Message/Author 


I ran a GMM with a distal outcome with three categories. The model ran well and I receive the expected outcome, including the latent class indicator model part in probability scale. Now, I have computed odds ratios. How can I tell whether my odds ratios comparing a given class probability to the reference class is significant? Where would I get the standard errors to compute a confidence interval? I'd appreciate your help greatly, as this is a question I anticipate a reviewer asking. 


I have a document that was prepared by another user that I can send to you that goes through this exercise. Email me at support@statmodel.com and I will send it to you. 

Rita posted on Monday, June 11, 2007  7:35 am



It seems that many people use LCA and there are studies including covariates and distal outcomes in the same model together with LCA. In different papers you could find various interpretations of such models (mediation, moderation, risk factors and outcomes). Basically it looks like mediation model, even though it is hard to interpret in that way. How could you interpret such model, if there is a way to name it? Let’s say you have a model where children’s temperament at age 3 (covariate) predicts parents behavior (that would be LCA) and both of them predict distal outcome – satisfaction with marriage at age 35. So, children’s temperament is a covariate and satisfaction with marriage – distal outcomes. In such model I have 3 classes or parents behavior and for each class I estimate distal outcomes. How should I interpret this model, especially if the links between predictor and latent class is significant and in the same class the mean value of predictor is significantly higher compare to other classes? It is a mediation? How should be interpreted if there is only one significant link between the covariate and distal outcomes for one class but not for others? I would be grateful for you advice. 


In a model where the covariate x influences the categorical latent variable c which then influences a distal outcome u and you are interested in the relationship between x and u, I would call this mediation. This type of model has not been written about much in the latent class literature. 

Rita posted on Monday, June 11, 2007  10:13 am



Thanks a lot! I agree that is not much about that, but some people when talk about latent class or profile analyses with predictor and distal outcomes, refer as mediation model. 


I want to use a binary distal outcome to assess the predictive validity of my LCA results (4 classes; 6 categorical indicators). I want to fix the classspecific item probabilities (i.e. values fixed at values from the 4 class model without covariates) to ensure the means are estimated based on classes I already have (re: I am referring to the strategy presented in K Nylun dissertation) My model runs but I am not sure of my syntax. Here is what I have: I wrote the thresholds for the 6 categorical variables and did nothing for the distal outcome. The distal outcome just show up in the USEVAR ARE and CATEGORICAL statements. ... USEVAR ARE cat1cat6 distal; CLASSES= c(4); CATEGORICAL = cat1cat6 distal; ANALYSIS: type=mixture; MODEL: %overall% %c#1% [cat1$1@1.811]; ... [cat6$1@1.352]; ... %c#4% [cat1$1@0.708]; ... [cat6$1@0.956]; Does this make sense? Tks 


Can you point me to the page in the Nylund dissertation where it suggests fixing the values of the thresholds. 


The information is on p.89, paragraph just below figure 3.2 


I see. This is one way of doing this. I would recommend not fixing the classes in this way. Although substantively you may think of the distal outcome differently than the latent class indicators, statistically the distal outcome is another latent class indicator. If this dramatically changes the classes, the reason for this should be investigated. 


I tried not fixing the classes and it does not substantively change the classes. Thank you! 


Hi again, I have an interpretation question. My 4 latent classes represent health status. I have age and gender as covariates and my distal outcome is mortality. In interpreting the ORs for the my distal outcome, is it correct to say that I've controlled for age and gender? For instance, the OR when comparing class 1 and class 2 is 4.5. Does it control for the age and gender differences between these classes as a multinomial regression would? It may sound like a stat 101 question but I am not sure how Mplus takes care of the covariates in this case. 


The answer is yes. It works the same way as logistic regression. 


I'm using multiple distal outcomes in a GMM, allowing the means (thresholds) of the two distal outcomes to vary across classes as follows: %c#1% [anychron$1]; [welfare$1]; When looking at the effects of class membership on the distal outcomes (in the output), do the effects of class membership on one outcome control for the effects on the other outcome (i.e. are they adjusted effects)? 


No, they are not adjusted effects. Note also that the conditional independence assumption applies to them. I would run the model with both distals and then with each distal and compare the results. 


I'm doing an LCA with a binary distal outcome and I'm not sure I understand how to interpret the output. The output excerpt below is from the 'Latent Class Odds Ratio Results' section. Am I correct in stating, based on this excerpt, that compared to Class 3, Class 1 is 2.179 times more likely to have an F1EVERDO score greater than category 1 than they are to have a score in category 2? Latent Class 1 Compared to Latent Class 3 F1EVERDO Category > 1 4.209 1.931 2.179 0.029 


The first number is the odds ratio. Compared to individuals in Class 3, individuals in Class 1 are 4.209 times more likely to have an F1EVERDO score greater than category 1 than they are to have a score in category 1. 

Lewina Lee posted on Wednesday, October 31, 2012  2:25 pm



Dear Drs. Muthen, I'm running an LPA with 2 binary distal outcomes: DEAD & CHD. In the M+ output for a 2class model, under "MODEL RESULTS" it lists the following thresholds right below the "Means" section: (For class 1: ) Thresholds CHD$1 0.919 0.105 8.790 0.000 DEAD$1 2.176 0.155 14.012 0.000 (For class 2: ) Thresholds CHD$1 0.997 0.102 9.732 0.000 DEAD$1 2.067 0.143 14.408 0.000 How should they be interpreted? Thank you for your help. Lewina 


The thresholds are the negative of logit intercepts, so that for a given class P(distal=1) = 1/(1+exp(threshold)). That gives the interpretation. 

Lewina Lee posted on Thursday, November 01, 2012  12:02 pm



Thank you! 


Dear Drs. Muthen, I ran a model with a binary distal outcome and am not sure how to interpret it. I got a mean of 2.33. Is the scale interpretable? Does this mean that C#1 has a sig lower mean that C#2 on the outcome? Thanks! 


It sounds like you have a mixture model for which a binary distal has a threshold parameter estimate for each class. The thresholds are related to the probability of the distal. The threshold differences can be tested for significance and so can the probabilities  you would use Model Constraint to do this. 


Dear Dr. Muthen, Thank you so much for your explanation  that the thresholds and thus, the probability of the distal outcome varies significantly between the two classes. Would model constraint tell me that the two classes have sig different probabilities or would it also tell me which class was higher (and the magnitude)? 


Dear Dr. Muthen, Looking at my output further, I found: EQUALITY TESTS OF MEANS ACROSS CLASSES USING POSTERIOR PROBABILITYBASED MULTIPLE IMPUTATIONS WITH 1 DEGREE(S) OF FREEDOM FOR THE OVERALL TEST CERTAINA32 (my binary outcome) Mean S.E. Class 1 0.387 0.036 Class 2 0.589 0.119 ChiSquare PValue Overall test 2.549 0.110 So, this output suggests that there really is no significant difference in means of the outcome variable across the two classes, right? Thanks, Mary 


Are you using the Auxiliary option "e" for your binary distal, that is, a 3step run, or are you doing a 1step run? Perhaps you'd better send your output to Support. 

Back to top 