|
|
 |
| FIML or multiple imputation for LPTA? |
 
|
| Message/Author |
|
|
|
|
I am working on latent profile transition analysis with 4 continuous class indicators, measured at 3 time points. As a first step I want to estimate the measurement model (latent profile analysis) for each time point. However, there are quite some missing data, especially at the 2nd and 3rd time point. It seems to make sense to use FIML, with the same indicators measured at the earlier time points as missing data correlates for the T2 and T3 model. However, auxiliary (M) does not work with mixture models. It seems that H0 multiple imputation based on the prespecified latent profile analysis model may be an alternative solution. For the T1 model this would be fine, but for T2 and T3 it is difficult to identify the best fitting model precisely because there are missing data. The number of valid cases is much smaller at T2 and T3 due to missingness; no clear optimal number of classes emerges from the enumeration and fit statistics.
My question:
1. Is it right to use earlier measurements of the class indicators as FIML missing data correlates for missingness at later time points? Or would this be inappropriate given the subsequent latent profile transition analysis?
2. If appropriate, is there any correct way to do this in a latent profile analysis model?
3. If not appropriate, is there another solution? I feel like I may be overlooking something.
Thank you very much,
Kind regards,
Laurien |
|
|
|
|
I think you should estimate all time points jointly. That benefits from FIML and it also is a good way to assess the number of latent classes.
Note, however, that with low coverage - say less than 0.5 for most of the indicators - you are relying a lot on model assumptions relative to information from the data and should report this. |
|
|
|
|
Thank you so much for your quick reply, Dr. Muthen. I'm not sure that I understand what you mean by estimating all time points jointly. Do you mean fitting three LPA models, one for each time point, through one input file? How can it be done? I have tried to look for example syntax on how to do this, but can't seem to find any.
I will definitely keep the coverage in mind and may decide on dropping cases with too much missing data, thank you! |
|
|
|
|
| By all time points jointly, I mean doing the LPTA with all 3 times points, so analyzing your 12 variables with 3 latent class variables. |
|
|
|
|
Thank you for your reply. I understand that the LPTA should be done with all three time points, which would also solve the FIML issue. But LPTA should be preceded by LPA for each time point separately, to determine that the number of classes is the same per time point, right? That part is what my question is about.
Or is your suggestion to skip the LPA and instead estimate a series of LPTA models to see which number of classes has best model fit? Sorry for all these questions, I just want to make sure that I'm understanding you correctly! |
|
|
|
|
| Yes, if your concern about missing data is greater than your concern about different number of classes at different time points. |
|
|
|
|
| Right, that is clear. Thank you for the quick and helpful advice, Dr. Muthen! |
|
|
| Back to top |
|
|
