I have been running GGMM with an overall model (e.g., linear), and providing start values for the growth factors level and trend, as suggested. However, my variances are always equal between classes. I am modeling with continuous dependent variables. Should I get unique variance for each class, instead of equal variance for all calsses?
If you have more than two classes, add the other classes.
Nathan Stein posted on Wednesday, February 01, 2012 - 9:54 am
Great. Thank you. We used the i, l, and q parameters to graph a latent class trajectory. Now we would like to add two additional curves depicting within class variance, one that is 1 standard deviation above and one that is 1 standard deviation below the trajectory. How would we do this? Could you provide us with the steps that would enable us to calculate the parameter estimates for these two additional curves?
Would I simply add (or subtract) a standard deviation from each parameter estimate. For example,
If the intercept of a line is 28.71 (variance = 52.99) and the slope is -2.87 (variance = 1.20), would the line that is one standard deviation below this line have an intercept of 21.43 (28.71 - 52.99^.5) and a slope of -3.97 (-2.87 - 1.20^.5)?
You need to compute the predicted value of y at each time point. See Slide 45 in the Topic 3 course handout to see how to do this. This is where the values of your growth curve come from. Then you need to get the standard deviation at each time point and add and subtract to the predicted value at each time point. This gives the other curves. Note that the standard deviation is the square root of the variance.
LCGMs do not estimate variances and co-variances of growth factors (within-class variation assumed to be 0/is homogoenous). Does it then make sense to estimate the covariate (x) on growth factors in classes separately, as in doing so, we are assuming that x factor explains variance in growth within classes? Am I misunderstanding?
Thank you for the response, Dr. Muthen! I wonder if you can help me with the interpretation. What would be the difference in our interpretation of the model in my previous post (where the parameter of I and S on X is estimated within each class), compared to below (parameter of I and S on X held equal between groups?)