|
|
 |
| When entropy is low in LPA |
 
|
| Message/Author |
|
|
| Ppudah Ki posted on Wednesday, October 21, 2020 - 2:15 am
|
|
|
Hello,
I am just wondering whether it is okay to keep the LPA model with .616 entropy (or what to do when the entropy is low).
I have analyzed an LPA model, and the 3-profile solution was the most desirable model, considering BIC, LRT, BLRT, class proportions, and theoretical background. 3-profile model was the best model to conclude, except that it has low entropy, which is .616.
(Actually, the entropy scores of all the models (2-profile model, 3-profile model, 4-profile model) were quite low (.641, .616, .626 respectively).)
1) Thus I am just wondering if I can choose the 3-profile model with .616 entropy and use it in my paper.
2) If so, could you recommend any reference to back up the usage of low entropy?
Thank you! |
|
|
|
|
I think using a model with that entropy is fine. It just says that the classes are not as well separated as with say entropy = 0.8. An analogy is SEM with an R-square that isn't very high but the model fits well - this is the truth that we have to accept; this is how good the model can be given the data/the variables.
I know of no reference for this. |
|
|
| Jon Heron posted on Thursday, October 22, 2020 - 9:59 am
|
|
|
It's dead easy to demonstrate with simulation that an LPA with awful entropy can give you an unbiased estimate of a covariate effect.
I've shown in a lecture that with two classes underlying a single variate Y you can get entropy below 0.2 if the mean difference in Y between classes is comparable to the within class variance of Y.
No bias, just lots of uncertainty. Just don't go using modal class assignment |
|
|
| Back to top |
|
|
