Dear all, I'm struggling selecting models in latent profile analysis. Luckily, tidyLPA provides a lot of helpful fit indices for making decisions. Among them, there are two indices for average class probability (prob_min and prob_max). As the literature reveals, cut-off for average class probability can be set to 0.8. I'm wondering if this is based on the lowest or the highest value (i.e. prob_min or prob_max)? Or do these indices in tidyLPA mean something completely different? Any advice is highly appreciated. Many thanks!!
The average should refer to the average across all classes that were assigned. You want it to be as high as possible. If the average is .99, but you have a minimal value that is .60, then you would be worried about your model. On other hand, if your minimal is .98, then it would just be an indicator that you have a good model. Average, minimal, and max are metrics that can help you evaluate your models. Hope this is helpful
Thanks a lot, Gerard! Indeed, I have some models with max. > .80 but min. < .80/.70. As far as I see, there is min. and max. average probability in tidyLPA but no overall average. But still, this is very helpful for making decisions!
I always thought that the entropy can be a sort of average probabilities. See page 5 in this:
https://digital.library.unt.edu/ark:/67531/metadc822835/m2/1/high_res_d/dissertation.pdf page
"High entropy values indicate more precision in group membership classification, and therefore a clearer distinction between latent classes (Celeux & Soromenho, 1996; Kaplan & Keller, 2011). For instance, an entropy value of .85 is interpreted as 85% of the individual cases being classified accurately in a latent class."
It is, kind of. I found this in Collins & Lanza (2010): "It is useful to remember that E[ntropy] is essentially a weighted average of individuals' posterior probabilities." (p. 75).