This approach does not work with a continuous outcome. You might instead consider two-part modeling. See Example 6.16 in the Mplus User's Guide.
Gareth posted on Saturday, January 24, 2009 - 5:33 am
I have modified example 6.16, but is the floor variable treated as a latent class in its own right (so I should ask for classes=5), or is it present throughout, in the overall model (so I should ask for classes=4).
I anticipate four latent classes (increasing, decreasing, rapidly increasing, consistently low), in addition to the "floor" variable (no use)?
Ideally, I want to compare the four growth patterns to "no use". If this is not possible, I can use the consistently low class instead, but I would like to know how the zeros are accounted for in two-part mixture growth models.
Two-part modeling does not involve a class per se. So if you hypothesize 4 classes, then classes = c(4);
Only if you also hypothesize a class of zero (floor value) throughout would you specify a zero class for the u part; but that you have to add yourself.
Gareth posted on Monday, January 26, 2009 - 9:39 am
I hypothesize a class of zero throughout, so I have attempted to add a zero class for c#1. I am expected the other four classes to show different growth patterns, but a zero value for the first class throughout. Is this the correct syntax?
In the setup, I see: CLASSES = cg (2) c (2); KNOWNCLASS = cg (g = 0 g = 1);
There are then the indicators y1-y4.
1. Can this be used for LCA? That is, status on cg is known, and cg *may* determine the proportions of cases for each "known group" that fall under the "latent types" of c. However, the number and nature of "c" are NOT known. Is this correct?
2. If so, could we use this to examine multiple possible numbers of c? For example, change c(2) to c(3) and compare these models? Could one examine relative fit of these two models using TECH commands? (Also we'd ensure the chosen class solution is interpretable.)
3. Do the classes of "c" have the same "nature" across the 2 known class types (cg) -- referring to the same "loadings" for the indicators? But potentially just differ in proportions of cases falling into these "latent types"?
4) The manual states "The means of y1, y2, y3, and y4 vary across the classes of c, while the variances of y1, y2, y3, and y4 vary across the classes of cg." For dichotomous indicators, mean and variance are perfectly related. Can this statement be clarified for indicators that are dichotomous?
4. UG ex 7.21 is just one way to specify the model - for instance, you don't have to specify that the variances vary across the cg classes. For dichotomous indicators, you would only talk about thresholds ($). With measurement invariance across the cg groups, they vary across c classes only. If you let an indicator be non-invariant, you let it vary across the cg groups.