Message/Author 

Gary Chan posted on Wednesday, February 01, 2012  9:40 pm



Hi, I ran a latent class analysis on ten items (says, D1D2) with two covariates (X1 and X2). The followings were the estimated proportion of each class. FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS BASED ON ESTIMATED POSTERIOR PROBABILITIES Latent Classes 1 137.77276 0.09855 2 589.86515 0.42194 3 670.36208 0.47952 CLASSIFICATION QUALITY Entropy 0.821 I got very different estimates for class proportion when I ran the LCA without the covariates. The followings were the estimated proportion of each class. FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS BASED ON ESTIMATED POSTERIOR PROBABILITIES Latent Classes 1 28.24309 0.02014 2 1116.58716 0.79642 3 257.16975 0.18343 CLASSIFICATION QUALITY Entropy 0.907 Can anyone advise why this happen and how I should deal with this? It seems to me that the result from the LCA without the covariates is more correct and align with the substantive theory in my research question. Thank you very much! 


When this happens, it points to the need for direct effects between the covariates and latent class indicators. See the following paper which is available on the website for a discussion of this: Muthén, B. (2004). Latent variable analysis: Growth mixture modeling and related techniques for longitudinal data. In D. Kaplan (ed.), Handbook of quantitative methodology for the social sciences (pp. 345368). Newbury Park, CA: Sage Publications. 

Gary Chan posted on Thursday, February 02, 2012  6:46 pm



Thank you very much Prof. Muthen. I have one more question regarding covariates in LCA. For example, I now want to predict the latent class membership with 5 predictors (X1 X5). There are some missing values in some of the variables, says X1 and X2. When I ran my model, mplus excluded them from the analysis. I have already used the command Analysis: Type=mixture missing; Can you please advise how I can include cases with missing value in my model? Thank you very much! Best regards, Gary 


Missing data theory does not apply to observed exogenous covariates. If you don't want to lose the cases, you can mention the variances of your covariates in the MODEL command. Then they will be treated as dependent variables and distributional assumptions will be made about them. 


This is probably a naïve question related to my limited understanding of what happens in latent class (or latent profile) analysis. I identified a reasonable grouping of 6 continuous variables into 4 latent classes (MPLUS version 7). I was wondering if parameter estimates for classification equations are produced, and, if so, how to view them, so that probability of group membership can be estimated in other samples  like can obtain with a discriminant classification analysis. I could not find guidance in the MPLUS manual or on web site, and such may not be possible. Any guidance as to how or whether this is possible appreciated. Thanks! Michael 


You can save the parameter estimates from your 4class run using SVALUES. With a new sample  which can be just one person if you like  you can then change * to @ for the SVALUES so that you fix every parameter and use the analysis only to get posterior probabilities for the person(s) in the new sample. 


Thank you very much! That was what I was missing! 


Can latent variables be examined as predictors of latent class membership? If so, should the latent variables be defined in the %overall% command or elsewhere? When I define a latent variable in the %overall% command and then include the latent variable as a covariate (5 class model) I receive the following error message. I do not receive this message when I use the latent variable indicators as covariates (i.e., not as a latent variable). *** ERROR The following MODEL statements are ignored: * Statements in the OVERALL class: C#1 ON INTERN C#2 ON INTERN C#3 ON INTERN C#4 ON INTERN 


Please ignore my question. I was able to resolve the issue by adding algorithm = integration to my analysis syntax. Thank you! 

RuoShui posted on Wednesday, November 06, 2013  3:15 am



Hello Dr. Muthen, I want to conduct latent class analysis using students' responses to four constructs. Each construct was measured by four continuous indicators. I am wondering should I use 16 items or should I use 4 mean scores as indicators for my LCA? Thank you very much. 


Either approach is ok. With 16 items you can also include the 4 factors in the mixture model. Using 16 items you probably get more classes so if you want to keep it simple you use the 4 scores (assuming that you are sure the items follow a 4factor model). 

Alvin posted on Wednesday, November 13, 2013  10:59 pm



Hi Drs Muthen, I have a 3class LCA and I am regressing classes 1 and 2 on V1. I notice "reference to the slopes for the final class could not be done". Does Mplus treat the final class (class 3 in this case) as the reference group? Also is it possible to change the default reference group? Ideally I'd like to regress class 1 and 3 with 2 being my reference group. Many thanks 


Yes, the last class is the reference class. You can reorder the classes using starting values. Ask for the SVALUES option of the OUTPUT command and change the class numbers to what you want them to be. Use STARTS=0 when you estimate the model with starting values. 

RuoShui posted on Monday, November 25, 2013  8:26 pm



Dear Dr. Muthen, I want to run LCA using factors with indicators, but I am having problems with the syntax. Would you please let me know what is the problem with the following syntax? Thank you very much. USEVARIABLES ARE a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4 d1 d2 d3 d4; classes = c (2); Model: E1 BY a1 a2 a3 a4; E2 BY b1 b2 b3 b4; E3 BY c1 c2 c3 c4; E4 BY d1 d2 d3 d4; Analysis: Type=Mixture; Plot: Type is plot3; series is E1 (1) E2 (2) E3 (3) E4 (4); Output: standardized; 


You need %OVERALL% in the MODEL command for TYPE=MIXTURE. See examples in Chapter 7 of the user's guide. 


Greetings, I am attempting to relate classes of maltreatment (time 1) to classes of intimate partner violence (time 2) through a 3step approach. Subsequently to fitting separate latent class models, I used the posterior probabilities from the maltreatment latent variable to create dummy variables representing most likely latent class membership and modeled them as covariates to the intimate partner violence model using the 3step approach. I have a few questions: (1) How would you interpret the odds ratios (the probability of being in an intimate partner violence class (vs. an intimate partner violence reference class) as a function of being in a substance abuse class (vs. a maltreatment reference class))? (2) Are there any inherent problems with using this approach? (3) Do you know of any published studies that used this or a similar approach? Any insight you are willing to share will be appreciated. Thanks. 


*Typo in the last message: In question #1 I meant maltreatment class not substance abuse class. 


(1) You can use the logic used for transitions in LTA as described in the web note on our web site: Muthén, B. & Asparouhov, T. (2011). LTA in Mplus: Transition probabilities influenced by covariates. Paper can be downloaded from here. Mplus Web Notes: No. 13. July 27, 2011. (2) No. (3) None comes to mind. 

Back to top 