

MuthenRoy PMM: SDs/SEs for Timepoint... 

Message/Author 


I have conducted a series of longitudinal analyses whereby I am assuming MAR as a first step and NMAR as a second. There are 6 timepoints. For the first step, I am using the default estimator in a multigroup latent growth model. For the second second, I'm implementing the MuthenRoy PMM method. I now want to assess whether the difference I am seeing between the groups at various timepoints are significant (primarily, the final timepoint). For the LGM using FIML, I am succesfully doing this by making the final timepoint the intercept, and earlier timepoints fixed @ minus occasions. However, this is not successful within the PMM approach. Doing this within the PMM approach changes the latent class allocations and the trajectories of change. Is there a way to get the SDs or SEs for the means at each timepoint? 


You can use model parameter labels in Model Constraint to express the mean at any time point and in any class. This will give you the estimate and its SE. But I would first explore why you get different solutions with different time scores. Look for the best loglikelihood when using many random starts. 


Bengt, thank you for your response. I've checked the users guide and explored the internet for examples of the syntax I would use to use the Model Constraint to get the estimate and SE. However, I am still unsure what the exact syntax would be. My model is this: i s q w1_cesd@0 w2_cesd@2 w3_cesd@4 w4_cesd@6 w5_cesd@8 w6_cesd@10; s on i; How do I express, in the model constraint, the Mplus syntax to get the SE for the estimates at each timepoint, individually? 


You give labels to the i mean, the s intercept, and the slope: [i] (im); [s] (sint); i on s (slope); and then use those labels to express means in Model Constraint: New(sm); sm = sint+ slope*sint; ! the s mean and further lines using the Topic 3 handout, slide 98 to express the means of the outcome with your time scores at the different time points. 


Wonderful  this is extremely helpful and with a listwise deletion model, FIML, and LOCF I have my estimates. However, for my PMM (basic PMM, no LCs) the produced means are worrying. They track the means and SEs from listwise deletion (where I have lost over 3,000 cases). They are also quite different to the modelestimated and sample means shown in the figures. This leaves me wondering whether the dropout indicators need to be somehow built in to the model constraints, to ensure that the means produced are taking those in to account. Currently, this is my syntax for the PMM: i s  w1_cesd@0 w2_cesd@2 w3_cesd@4 w4_cesd@6 w5_cesd@8 w6_cesd@10; [i] (im); [s] (sint); i on s (slope); i on d1d5; s on d3d5; s on d1(1); s on d2(1); Model Constraint: New (sm w1m w2m w3m w4m w5m w6m); sm = sint+slope*sint; w1m = im+sm*(0); w2m = im+sm*(2); w3m = im+sm*(4); etc Should the im estimate actually be based around the regression of the intercept on dropout, and likewise for the sint? 


when you have i on d1d5; s on d3d5; s on d1(1); s on d2(1); you have to modify sm = sint+slope*sint; Just think how you would do that for regular regression. 

Back to top 

