I have samples of less than 200 subjects. I have conducted latent class analysis using Mplus and identified three subgroups. And now I encountered several new subjects (less than 30). So is there some way to tell which subgroup these new subjects belong to? Can Mplus give out some parameters or something like an equation so I can do some calculations to assign these new subjects to one of the three groups? If this is not available, is there another method to achieve this? Any help will be much appreciated!
Hi Liqi Wang
did you add 30 new respondents to your data set? If so, why not rerun the whole LCA with those 30 people included?
Best
Marcel
Marcel Grieger Hi! Thanks for your timely help!
I worried that adding new subjects might change the structure of the latent class model and even yield different results (like you get four classes instead of the original three classes). We have conducted some analyses based on this three group model and the results make sense. So we would rather not to change the whole model.
But I will try adding these subjects and constraining the class number as three to see how it goes. Nice advice! And much thanks!
I see your dilemma. You may still try it and in case your three-class-model cannot handle 30 more people there is always the option to sweep those 30 under the rug, so to speak. However, perhaps even a different solution than the original three might make sense. Constraining the number to the desired outcome certainly is a good place to begin with.
Best of luck to you!
Marcel
Marcel Grieger Hi I think i've found the answer in the thread posted on http://www.statmodel.com/discussion/messages/13/8788.html?1602972708
It turned out the solution is quite simple: get the parameters of the original model using SVALUES command and subtitute all the * with @ . Use these syntaxes to constrian the new model with newly added subjects.
I did a little experiment by combining the old sample with these new subjects and rerun the analysis. It turned out the memberships of the old subjects remained the same and the new subjects were assigned to the three groups. I'm not exactly clear about what these meant but the result is fine so I'm OK with it (it will be great if anyone can explain it though).
Much thanks to you for the inspiring answer and Bengt O. Muthen for the original post!
Liqi
Liqi Wang Marcel Grieger The SVALUES "trick" is a neat option. However, I am somewhat skeptical about the use of fixed parameters in a small sample like yours (typically, we recommend sample sizes of > 300 for LCA, although there are situations where smaller samples will work OK).
The issue is that your parameter estimates were derived empirically (in a data-driven way) based on a fairly small sample (you mentioned "less than 200 subjects"), meaning the estimates are prone to a lot of sampling error. You then apply these "error prone" estimates to your new/additional 30 participants as if those (probably fairly imprecise) estimates were the "true" population parameters. That way, you carry over the sampling error from the original small sample to a new larger sample. I personally would instead re-estimate the model with the new complete sample (~200 + 30 subjects) as suggested by @Marcel Grieger to reduce sampling error and make use of the entire sample for the sake of more precise parameter estimation. If the solution changes when adding the 30 cases, then that is what it is. This may indicate that there was substantial sampling error in your original (< 200) cases sample, which you would want to know about.
- Badges
- Science topic
- Similar topics
- Psychology