Message/Author |
|
mpduser1 posted on Thursday, March 28, 2013 - 7:28 am
|
|
|
Is it possible to estimate complementary log-log logistic models in Mplus? Further, do you imagine the possibility of incorporating rare-events-type framework into the construction of latent class models in Mplus (if such methods currently exist)? Thank you. |
|
|
No, Mplus does not estimate a log-log logistic model. We do have several models for count variables. See the COUNT option. |
|
mpduser1 posted on Thursday, March 28, 2013 - 12:45 pm
|
|
|
Okay. Thank you. Don't see COUNT models being the same in this context. |
|
|
What do you mean by rare events type framework? |
|
mpduser1 posted on Friday, March 29, 2013 - 1:18 pm
|
|
|
Very low probabilities of occurring; very small (but still salient and meaningful) incidence rates. |
|
|
That should work with regular latent class modeling as long as you have a sufficient sample size. I am not aware of special methodology for this, but please educate me if you are. |
|
mpduser1 posted on Tuesday, April 02, 2013 - 12:05 pm
|
|
|
I am not aware of a special methodology either; but know that your team is at the forefront of LCA software and modeling and so didn't know if it was on your radar screen. My thought was that the C L-L framework might open up some interesting possibilities for smaller sample sizes and response assumptions. I'm sure the interpretation of model parameters would be much more difficult, however. I was also interested in the C L-L more generally, which Linda indicates is not currently available. Thank you. |
|
|
Actually there is a specific methodology for rare vents. Gary King and colleagues have developed some stuff in this direction (see King & Zeng 2001, http://gking.harvard.edu/files/0s.pdf) I was wondering whether it would be possible to do it in mplus to correct LCA in which the best solution is for classes with very few cases contrasted with classes with many cases. |
|
|
This reminds me of case-control data (where you use all cases and a random sample of non-cases) and weighted logistic regression, which I think can be done in Mplus although I haven't looked into it yet (there is an old Satorra-Muthen tech note on it). But that's not in the context of LCA. |
|
|
You are right Bengt, it is used also for that. Can you give me the exact reference. I couldn't find on the website. I was also wondering whether there might be some other solution to the problem. I would use bootstrapping for instance, but in my analysis I have an LCA with covariates and I have to use Montecarlo integration. Would increase the number of initial stage starts and final stage optimizations create more robust analyses under the condition of small classes |
|
|
I'll email our case-control paper to you. To answer your last question, no but you would be more sure you got the ML solution. |
|
|
I have a dataset where 0.7% of the outcomes are 1's and the rest are 0's. In other words, very rare occurrences (46 out of 6,945). Does Bengt's 1997 paper "Robust Inference Using Weighted Least Squares..." provide an adequate reference to argue for using WLSMV estimator here? Is WLSMV in fact the appropriate estimator here? I have 3 covariates in the model with relatively normal distributions. Many thanks as always. |
|
|
With such rare outcomes the latent sample correlations that WSLMV uses are often not well estimated. I would use ML (or Bayes). |
|
Back to top |