

LCA Measurement Invariance  Trend An... 

Message/Author 


Dear Professor Muthen, dear MPlusUsers, I am currently working on a Latent Class Analysis, studying types of employment in Germany between 2006 and 2014. In the end I want to do regressionanalysis for the years 2006/2011/2014 using my LC's as independent Variables. Furthermore I would like to give descriptive evidenve of the relative sizes of my LC's from 2006 to 2014. I work with PanelData (SOEP), but I do not intend to compute LatentGrowthModeling or MarkovChains, but rather do TrendAnalysis crosssectionally. Therefor I do need to assess measurement invariance for my LCAModel before pooling my data (=restraining the conditional probabilities to be equal across years, the number of classes has already been assured for each year): My results only work if I measure the same classes for each year of my analysis. Is there a way/method of assessing this measurement invariance for LCA in Mplus? Maybe there is a direct command to check for this? I refer to Finch (2015): Model Invariance in Latent Class Analysis Kankaras et al. (2010): Testing for Measurement Invariance With Latent Class Analysis I am looking forward to your responses! With kind regards, Nils Teichler from Berlin (Germany) 


I assume your samples are independent (different subjects) so you can represent the years as Knownclass in your mixture analysis. That is, you work with classes = cg(3) c(?); where cg is declared Knownclass (see UG) and c is your LCA class variable. In the Model command you can then use the dot approach %cg#1.c#1% [y1#...] (1); %cg#1.c#2% [y1#...] (2); %cg#2.c#1% [y1#...] (1); %cg#2.c#2% [y1#...] (2); etc where you put the equality constraints across the cg classes where you want them. 


Dear Professor Muthen, thank you very much for the prompt and constructive answer! My samples are actually dependent. It's a Panel study. Still I want to treat them as TrendAnalysis, not making use of the panel structure of the data. Does this then only change the use of weights or the use of clustered standard errors? With kind regards, Nils Teichler 


Alright, so you have panel data and can approach this as a singlelevel model in wide format. This means that you have an LCA model for each time point and you have a dependent variable at each time point that is influenced by the latent class variable at that time point. You should set this up as a Latent Transition model for the 3 latent class variables and their indicators, applying measurement invariance across the 3 time points. There are lots of examples of that including our User's Guide and the papers we have posted under LTA. The influence of these 3 latent class variables on the 3 DVs is such that the DV means change over the latent classes at the respective time point (that is, you don't say y on c). You may or may not want to use the new 3step approach for this. See, e.g., NylundGibson, K., Grimm, R., Quirk, M., & Furlong, M. (2014): A latent transition mixture model using the threestep specification. Structural Equation Modeling: A Multidisciplinary Journal, 21, 439454. and Asparouhov, T. & Muthén, B. (2014). Auxiliary variables in mixture modeling: Using the BCH method in Mplus to estimate a distal outcome model and an arbitrary second model. Web note 21. This will give you ideas for the Mplus scripts you want to use. 


Thank you very much! I have one additional problem. As an identifierVariable for my longdatafile I have a variable that can have up to 12 digits, containing ID and year numerically. After my analysis when creating a datafile with savedata (I want to do further analysis with Stata), it seems that Mplus changes some of my values of my IDVariable, resulting in this change of values. It seems as if Mplus rounds up my values beginning at 100000000. Is there any way around this? 99012006. 99012007. 113000000. Thank you! With kind regards from Berlin, Nils T 


Please send input, output, and data to Support along with your license number. 


Hello Dr. Muthen, I got a question on the correct syntax in order to run invariance tests for a multigrouplatentprofileanalysis. I am working with version 8.1, base program and combination add on. I have ten continous indicators to identify my subgroups. I observed a 3class model "c" fits the data best. I want to use country "co" in order to test for measurement invariance to examine differences in the latent classes for the two countries involved in our study. Below is my model to accomplish the task. I have two questions: 1.) Is the model correct to accomplish the task? KNOWNCLASS= co(co=1 co=2); classes = co(2) c(3); ANALYSIS: TYPE = MIXTURE; classes = cg(2) c(3); MODEL: %OVERALL% c on co; Output: SAMPSTAT tech11 tech14 tech8; 2) When I run the model specified as under 1) I don't identify in the output any information on measurement invariance. So, is the model syntax wrong or did I miss to specify the correct output needed? Thank you very much in advance for your support, Wassilis Kassis 


The model of "c on co" says that the class probabilities can vary across co classes. But the measurement part is the same. If you want the measurement part to be different, you can use the construction Model co: %co#1% measurement par's in class 1 %co#2% measurement part's in class 2 You have several UG examples of this. 

Back to top 

