SEM FIML and Binary outcome

[Anne Tremblay]

Hi,

I have a variable with 38% missing data (in a study with over 50 000 participants).This variable is a score on a physical test so some people don't know their score because they simply didn't pass the test. I would like to use use the Full information maximum likelihood (FIML) in SEM to deal with the missing data. However, I would like the dependent variable to be binary (above vs below a threshold) or categorical (high score, medium score and low score). Is this possible? Does the dependent variable absolutely need to be continuous? Any advice?


Thanks,
Anne

[Tiago Pereira]

Hi, Anne.

FIML can be used when we can assume that data are missing at random (MAR). Alternatively, you can also use multiple imputation. There is a variety of sources for multiple imputation. Suggestion:
https://stats.oarc.ucla.edu/stata/se...stata_pt1_new/

Check these posts as a start point (Andrew Musau is absolutely amazing with his simple and didactic examples).

https://www.statalist.org/forums/for...te-information

[Leonardo Guizzetti]

I don’t believe FIML (in Stata, this is called MLMV) is supported for outcomes other then continuous because of the multivariate normal distributional assumptions. So in that sense, you can’t strictly do what you are asking.

However, FIML and multiple imputation methods are asymptotically equivalent under MAR, so I would consider an imputation approach.