I am about to start an LPA analysis and have a categorical distal outcome that I would like to relate the classes to as well as other covariates that I would like use to predict class membership. I have read the article from 2014 on auxiliary variables in 3 step approaches, but it is not clear to me if it is possible to do this all together, particularly including the other covariates as predictors of class membership. Can you please clarify?
Thank you. I am familiar with BCH and had planned on using it, however I was confused by table 6 of webnote 21 that recommends BCH for continuous and DCAT for categorical. Are you suggesting that I try BCH even though my distal outcome is categorical?
I have run the LPA (with 3 classes) as mentioned above to a categorical distal outcome and in the output for one of the classes I am getting an OR of 1 with a 95 % CI of (1.0, 1.0). The other 2 classes are fine. Do you know why this may be the case?
Thank you that was very helpful. In this particular case, my last class is the one whose OR I am most interested in. Is there a way to change the order of how the classes fall out? I have read about using DEFINE but I am not sure if that could work.
You use starting values to change classes around. Request SVALUES in the output from your original run and use them in a second run where you switch them for the classes so you get the last class you want.
lisa Car posted on Thursday, November 17, 2016 - 10:45 am
Seeing as the recommended model for binary outcomes cannot accommodate covariates, I am wondering what your opinion is on using logistic regression after the fact as an option to explore them?
You can use the manual 3-step approach described in our papers.
lisa Car posted on Friday, November 18, 2016 - 10:19 am
I have tried using the manual 3-step as suggested and when I get to the final step I cannot get the model to converge. My model has 3 classes, a number of covariates and 2 binary distal outcomes. The following error message appears. THE ESTIMATED COVARIANCE MATRIX FOR THE Y VARIABLES IN CLASS 1 COULD NOT BE INVERTED. PROBLEM INVOLVING VARIABLE INCIDCFK. COMPUTATION COULD NOT BE COMPLETED IN ITERATION 3. CHANGE YOUR MODEL AND/OR STARTING VALUES. THIS MAY BE DUE TO A ZERO ESTIMATED VARIANCE, THAT IS, NO WITHIN-CLASS VARIATION FOR THE VARIABLE.
I have tried increasing STARTS up to 800 100 with no avail.
Send output to Support along with your license number.
lisa Car posted on Monday, November 21, 2016 - 1:37 pm
So the model works with a continuous distal outcome but not a categorical one however I am getting class shifting. Does something special need to be done for a categorical distal outcome in the manual 3 sep approach?
I have a 10-class LCA model with two binary distal outcomes. I'm using the Auxiliary with DCAT, but getting the following error message:
THE ESTIMATED COVARIANCE MATRIX FOR THE Y VARIABLES IN CLASS 1 COULD NOT BE INVERTED. PROBLEM INVOLVING VARIABLE NICU. COMPUTATION COULD NOT BE COMPLETED IN ITERATION 12. CHANGE YOUR MODEL AND/OR STARTING VALUES. THIS MAY BE DUE TO A ZERO ESTIMATED VARIANCE, THAT IS, NO WITHIN-CLASS VARIATION FOR THE VARIABLE. THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY. ESTIMATES CANNOT BE TRUSTED.
What does this mean? There are certainly several class indicator variables for which there is only a single value in a given latent class. Is this causing my problem and what can i do about it?
"You use starting values to change classes around. Request SVALUES in the output from your original run and use them in a second run where you switch them for the classes so you get the last class you want."
Where can I find example code to show me how to do this?
Ok, so I think I see what you are saying. I have the starting values for all the variables in each class. I just specify these in the input re-ordering the classes the way I want them? i.e. Class 8 starting values become class 1 starting values, for instance?
Also, hopefully a quick question. With the DCAT output, I'd like to get the OR or RR comparing each class to every other (pairwise comparisons), rather than just the chi-squares. Is there a way to do this?
Overall test 228.783 0.000 2 Class 1 vs. 2 19.018 0.000 1 Class 1 vs. 3 0.079 0.778 1 Class 2 vs. 3 26.816 0.000 1
The pairwise comparison and the OR are given in that output, i.e., the OR comparison for Class 1 and Class 2 has a p-value of 0.000, although the P-value is coming from the test of Prob in class 1 - Prob in class 2 = 0 rather than odds ratio.
Thank you, the chi squares are not particularly useful to me, but are you suggesting that I can create confidence intervals for each of the ORs in the upper part of the output, and compare them to each other that way? i.e., if the OR for class 2 is not inside the CI for the class 1 OR, then they are significantly different?
In my example above the T-statistics for class 1 v.s. class 2 is sqrt(19.018)=4.361.
The point estimate for the log odds ratio is log(0.487/0.513)-log(0.254/0.746)=1.025
From here you can get an estimate for its SE 1.025/4.361=0.235 and a confidence interval for the log odds ratio as [1.025-1.96*0.235, 1.025+1.96*0.235]=[0.565,1.486]
The problem of this approach is that the T-statistic is coming from a different scale for the same test - so it is up to you if you want to use it but this is what is available at this time.
The T-statistic is for the test P11-P12=0, while what you are using it for is log(P11/(1-P11)) - log(P12/(1-P12))=0.
Even though the two tests are equivalent and asymptotically they will yield the same conclusion because of the different parameterization scale they are expected to have slightly different T-statistics. If you decide to use the above approach to construct CI for the log odds ratio you have to treat them as an approximation.
I don't agree that the chi-squares are not useful. They do the same thing that the log OR confidence intervals do - they are simply in a different parameterization scale, but there is no evidence that one test is superior than the other.
Actually I just realized that you can indeed obtain the exact SE and confidence interval. It is a bit more complicated however.It goes like this Var(P11-P12)=((0.487-0.254)/4.361)^2=0.003049 Since Var(P11)=0.045^2 and Var(P12)=0.032^2 you get Cov(P11,P12)=-0.000097. You can then apply the delta method for log(P11/(1-P11)) - log(P12/(1-P12)) to obtain the exact value.
*** WARNING in MODEL command All variables are uncorrelated with all other variables within class. Check that this is what is intended. *** ERROR in MODEL command Variances for categorical outcomes can only be specified using PARAMETERIZATION=THETA with estimators WLS, WLSM, or WLSMV. Variance given for: ANYCPS3T
Change from anycps3t to [anycps37$1]. This is the threshold for a binary variable. With an ordinal variable you can have ..$2, ...$3 etc.
ksk posted on Friday, November 15, 2019 - 12:02 pm
Hello Drs. Muthen,
I'm trying to look at whether class memerbership from LPA (4 classes) would be associated with a categorical distal outcome (coded 1, 2, 3). Using DCAT in latent profile analysis, I got odds ratios shown in the EQUALITY TESTS OF MEANS/PROBABILITIES ACROSS CLASSES section. I checked FAQ "Odds ratio interpretation for categorical distal outcomes using DCAT." However, I still have questions.
1) How can I know whether the odds ratio is significant? I see 2.5% CI and 97.5% CI but not sure how to use them to figure out which odds ratio is significant.
2) Each class has odds ratios for 3 categories of a distal outcome, as the distal outcome was coded 1, 2, and 3. As far as I know, Mplus set a category with the highest value as a reference group in multinomial regression. Does it work the same in the multinomial regression using DCAT?
3) In the last class, all three categories of the distal outcomes have 1 as an odds ratio, as shown below. Did I get this because something was wrong in the syntax?