|
|
 |
| Covariate missingness with WLSMV esti... |
 
|
| Message/Author |
|
|
| may_k posted on Tuesday, February 13, 2018 - 8:48 am
|
|
|
Dear Drs,
I have a model with four latent factors for which the 17 indicators are ordinal, one observed outcome (continuous) and two observed covariates (dichotomous). I am using type=complex to account for the clustered nature of my data (students in schools).
Missingness is not possible for the outcome, and is low for each indicator (around 1% of cases). I understand that the WLSMV estimator handles this missingness using pairwise deletion. My problem is that one of my covariates (an SES variable obtained from a parent questionnaire) has a higher level of missingness (around 9%).
I tried to run the model using the MLR estimator so that I could bring the covariates into the model by mentioning their variances on the model command, thus retaining the cases in the analysis. However, I get the following message: "WARNING: THIS MODEL REQUIRES A LARGE AMOUNT OF MEMORY AND DISK SPACE. IT MAY TAKE A LARGE AMOUNT OF TIME TO COMPLETE." And then it doesn't complete.
Am I correct in thinking, therefore, that my remaining options are to a) lose 9% of my cases, b) use multiple imputation with WLSMV, for which there are no fit statistics or c) treat my ordinal indicators as continuous and use MLR (and therefore FIML)? If so, are all of these options defensible?
I would appreciate any advice on this and apologise if the issue has been dealt with elsewhere. |
|
|
| may_k posted on Tuesday, February 13, 2018 - 8:52 am
|
|
|
| Apologies, this was my first post and I see I have posted this in the wrong discussion topic! |
|
|
|
|
| I think I'd go with a), but if the ordinal variables don't have strong floor or ceiling effects, I could go with c). |
|
|
| may_k posted on Wednesday, February 14, 2018 - 2:00 am
|
|
|
| Many thanks. |
|
|
| Back to top |
|
|
