Dear all,
I am trying to find the least biased approach to connect results from latent-class analysis with covariates. However, I do not focus on class assignments... Instead, I would like to use posterior latent-class membership probabilities for class#1 as a dependent variable -> "Regression of an individual’s logit-transformed posterior probability to be in a given class on the covariates" as defined by Clark and Muthén in their article "Relating Latent Class Analysis Results to Variables not Included in the Analysis".
For example, I run the LCA without covariates in Mplus, save posterior probabilities and used individuals' posterior class membership probabilities for class #1 as a dependent variable (log-transformed, e.g. in Stata). However, in this case, I faced bias because standard errors for individuals' posterior class membership probabilities are not considered in the analysis... Generally, a bias-corrected three-step approach is recommended, but is it possible to implement it when I use probabilities as DV?
I will be grateful if you can share recommendations/articles that focus on the issues on how it is better to run regressions with posterior probabilities for one class LCA as a dependent variable? Is there any way to use the three-step approach in this case? Thank you for the advice!
Rosner Fundamentals of Biostatistics which is available in the z-library covers types of regression methods that may be helpful to you. I recommend that you you look there. Best wishes, David Booth
Dear David Boyda
Thank you very much for your answer and invaluable recommendations. If I understand correctly, R3step implies a bias-corrected multinominal model where DV = nominal scale of classes. Many simulation studies confirmed the relatively good performance of this approach. However, it is not possible to use if I need to use probabilities to be in class #n as DV in the linear regression model.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression