Message/Author |
|
anonymous posted on Tuesday, April 24, 2018 - 11:32 am
|
|
|
Does anyone have examples of syntax for the two-step method for testing predictors of latent classes recently published http://stats.lse.ac.uk/kuha/Publications/Bakk_Kuha_Two_step_latent_class_analysis.pdf ? Can this be done while estimating missing data with maximum likelihood? |
|
|
Missing data is not a problem. Here is an example for you. This is the example from Section 3 in http://statmodel.com/download/webnotes/webnote15.pdf The syntax for step 1 is the same as the syntax for step 1 in the three stage procedure - Appendix D http://statmodel.com/download/AppendicesOct28.pdf The second step input file is this (much of it is automatically generated by Mplus if you add output:svalues; to the Appendix D run): variable: Names are u1-u10 y x p1-p3 n; usevar are u1-u10 y x; categorical=u1-u10; classes = c(3); data: file=man3step2.dat; Analysis: Type = Mixture; starts=0; Model: %OVERALL% Y on X; [ c#1@0.05439 ]; [ c#2@-0.03804 ]; %C#1% Y on X; [ u1$1@-0.93113 ]; [ u2$1@-0.77363 ]; [ u3$1@-0.85349 ]; [ u4$1@-0.94545 ]; [ u5$1@-0.66272 ]; [ u6$1@-0.71896 ]; [ u7$1@-0.93411 ]; [ u8$1@-0.81936 ]; [ u9$1@-0.95383 ]; [ u10$1@-0.70511 ]; %C#2% Y on X; [ u1$1@0.82326 ]; [ u2$1@1.21277 ]; [ u3$1@0.76346 ]; [ u4$1@0.84224 ]; [ u5$1@1.17576 ]; [ u6$1@0.89084 ]; [ u7$1@0.87685 ]; [ u8$1@0.93901 ]; [ u9$1@0.99801 ]; [ u10$1@0.90357 ]; %C#3% Y on X; [ u1$1@1.28491 ]; [ u2$1@1.11949 ]; [ u3$1@1.12317 ]; [ u4$1@1.44773 ]; [ u5$1@0.83887 ]; [ u6$1@-1.15671 ]; [ u7$1@-1.27923 ]; [ u8$1@-0.91518 ]; [ u9$1@-1.13319 ]; [ u10$1@-1.01133 ]; |
|
|
I would like to use the new two-step procedure suggested by Bakk & Kuha (2018, Psychometrika) for an LTA. For this purpose I have estimated two separate LCA’s (input and output files attached). However, when I feed the starting values from these (using the output from “svalues”) into a LTA, the classes change. To clarify: in this second step I am only interested in the structural parameter linking the two LCA’s (c1 and c2). I wonder if perhaps you could you provide some help with the correct syntax for the second step? Thank you very much. |
|
|
Does this article deal with LTA or are you trying to generalize it from LCA to LTA? |
|
|
The example in the article primarily deals with LCA. However, they note (on p. 877) that the method easily extends to more complicated cases. They write: "For instance, suppose that there are two latent class variables X1 and X2 with separate sets of indicators Y1 and Y2, and the structural model is of the form p(X1)p(Z1|X1)p(X2|Z1, X1)p(Z2|X1, Z1, X2). In step 1, we would then estimate two separate latent class models, one for X1 and one for X2 (and both again without Z = (Z1, Z2)). Step 1 parameters θ1 would be the measurement probabilities of X1 and X2 and the parameters of p(X1), and step 2 parameters would be those of the rest of the structural model apart from p(X1)." |
|
|
(in the version of the paper linked to at the top of this discussion the quoted text is on p. 10) |
|
|
Yes, looks like LTA might be covered. I haven't tried out this approach but send your output from the first and second step to Support along with your license number and we'll try to help you. |
|
Sara Suzuki posted on Sunday, December 29, 2019 - 9:10 pm
|
|
|
Is the Bakk & Kuha method better than DCAT even for categorical distals? |
|
|
We haven't seen any studies that indicate this. The DCAT approach appears to work quite well, see Table 8 http://www.statmodel.com/download/3stepOct28.pdf In most situations, I would expect the DAT method, the two-step method, the three-step method and the BCH method to yield similar results. If you have an example where that is not the case send it to support@statmodel.com |
|
|
Thank you for providing the code for the two-step method. Is it possible to also include the SE correction (pg.12 in the paper linked up top) using mplus syntax? |
|
|
Not easily. All the matrices in formula (5) can be obtained with output:tech3, but this still looks a bit hard to do manually. We have documented one simulation study for that method, see Table 5 https://www.statmodel.com/examples/webnotes/webnote21.pdf and the bias overwhelms the need for SE adjustment. |
|
Back to top |