genwei posted on Wednesday, September 02, 2015 - 8:49 pm
I am doing latent class modeling with IRT. The "CLASSES = essay(6);" was used. A IRT model (like graded response model) was used to link the essay variable with latent trait "factor". But, I found the essay is nominal and latent. The "factor BY essay;" cannot work. I know that it could use "essay ON factor;". However, Mplus adopted the last category as reference category and doing multinomial regression model. In my model, I have to adopt the first category as reference category to count from 0 to 5 for ordinal intensity. Any workaround to make "factor BY essay;" work? Thanks.
You can change the reference class by using starting values and STARTS-0; to order the classes the way you want.
Dena Pastor posted on Wednesday, March 01, 2017 - 4:20 pm
I have dichotomous indicators and am interested in 2 different kinds of 3-class models: 1) classes are allowed to be unordered (no constraints) and 2) classes are ordered (such that all thresholds for class 1 are larger than class 2; all thresholds for class 2 are larger for class 3). I was able to run the first 3-class model and obtained a stable solution. As well, the thresholds were ordered in a way that aligned with the second 3-class model. I used the parameter estimates from the first model as starting values for the ordered 3-class model and used the MODEL CONSTRAINT command to constrain the thresholds in the manner outlined above. I am having trouble with estimation and obtained the following error message:
2000 perturbed starting value run(s) did not converge. Final stage loglikelihood values at local maxima, seeds, and initial stage start numbers:
-89160.500 unperturbed 0
99 perturbed starting value run(s) did not converge.
The LL here is the same as the LL from the unordered 3-class model. Any advice for obtaining results? My next step is to add more restrictive model constraints where the thresholds for the 3 classes fall in between certain values (e.g., t1>-3>t2>-1>t3), so any advice on facilitating the estimation process with ordinal LCA (with constraints) would be helpful. Thanks!