Message/Author 

krishna rao posted on Wednesday, October 31, 2001  4:30 pm



I ran into the following problem while trying to model a nominal variable with 3 classes. I dicotomized each of the three classes giving me 3 binary variables. Now in regression analysis one would use only two of the variables keeping the third as a reference. Would one do the same in LCA or do we include all three. This was a topic of discussion among my collegues and I was hoping for some of your wisdom on this. Many thanks.... 

bmuthen posted on Thursday, November 01, 2001  12:32 pm



You don't want to use all 3 binary variables because then you would not have conditional independence given class, but an extra direct relationship between the 3  just like regression. 


I have an LCA model with binary, categorical and nominal variables. I am trying to decide whether the best model is a 4class or a 5class model. How do I get the different estimated probabilities plots for the different classes when I have categorical and nominal variables? How do I know to wich category the plot is referring to for a specific variable? Than 


Mplus does not give estimated probability plots for nominal variables. 

Fiona Shand posted on Monday, October 20, 2008  8:54 pm



I'm running LCA with 4 classes, 11 dichotomous variables and 5 demographic covariates. In the summary of categorical data proportions, one of the variables shows 5 categories even though it's a dichotomous variable. I've checked my data and it's all 0 or 1. Can you shed any light on what might be happening? Thanks. 


It sounds like the data are being read incorrectly. This can happen when there are blanks in the data and free format is used. Blanks are not allowed with free format data. This can also happen if the NAMES statement does not match the columns of the data set. If you cannot figure this out, please send the input, data, output, and your license number to support@statmodel.com. 


I wonder can I build a "multidimensional" latent class (latent profile) model? For example, I would like to assign a adolescent to one of three categories in terms of overall academic performance based on an instrument with three domains (math, science, and reading). Thanks! ICHUANG 


You can do that using several (in your case 3) latent class variables. It requires a higher level of LCA skills, however. You formulate a "confirmatory LCA" where you let a certain latent class variable influence only certain items. See LTA examples in the UG for ideas on how to do this  the time points correspond to your different domains. 

Andy Daniel posted on Monday, November 19, 2012  7:22 am



I'm running a (longitudinal) LCA with one nominal variable (with up to 9 categories) measured at four timepoints. The output provides "means" like: Latent Class 1 Means Occu_t1#1 5.658 Occu_t1#2 0.161 Occu_t1#3 0.026 Occu_t2#1 15.000 Occu_t2#2 11.101 Occu_t2#3 11.014 Occu_t2#4 10.509 Occu_t2#5 8.116 Occu_t2#6 11.655 Occu_t2#7 8.930 [...] I do not understand how to interpret the means. I guess they are logits, but it would be nice to have a example for the interpretation. Another possibility could be that I'm completly wrong and it isn't possible to model a LCA like this and thus also the estimates are uninterpretable at all. Thank you very much for your help!! Best regards Andy 


Yes, they are logits. You can turn them into probabilities. See the example in the user's guide on page 495496 where all covariates are zero. 

Ali posted on Monday, January 25, 2016  9:07 am



I am using LCA. There are four responses which are nominal variables. First, I ran the model without any problem. Later, I add two covariates country gender, but it gave the error messageUnknown variable(s) in an ON statement: COUNTRY. Here is my syntax: VARIABLE: NAMES ARE country SUBNATIO SCHOOLID StIDStd Student_ID_2 gender u1u4 w; USEVARIABLES ARE country gender u1u4; MISSING ARE ALL (9); CLASSES = c(3); CATEGORICAL=country gender ; NOMINAL = u1u4; WEIGHT=w; auxiliary=country gender ; ANALYSIS:TYPE = MIXTURE; Model: %overall% c#1 on country gender; c#2 on country gender; 


Remove country and gender from the CATEGORICAL list. This list is for dependent variables. 


Hello, I am conducting LCA in MPlus using 3 binary variables (LINF, MICRO, SCL) where 1=YES and 2=NO, and 5 categorical nonordered variables (BRAAK, NEUR, AVAS, ARTER, AMY). My syntax is: Variable: NAMES ARE ID SEX RACE BRAAK NEUR LINF MICRO SCL AVAS ARTER AMY DAGE; USEVAR ARE BRAAK NEUR LINF MICRO SCL AVAS ARTER AMY; IDVARIABLE = ID; CLASSES = c(3); NOMINAL ARE BRAAK NEUR AVAS ARTER AMY; CATEGORICAL ARE LINF MICRO SCL ; Savedata: FILE is LCA_cprob; SAVE=cprobabilities; Analysis: TYPE = mixture; Output: tech11 tech14; In the "RESULTS IN PROBABILITY SCALE" section of the output, I see: Latent Class 1 LINF Category 1 0.142 0.054 2.613 0.009 Category 2 0.858 0.054 15.742 0.000 [...] 1) Am I correct in interpreting this as follows: There is an 85.5% probability that people in Class 1 will have LINF="no"? 2) Why are the Pvalues from Category 1 and 2 different? 3) The output is only listing "Results in probability scale" for the binary variables. How can I see results for the nominal variables? Thanks so much! 


1) Yes. 2) Because the ztest of zero comes out differently. 3)Not provided  you would have to use the formulas at the end of Chapter 14 and use Model Constraint to express them. 


Hello I have a basic question. I have categorical and continuous variables as indicators for latent class/profile analyses. What is terminology to be used for this statistical approach as I have both continuous and categorical indicators? Thank you 


I would use latent class analysis, LCA. Actually also when all indicators are continuous  I don't think there is a need to use the term LPA. 

rgm smeets posted on Thursday, February 07, 2019  1:33 pm



Dear Mister Muthen, I ran a LCA with continuous, binary and nominal variables. I would like to know the probabilities and standard errors for the nominal variables. Is there a simple formula to compute these? 


See end of chapter 14 where the multinomial regression model is described. You would consider the case of no covariates. Then use those expressions in Model Constraint. 


Is there a reason why "Results in probability scale" are not provided for nominal latent class indicators? I am only wondering whether I ought to be cautious in how I interpret and use these results once I have calculated them manually? 


You can get these computed in Mplus using the option output:residual; 


Thanks, that's very helpful. Is there any evidence that when the latent classes are identified with very high entropy >0.95, it is sufficiently accurate to use the 'most likely latent class membership' in future analyses, and the more complex R3STEP is not needed? 


I think this is probably a good cutoff value. See Table 7 https://www.statmodel.com/download/relatinglca.pdf The table doesn't use that entropy value of 0.95 but the bias is noticeable at 0.8 and it would be safe to assume that it is minor at the 0.95 level but it does depend on various other things. You can easily run the R3step method by adding auxiliary=x(r3step) 


Thanks for the link. Yes, I guess the complexity is not the r3step method, but adding in additional things that I want to do in this analysis like multiple imputation, propensity weights, mediation etc. that I haven't learnt how to do in MPlus yet. Eleanor 

Back to top 