Message/Author 

V X posted on Tuesday, February 06, 2007  12:01 pm



Hi, Linda, I used to use Mplus v3 for the twopart growth model and I am now using v4. I noticed that new commands have been added for the model (DATA TWOPART) in the latest version of Mplus and the original data are log transformed automatically for the continuous part of the model. I am wondering, what if I ignore the DATA TWOPART command in Mplus v4 enviroment, will the log tranformation automatically apply to the analysis? Thank you. 


No, if you don't use DATA TWOPART, the log transformation will not be done. But you can do in using DEFINE. 

V X posted on Friday, February 09, 2007  5:15 pm



Thank you,Linda. This is of great help. I have another question with regard to Monte Carlo simulation study for a twopart growth model for a continuous outcome. I plan to study the impact of alternative distribution assumption to parameter estimation. So, can I use Mplus v4 to generate a longitudinal data set with alternative assumption, say, gamma distribution for the continuous part? If not, would you give me some suggestions on how to do it? Have a nice weekend ! 


Mplus does not offer the gamma distribution. But using mixtures you can obtain almost any distribution you like. For example, using Mplus to generate data with 2 latent classes, each with a normal distribution, the mixture has a skewed distribution. We used that approach in our 2002 article which you find on our web site: Muthén, L.K. & Muthén, B.O. (2002). How to use a Monte Carlo study to decide on sample size and determine power. Structural Equation Modeling, 4, 599620. 

Nicki Bush posted on Monday, March 12, 2007  8:06 am



I am examining interactions (between 2 observed variables) within a 2part model. Are there special considerations on how to probe significant interactions in this contextI mean, given that there are significant interactions predicting both the binary and continuous parts, do I do anything other than calculate log odds for the binary outcome ones and plot 1SD below/above mean on the continuous outcome ones (ala Aiken & West)? Also, I am unaware of any papers that have examined interactions in a 2part model. Can you refer me to any writings or examples of this? Thank you so much! 


There is nothing special about the observed variable interactions in a twopart model. The issues would be the same as for regular regression and the discussion in any regression text would apply. 


Hello, I want to run a twopart model with covariates on crosssectional data (re: analyzing health care costs data with a preponderance on zeros). I already created my binary and continuous outcomes (but I know it is not necessary since V4) Is e.g 6.16 the best starting point? If yes, how do I specify the model? Or should I treat my problem as a censoredinflated regression WITH correlation between the two regression (e.g. 3.3)? If so, how can I make sure the censoring limit is zero? For the correlation, do I just add y1 with y1#1 in the model? Tks. 


The residual covariance between the two parts of a regression model is not identified so it doesn't matter if you run the two parts separately or together. You might was to see work at the Rand Corporation by Naihua Duan on twopart regression modeling. 


Thank you Linda. 

Steven Haas posted on Tuesday, March 11, 2008  4:43 pm



I have a question about the transform option in the data twopart command. The mplus guide lists a log transformation as the default. It also lists a transform=none as an option. If I do not include the transform option statement does mplus automatically log the continous part of the model? 


Yes. 


In a twopart growth approach, how would you label the growth curve of the ypart, when "type = missing" was applied? Normally, one labels this curve "frequency of behavior for those engaging in the behavior in question". However, since I'm using type=missing random "nonusers" are not excluded from the estimation of growth means of the ycurve. If a person was nonuser at t1 and then user at t2, according to FIML (type=missing), a sort of "imputation" is done with the missing yvalue at t1 (is that correct), despite the fact that this person is a nonuser!? Thanks, Michael 


Please see the following paper for a discussion of these issues: Olsen, M. K, & Schafer, J., L. (2001). A twopart random effects model for semicontinuous longitudinal data. Journal of the American Statistical Association, 96, 730745. 


Dear Linda and Bengt, I am trying to setup a twopart growth mixture model with two categorical latent variables, i.e., where c1 depicts developmental heterogeneity in the probability and where c2 depicts this for the conditional mean of y. While the model converges, it seems that my model set up is incorrect because for both c1 and c2 both measurement models (i.e., u and y) are used to define the classes. How can I keep these two part of the model separate and allowing for different number of classes for u and y. Thank you for your feedback. Below is my abreviated syntax: Model: %overall% iu su qu cu  etc. iy sy qy cy  etc. c2 on c1; Model c1: %c1#1% [ iu su qu cu]; %c1#2% [ iu su qu cu]; %c1#3% [ iu su qu cu]; Model c2: %c2#1% [ iy sy qy cy]; %c2#2% [ iy sy qy cy]; %c2#3% [ iy sy qy cy]; 


This looks correct. Try it out and if you have a problem send it to support. 


Hello! Is it possible to run a twopart longitudinal model with individually varying times of observation using Mplus? e.g., MODEL: iu su  b1 b2 b3 b4 AT a1 a2 a3 a4; iy sy  c1 c2 c3 c4 AT a1 a2 a3 a4; Thank you! 


Yes, this should be possible. 


Hello! I have done twopart growth modeling for 3 ethnic groups separately. Is it admissible to test for differences in intercepts and slopes between ethnic groups from these separate models using simple tTests or ANOVAs? I know it's theoretically possible to use Mixture Modeling with known classes to compare groups. But I don't want to go down that road unless absolutely necessary. 


Yes, this is ok because the estimates are uncorrelated. 


Dear Bengt, Thanks a lot for your answer. Would you perhaps have a reference to justify this procedure of model comparison? Philipp 


No, but it follows from first principles. You have independent samples in the different groups and you have no acrossgroup parameter equalities which means that the estimates are uncorrelated. To check that you get the right results you can run the 3 groups in one multigroups run with no equalities and use MODEL Constraint to express whatever difference you are interested in  it gives you SEs and zscores. 

Back to top 