Corinna posted on Saturday, April 11, 2015 - 2:37 am
Dear Linda & Bengt, I want to include some predictors in my model. However, some of them were only assessed at two of the four time points. Can I still include them or is it necessary that they were assessed at all time points?
Moreover, I calculated unconditional models with linear and quadratic slopes and variances of intercept and slope fixed or freely estimated. However, I do not get a clear answer concerning the best fitting model. For example, in one model the BIC is better than in the other but entropy and the posterior class probabilty are worse. How do I decice on the best model then?
Additionally, my results indicate that e.g. in a 3-class-model the quadratic term is only significant for one of the classes. How do I then fix the quadratic slope in the other two classes to zero? (I got an ordered categorical outcome variable).
Thank you very much for your help, I really appreciate this.
I am new to both GMM and MPlus so sorry if this has either been covered or is simple.
I have longitudinal data reflecting children's scores on a test. They were tested 2 X weekly, for 8 weeks (16 scores, with missing data for sickness so random). The issue is that the children have varying weeks since starting school, however their time during the assessment is the same (eg. 0, .5, 1, 1.5...7.5). I need to control for this variability in case this is what explains the different latent class trajectories.
For a TYPE=MIXTURE model, how would I do this? I have searched widely, and have not found anything to help. I have tried this as a covariate, but should I be using the Y1-Y16 AT T1-T16? If so, how do I do the quadratic?
I'm just not sure what I should do, so any help would be greatly appreciated.
Okay, so far I've used the: i BY Y1-Y16@1; s BY Y1@0Y2@.5 .... Y16@7.5; q BY..... i ON x; !WksSinceStartSchool s ON x; q ON x;
I used a covariate as only the weeks since starting school were different for the children, whereas the schedule of assessments were all the same (ie. 0, .5, 1.....).
But reading between the lines of your statement, should I instead be using.... .... USEVAR = Y1-Y4; TSCORES = T1-T16; MISSING = ALL(9999); ... TYPE: MIXTURE RANDOM; ... %OVERALL% I s q | y1-y16 AT T1-T16; !where each student has their own unique T values. !then look at each re. predictors in a later analysis... I ON x; s ON x; q ON x;
Thanks for your help. I just want to ensure that a) we are on the same page re. my data, and b) I understand your answer. :-)
Oh an addendum to the last post, the T1-T16 values for each student do not start at time 0, eg. student x may have 51, 51.5, 52... whereas student y may have 5, 5.5, 6... etc. So the first number reflects number of weeks in school until first assessment, then each one after that is +0.5 reflecting the bi-weekly testing.
This isn't an issue is it?
Also, am I better to remove missing test values and just alter the T to reflect when tested eg. student z have T values of 5, 5.5, 6, 7... (as 6.5 time point student was sick).
Lastly, students have varying numbers of assessments, though most have 8 weeks, some have less. So I presume for the students with less than 8 weeks of assessment, I just give them incremental T values for all T1-T16, and put the MISSING 9999 value in their Y parameters for the last few that are missed?
You can use the AT approach or you can use the multiple-cohort approach (see UG), but the most important aspect is what your time axis is supposed to represent. The bi-weekly testing would seem to have a learning time axis which is quite different from the longer time axis of which week the student started. But that is a substantive question that other venues such as SEMNET is better suited for.
I have a question about reporting results. For a 1-4 class model I report information in a table including the N and % per class. I provide the N and % as they are displayed in the output under "FINAL CLASS COUNTS [...] BASED ON THE ESTIMATED MODEL" because I believe this is what a lot of the output is based on, right? (e.g.parameter estimates reported in "model results", BIC, entropy, the output that can be requested through TECH7 etc) For this reason, I believe that I should report these N and %. However, as a next step I assigned each person to their most likely latent class (i.e. "FINAL CLASS COUNTS […] MOST LIKELY LATENT CLASS MEMBERSHIP") and this makes the N and % per class change somewhat. Because I use the 3-step procedure in Mplus I adjust for the classification uncertainty, but still these "new" N and % is what I perform my subsequent analyses with. This made me a little confused as to whether I maybe should have reported these N and % reported under MOST LIKELY LATENT CLASS, instead of the ones under THE ESTIMATED MODEL. But I feel when I do that, it does not make sense to report the parameter estimates, BIC, etc as these are based on the N and % under ESTIMATED MODEL? Maybe I should provide both, and report that after model selection women were assigned to their most likely latent class and that this led to a slight shift in N and % per class? How should I deal with this? Thank you so much!