I am running a LPA with covariates using the R3step approach and ran into a problem concerning missing data.
I know that FIML cannot handle missing data on the covariates. So I imputed three of my covariates with multiple imputation in SPSS, the third one only had 2 out of 134 observations missing so I left these missing. I did not impute missing data on the latent categorical variable indicators, but used MLR estimation.
When running this analysis I see that substanively more observations have been deleted listwise than the 2 that I expected. Does this mean that MLR does not work at all in the 3 step approach an I should also impute the indicator variables?
Also, is there a good source explaining why FIML cannot handle missings on covariates? or is this 'just' a technical problem?
Relevant part of my syntax (with low, interm and ADHD imputed, and gender only has 2 missings out of 134)): Auxiliary = (R3step) gender Low Interm ADHD ; Classes = c(3); Missing = All ( -999- -10) Define: Low=0; if (SES EQ 1) then low=1; Interm=0: if (SES EQ 2) then Interm=1;
Jon Heron posted on Wednesday, August 23, 2017 - 9:20 am
I wasn't aware that r3step worked with multiply imputed datasets - are you sure that it does?
To answer a question you didn't pose, you might want to read
Enders, C. K., & Gottschall, A. C. (2011). Multiple Imputation Strategies for Multiple Group Structural Equation Models. Structural Equation Modeling: A Multidisciplinary Journal, 18, 35-54.
specifically the "cautionary note" describing why the imputation you have just done may not be a good idea. In other words, your theory says you have multiple sub-populations but you imputation model is only assuming one.
More generally, both Craig Enders and John Graham have written good stuff about FIML and missing data.
Thanks again for helping me sofar and the interesting sources!
Jon Heron posted on Thursday, August 24, 2017 - 6:18 am
i'm not sure, I was just a little surprised at the news that it might be possible + implemented. Do you get the standard r3step output? I would be inclined to go back to SPSS and derive a binary flag for people you think should be in the model and see if that agrees with Mplus. It maybe you have more missing data categories than you realised or something.
When you estimate the variance of a covariate you are changing its status from exogenous to endogenous so that it is included in the FIML estimation rather than conditioned on. At this point you are making additional distributional assumptions so it's perhaps only a good idea for continuous covariates. I think this method is attributed to Graham, although we in Bristol call it "Liam's method" for reasons I wont bore you with.
the reason for your error in that code is that you have gender in the aux command whereas it should be in the usevariable list. It's a little unusual to have missing data on gender, but it's more unusual to suspect that gender may be normally distributed which is what you are assuming there.