Message/Author |
|
|
Hi Linda, I'm trying to fit a two-part LGM to a data set with a preponderance of zeros. I am using the DATA TWOPART: syntax to model this, however the indicator for each timepoint is comprised of 4 categorical variables which I had planned to create second order factors from. Should I model these second order factors using the continuous variables from DATA TWOPART:, then estimate the two LGMs with one based on the binary variables from DATA TWOPART: and the other based on the second order factors? Or is there a way to pass the second order factors to DATA TWOPART:, VARIABLES=? Ideally I would simply use a composite variable for each time point, however as the indicators are categorical I have not been able to calculate appropriate weights for creating such variables. This is further complicated by the scale used being changed between the first and subsequent timepoints. The MODEL: section from my non-twopart model is as follows: MODEL: t1 BY t11*-t14; t2 BY t21*-t24; t3 BY t31*-t34; i s | t1@0 t2@1 t3@2; Thanks for your help, Justin |
|
|
Why do you think your categorical variables need twoo-part modeling. This is usually for count variables. |
|
|
Hi Linda, The variables reference frequency, i.e.: 1 = Never 2 = Once or twice 3 = Three to five times 4 = Six to ten times 5 = Eleven to twenty times 6 = 21 to 30 times 7 = 31 to 40 times 8 = More than 40 times with the majority of values being 1 (Never). What would be the correct approach to fitting an LGM to data such as this if not two-part? Thanks, Justin |
|
|
The variables referenced above only apply to t2-t6 of the data. t1 and t7 use a different scale (3 point). 3 point variables are available for all timepoints, i.e.: 1 = Never 2 = Once or twice 3 = Three or more times to capture the first and last timepoints. Thanks, Justin |
|
|
In your case, you can treat the variables as categorical by placing them on the CATEGORICAL list. Categorical data methodology can handle the piling up. It is when there is a preponderance of zeroes with count or continuous variables that two-part modeling is helpful. |
|
|
Thanks Linda, When I run the model with the variables specified as categorical, and the LGM fitted to the second order factors for each time point, I receive errors regarding convergence, non-positive definite, etc.. I am using the syntax: MODEL: t1 BY t1q1* t1q2 t1q3 t1q4 ; !Time 1 Factor t1@3 ; !To scale by original metric ... t7 BY t7q1* t7q2 t7q3 t7q4 ; t7@3 ; i s | t1@0 t2@1 t3@2 t4@4 t5@5 t6@6 t7@8 ; Is this correct? Thanks, Justin |
|
|
Follow Example 6.15 for multiple indicator growth. See also the example in either the Topic 3 or 4 course handout on the website. You must establish and specify measurement invariance of the factors across time before you do a growth model. |
|
|
Hi Linda, Thanks for those references, I read and listened to the Topic 3 & 4 sections, and reviewed example 6.15. Unfortunately, even with measurement invariance specified the model still won't converge. Most researchers using this data set in the past have treated the variables in question as continuous, and rarely as count. I have had much better luck with using continuous composite variables with this model when run as a two-part model, and was wondering if using DATA TWOPART: I could fit a multi-indicator model? The issue I am facing here is the requirement for both ALGORITHM=INTEGRATION and scale factors. |
|
|
Sorry, I forgot to mention that the point I'm stuck at is running the binary part of the model: is it necessary to use scale factors for the binary model? |
|
|
Two-part modeling can be used for any model. |
|
|
Thanks Linda, So it's not necessary to specify the scale factors as per 6.15? I assume the section specifying measurement invariance should stay. Thanks, Justin |
|
|
You can use only ML for two-part so scale factors are not an issue. |
|
Back to top |