Message/Author 


Hi, i am VERY VERY new to MPlus. Is there a good "beginners" article to explain Parallel Process modeling? I'm hopeing to examine the relations between depression (using dimensional measures) and substance use (also dimensional) with a number of covariates as potential mediators and moderators. Also, I have 4 time points and an n=450. Any issues with this? 

bmuthen posted on Sunday, March 05, 2006  11:57 am



I can't think of an applied article off hand, although they probably exist. We teach this topic in our annual 5day course (Day 2)and have a handout for that day which can be requested off our web site. We also have an example in the User's Guide. Essentially, here are the steps we recommend: 1. Analyze each process separately, without covariates and then with covariates. 2. Analyze the 2 processes jointly without covariates. Remember that the correlations between the outcomes of the 2 processes are due to (a) correlations between the growth factors and also (b) more directly via correlated residuals for the 2 processes, particularly for residuals at the same time point. Point (b) is often ignored which leads to too much correlation being channeled through the growth factors, sometimes making them correlate 1. 3. Add covariates. 


See Example 6.13 in the Mplus User's Guide. This shows the input for a parallel process growth model. 


I'm trying to figure out how to plot the two processes on the same figure. When I run the imput below I get two seperate plots for process 1 and process 2. Is there anyway to combine the two processes onto one figure? Thanks, Matt TYPE = plot3; series = copeblcopef(s1)  thrtblthrtf(s2); 


There is no way to do this in Mplus. 


Hi, Should we have the same number of time points for both outcomes? Thanks! 


This is not necessary. 


Hi, I have a correlation of 1 in a multivariate growth model. I have already added correlations for residuals at the same time point between the two processes but the problem persists. Modification Indices suggest to add one lagged correlation between the residuals of the two processes (adjacent time points). I have tried that and the model is now admissible. Would that be o.k. to include this single lagged residual correlation (besides correlations at thesame time point), since I have only heard of contemporaneous correlations between residuals in multivariate LGM's until now!? Thank you Michael 


In principle, yes. But you wonder why the growth model is not able to cover more of the acrossoutcome correlations  perhaps one outcome influences another directly. 


Hello – I am running parallel process growth models that use a sampling weight and five covariates. All of these covariates include missing data to some degree. To address the comments of reviewers for our article, I would like to compare the results of a model where I bring in all possible cases by asking for the variance of the covariates to be estimated in the models, as well as run a model where the missing covariate cases are excluded but I introduce attrition weights. However, each time point would need to have its own attrition weight. My questions are: a) Is there a way to use timepoint specific attrition weights in a parallel process model? If so, how is this specified? b) Can I run what is essentially a parallel process model with the data in a ‘long’ format so that each case can have one variable that would be for the attrition weight, but it could vary across time? If so, is there an example paper you could point me to, or help me with specifying the appropriate model? Many thanks – Yvonne 


The answer to both questions is to rewrite your model as a twolevel model. Take a look at example User's Guide example 9.16 for how Mplus can help you transform the data and model properly. Once the data is transformed you can use it to write any twolevel model. Using weight and bweight command in twolevel models you can specify both weights. Note however that if you have time varying covariates with missing values due to (all) observations missing at some time points you can simply set these covariates to 0 instead of the missing values  and thus easily resolve your problem with the missing covariates. Note that those zeros will not be a part of the likelihood. 


Many thanks. One additional question to followup. The parallel process growth models I had been using specified piecewise slopes. Is there a way to specify piecewise slopes in a twolevel linear growth model as described in User’s Guide Example 9.16? 


You can do this is long format by creating two time variables that contain the time scores for the two pieces. 

Joe posted on Thursday, July 26, 2012  5:56 pm



Hi, I have two general questions about the parallel process model (ex 6.13). (1) Why is it customary, or recommended, to regress each slope on the other intercept? (2) Why can't I estimate the following model? i1 s1  y11@0 y12@1 y13@2; i2 s2  y21@0 y22@1 y23@2; y11 y12 y13 (1); y21 y22 y23 (2); i1 WITH s2@0; i2 WITH s1@0; The purpose of this is to have only the following covary. i1 WITH i2; s1 WITH s2; i1 WITH s1; i2 WITH s2; It seems intuitive to just model two growth models simultaneously, and also allow the intercepts to covary and the slopes to covary. Is the answer to (1) related to that of (2)? Thank you! 


Example 6.13 is an example of one thing you can do. The model you suggest is also fine. The model should reflect your research questions. 

Joe posted on Friday, July 27, 2012  3:33 pm



Thank you. But when I run that model, I receive the following: WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. Any suggestions? 


This is due to some mismatch between the model and your data. Please send the output and your license number to support@statmodel.com. 

Back to top 