Hi; Hoping you can give me an anwer. I have a 3-variable LCA w/sampling weights. I output the cond. probs to a file, and noticed that the weights in the output file are diminished by a small factor (.000623)for each. Is this to adjust the sample size in some way? Since it's a constant factor I don't think I have to worry about it, right? Thanks in advance,
Perhaps your weights do not sum to the sample size. If they don't, Mplus makes an adjustment. I assume that must be what you are seeing.
John C posted on Tuesday, February 21, 2012 - 11:34 am
Hello, When doing latent class analysis using a weight variable I get different class membership probability distributions in the saved file than what is reported in the out file. Further, I cannot see any common pattern in the two reported distributions. Can anyone explain why this is the case?
If you want to save and use the posterior probabilities in another analysis, multiplying them by the weights will result in the same class proportions as in the analysis.
John C posted on Wednesday, February 22, 2012 - 5:01 pm
I'm not sure I follow. Multiplying each of a set of probabilities by a constant is not going to change which one is largest and therefore the class assignment for that individual. Am I misinterpreting your suggestion?
When doing latent class analysis using a weight variable I get different class membership probability distributions in the saved file than what is reported in the out file.
The saved posterior probabilities need to be weighted.
Was this not your question?
John C posted on Thursday, February 23, 2012 - 7:59 am
No, my original question was why are the two distributions different? (see above)
Perhaps we can take a concrete example. Suppose I run a latent class analysis and save the class membership data (which you agree is different from what is reported in the out file).
Then I decide to use this class information in a different analysis. Do I just need to specify that the second analysis is executed using weighting? Or is there an intermediate step I must perform before running the second analysis?
I'm not sure, but perhaps there are two different stories here.
The way I read it now, you are asking about the most likely class designation. This is correct in the saved file for all subjects despite weigthing not being done in that file, the reason being that the weight is a person-specific matter, not class-specific.
Earlier, I thought the question was why the model-estimated class proportions in the output were not the same as what you get from the save file. There the answer is that the weighting needs to be done to the save file for agreement.