Message/Author 

Mark Prince posted on Thursday, April 18, 2013  8:54 am



Hi, In a growth mixture model tech 7 provides means and covariances for the variables in the model separated by each latent class. Is there any way to request or even calculate by hand the standard errors or variances of the means for each latent class? thank you, Mark 


You would have to do this using MODEL CONSTRAINT to express the means in terms of model parameters that are labeled. 

Mark Prince posted on Thursday, April 18, 2013  1:08 pm



Thank you for the reply, but I am unsure how to use the model constraint command in the way you are suggesting. Here is my code: USEVARIABLES ARE x1 x12; CLASSES = c (4); ANALYSIS: TYPE = MIXTURE; STARTS = 50 10; MODEL: %OVERALL% i s q  x1@0 x2@1 x3@2 x4@3 x5@4 x6@5 x7@6 x8@7 x9@8 x10@9 x11@10 x12@11; OUTPUT: TECH7; How would I use the model constraint command to get the standard errors for x1x12 within each of the 4 classes? Can you give me an example of the code I would add? Thank you. 


Take a look at our short course teaching of Topic 3, particularly slide 98 showing how to compute estimated means at different time points. Drawing on that: add in the MODEL command [iq] (p1p3); and using the example of the outcome mean at time point 3: MODEL CONSTRAINT: New (m3); m3 = p1+p2*2+p3*9; that is, multiply by the time scores and their squares. This gives you the mean estimate of m3 and its SE under New parameters. 

Mark Prince posted on Saturday, April 20, 2013  10:36 am



Hi  one last time on this question. I appreciate the help thus far, and I was able to use your example to get the means and standard errors for each time point; however, I wasn't able to get the means and SEs for each time point within each class. The code you provided only gave me estimates for the overall model. I tried using the %c#1% statement but then I didn't know exactly what to call out, because if I put m1m12 it was an unbalanced equation, and if I repeated the equations I used in the overall model, my model was over specified. Can you help me understand the final steps I need to take to get the separate estimates for means and SEs at each time point within each latent class? Thank you for all the continued help. 


You have to do the labeling in each class and compute this for each class. Don't do it in the Overall. So for class 1, 2, etc: %c#1% [iq] (p11p13); %c#2% [iq] (p21p23); etc And in Model Constraint you define all your means, for each class. 


Thank you for the helpful discussion so far. I have been able to estimate the means and SEs for each time point within each class using the syntax you provided; however, I am finding that the means calculated using the model constraint procedure are different from the TECH7 estimates  with some mean estimates extending beyond the possible range for the variables included. Ultimately, I would like the standard errors of the TECH7 estimates  so this might be a standard error weighted by the posterior probabilities. Can you shed some light on the discrepancies I am finding and/or lead me in the right direction to get the SEs I am looking for? Thank you. 


TECH7 does not give modelestimated means. Instead, it is an attempt at creating "sample statistics" with mixtures. As an aide, this is often used in statistical articles, but is not a perfect sample statistic for means in mixture modeling since we don't observed the latent classes. But it is as close as you get; it is not perfect since you use modelestimated posterior probabilities. If the Model constraint approach doesn't seem to give the right values, please send to support. 

Back to top 