Hello, I am working on a latent profile analysis project via R package tidyLPA. The problem is that I found it is difficult to determine the number of profiles. For 1 to 10 profiles, the BIC continuously decreased when the number of profiles increased. And both BLRT and LMR calculated p values were significant (p
Article Statistical Power to Detect the Correct Number of Classes in...
Article Deciding On the Number of Classes In Latent Class Analysis a...
http://www.statmodel.com/bmuthen/articles/Article_111.pdf
I will say it upfront: I only have experience with Latent Class Analysis, but the problem of choosing the number of classes there is similar.
Based on the information you provided I would go for three classes - which is where the inflection point in the plot is located.
I would suggest running some clustering algorithms or dimensionality reduction techniques and then see if three profiles is the optimal number. If the data is very heterogeneous it might be a good idea to reduce it (for example with PCA or even maybe kernel PCA) and then run the analysis. You will have some more information to back up your decision.
But you mentioned "clinical features" so I assume it is a "real-world" problem. I think here more important than the BIC values will be the ability to interpret the profiles and use them in clinical practice.
I think you can also rely on the existing literature/theories to see how many profiles you can expect. Is it closer to 10 or closer to 3? If it is closer to 10 then maybe a bigger sample size would be needed?
I hope this helps in some way.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R Statistical Package