Hi all I am working on creating a general framework to work with mxiture models, specifically with Latent Class Analysis (LCA) and its extensions. Now, I am working on a LCA model, using the bernoulli_logit_lpmf functions from Stan. This model so far works well when I estimate 2 clases, but when i…

[image] spinkney: simplex[C] pi[I]; Unfortunately that is another identifying assumption that substantially changes the meaning of the model. It assumes that the class probabilities for each item sum to 1. For example, if Class 1 has a 90% chance of answering item 1 correctly, Class 2 has a 1…

I get unimodal distributions when I drop the model-estimated prior for pi in favor of a preset one. So that remains my best guess why your specific example ends up multi-modal. I doubt that it fully solves the problem of identifying the model, though. (Note that the x-axis should actually read pi, i…

@spinkney sadly the ordering of pi doesnt work theoretically as it changes the structure of the model. Thanks @simonbrauer , I tried with the change in priors, and result in some runs working fine and other dont. Still sensitive to starting values between runs

Coming back at this, saw this other post in the forum Latent class previous post where thet adjust the DINA model code And after testing found that this method to estimate latent classes from binary data works well for increasing numbers of classes. With a small change in the constraints between C …

Interesting! I’m surprised it works better than other methods too. Interested in hearing what you learn as you extend it. [image] Mauricio_Garnier-Villarre: pi[i,c] = master_pi[i]^(c - 1) * nonmaster_pi[i]^(1 - (c - 1)); I would be leery of this line for C > 2. It assumes an odd structure fo…

Removing that restriction and coding the model seems to recover the poLCA parameters and with no issues detected. But it’s so much slower (takes about 14 mins on my laptop). data { int<lower=1> I; // number of items int<lower=1> J; …

This simplifies the model and only tries to find parameters for the item by class interactions. It takes four min 2 min on my machine. Though you still may not like the ordered vector restriction. data { int<lower=1> I; // number of items int<lower=1> J; …

@simonbrauer thanks for pointing out the issue with that part of the code. @spinkney these are very intereting changes to the code for the model. I will play with them, thank you very much

@Mauricio_Garnier-Villarre Did you happen to find a solution for polytomous variables?

@Kevin_Linares no, I havent been able to extend this for polytomous items. Also, to be fair, for the sake of the project I am working on I drop this so I havent worked on it for a bit If you want to try this, I would susgget to take this code, that nis written for the log probability of answering …

Identification issues with Latent Class Analysis/Mixture model

Modeling

Mauricio_Garnier-Villarre August 15, 2022, 10:00am 21

@spinkney I would like to join you on this. This fits my work and current projects.

How to code latent class model w/ > 2 classes

Topic		Replies	Views
Identification of mixture of multivariate normal distributions Modeling fitting-issues	5	2645	June 11, 2018
Help writing latent class multinomial logit model (for conjoint data) Modeling fitting-issues , specification , performance	21	3734	October 16, 2020
Bayesian Hierarchical Latent Mixture Model Modeling	2	467	June 29, 2023
Factor Mixture Modelling Modeling	5	1947	May 31, 2017
Identifiability of Gaussian mixture mode Modeling	24	3188	October 29, 2017

Identification issues with Latent Class Analysis/Mixture model

Related topics