linda beck posted on Monday, October 20, 2008 - 11:03 am
How can one test if a linear model fits better as an intercept-only model? I specified this in one class of two of a two-part LGMM. This was a normative class with a relatively flat growth in the quantity-part (y-part), so i set [sy@0] and sy@0 in this class. Do I use the chi-square diff test or BIC to compare with a model applying a linear setting in the y-part? I want to compare on the basis of entire two-part models not only the y-part. The u-part has the same growth factors in both models.
One approach would be to estimate a linear model and look at the significance of the mean of the slope growth factor.
linda beck posted on Tuesday, October 21, 2008 - 11:40 am
yes, this was a silly question ;-), sorry. It's like the quadratic which is not needed when its mean is insignificant.
That leads me to a final question relating this issue. One commonly models LGMM with the growth factor setting derived from LGM. I also did (quadratic). But after that I found relatively flat curves in some subgroups, better captured by linear growth. Once one has established the final number of classes with the quadratic setting, can I restrict to linear growth in some groups AFTERWARDS and rely on that nothing changes with regard to the correct number of classes and central parameter estimates? What is changed and critical when doing so!?
If my approach isn't appropriate, how can one handle different growth shapes across groups in LGMM? I find it very hard to do that withing the class finding process, because sometimes the ordering of groups is changed and one doesn't know in which group to put the model specific commands.
I think it is often just as well to keep the quadratic in all classes even if some classes turn out close to linear. Unless you have a good theory for why a class should be linear. If you want a class to be linear, you should give a sufficient amount of good starting values for that class.
linda beck posted on Wednesday, October 22, 2008 - 6:28 am
I have a class with a flat increase and it had insignificant linear and quadratic growth means in an unconditional model. I restricted then within this class to a linear growth function and found a sig. mean for the single linear slope, indicating growth. So I just want to show, that there is sig. growth within this class for pedagogical reasons (or reviewers :-)).
My (sig.) effects on the linear slope in a conditional model did not differ between growth factor settings, as expected. When you say from your experience, that class finding process is not influenced by quadratic or linear setting, I would say my linear setting in one class is something for a footnote. Since I can imagine, that implementing an extra linear class is more challenging when moving to 3 classes (I have two classes so far), which were rejected so far in my LGMMs with a quadratic setting.