|
|
 |
| Large amount of missing on predictor |
 
|
| Message/Author |
|
|
| Sun Kim posted on Friday, November 09, 2012 - 9:21 am
|
|
|
Dear Dr. Muthen,
I am running a mediational model with around 50% of missing data on one of the predictors (I have 5 predictors, 3 mediators, and 1 outcome). Is this a serious proble with FIML estimation? I could drop all those missing on this specific predictor from the whole analysis (which cuts my sample in half), but I am not sure this is the best approach.
Also, what is the "number of observations" in the output? Is it the total N of the model (which indicates that none of the data has been listwise deleted)?
Thank you.
Sun |
|
|
|
|
The N shown in the output is the total N for the analysis. Listwise deletion is not the default.
If the predictor is not that important, I would not include it in the analysis. One would hope for less than 20% missing. |
|
|
| Sun Kim posted on Friday, November 09, 2012 - 6:43 pm
|
|
|
Dear Dr. Muthen,
Thank you so much for your quick response.
The predictor is actually important (central to the research question), and I am not sure what to do in this case-- should I drop the cases missing on this predictor or still just run everything on Mplus since it does do FIML estimation.
Another question is about the indirect effects-- is "MODEL INDIRECT" giving me estimates equivalent to what I would obtain if I conducted the Sobel test of mediation (so I read in another article)?
Thank you so much for your help.
Sun |
|
|
|
|
| You would need to decide whether you want to do listwise deletion or bring all of your covariates into the model and make distributional assumptions about them. Having 50% missing is not desirable. |
|
|
| Back to top |
|
|
