Message/Author |
|
Jan Ivanouw posted on Sunday, October 15, 2017 - 6:39 am
|
|
|
Hi Is there a way in which I in the same model can combine some measurement models for latent variables, and also use these latent variables as input in a latent class analysis (Latent Profile Analysis)? |
|
Jan Ivanouw posted on Tuesday, October 24, 2017 - 10:51 am
|
|
|
In order to explain my question better: I can use observed indicators (items) as basis for a latent class analysis. I can instead use a measurement scale (like a sum score) as basis for a latent profile analysis. However, there are instances in which there are very many items making latent class analysis cumbersome, and maybe not practicable without a really large sample. I could use a two step procedure, first creating factor scores from a measurement model, and then use these scores in a profile analysis, but since factor scores are only approximations to the latent variables, I wonder if it is possible to combine the measurement model with a latent class analysis, and so skip the estimation and use of factor scores as intermediates. |
|
|
Yes, this is possible. Your Model statements should make sure that you have measurement invariance over the classes (using intercept/threshold equality labels) whereas the factor means (and perhaps variances) are allowed to vary over the classes. |
|
Jan Ivanouw posted on Tuesday, October 24, 2017 - 3:29 pm
|
|
|
Thank you very much. Is there an example on how to make the input file? |
|
|
Not that I can find right now. But just go ahead and try it - first specify the Overall model as a factor model. You know that the defaults with mixtures let the means and intercepts vary freely across classes while loadings and variances don't, so you just have to restrict the measurement intercepts to be equal across classes using the usual mixture approach: %c#1% [...] (1); .. %c#2% [...] (1); |
|
|
Dr Muthen, I am trying to specify a LCA model from latent variables. Your recommendation above indicates that the measurement intercepts should be equal across classes. I was wondering if the input below is the best way for me to specify that within a LCA model with covariates from latent variables? (I examined 3 classes for now). Thanks so much, Kalina CLASSES = C (3); CATEGORICAL = ST48Q01 - ST48Q03; Weight = W_FSTUWT; ANALYSIS: TYPE=MIXTURE; ALGORITHM=INTEGRATION; STARTS = 100 10; MODEL: %OVERALL% Attitude BY ST89Q24* ST29Q08 ST29Q02; Attitude@1; SubjNorm BY ST35Q46* ST79Q115 ST79Q517; SubjNorm@1; Control BY ST43Q15* ST43Q36 ST37Q14; Control@1; Intention BY ST48Q01* ST48Q02 ST48Q03; Intention@1; Behavior BY ST46Q12* ST46Q35 ST46Q67; Behavior@1; C#1 ON GENDER RACETHC ESCS; C#2 ON GENDER RACETHC ESCS; Attitude (1) SubjNorm (1) Control (1) Intention (1) Behavior (1); |
|
|
You hold measurement intercepts equal across classes by using statements like [ST....] (list of labels like p1-p5); in each class. |
|
|
Hi there, Just to add to this post. I'm having trouble working out the code with my analysis. I want to estimate a LCA using 1 latent variable (factor of mralmon9 mraltue9 mralwed9 mralthu9 mralfri9 mralsat9 mralsun9) and 2 manifest variables. I've looked at the output, user guide and various TECH's but i'm not sure if this is right. Any information on this would be very welcomed. Many thanks, Emily Code: AUXILIARY = unqid; USEVARIABLES ARE mralmon9 mraltue9 mralwed9 mralthu9 mralfri9 mralsat9 mralsun9 mpalmon9 mdruyn9; COUNT ARE mralmon9 mraltue9 mralwed9 mralthu9 mralfri9 mralsat9 mralsun9 (nbi); CATEGORICAL ARE mdruyn9; NOMINAL ARE mpalmon9; MISSING ARE ALL(-99); CLASSES = c(2); ANALYSIS: ESTIMATOR = MLR; TYPE = MIXTURE; STARTS = 20 4; STITERATIONS = 10; MODEL: %OVERALL% mweekal9 BY mralmon9 mraltue9 mralwed9 mralthu9 mralfri9 mralsat9 mralsun9; %c#1% [mweekal9*1]; |
|
|
We need to see your full output - send your output to Support along with your license number. |
|
Back to top |