC across the different models and see which one is the lowest? Alternatively, should I go with the 2 class model as it allowed for different error variances across classes and allowed for the indicators to be correlated?
Thank you for your help in advance!
