|
|
 |
| Missing Observed Covariates |
 
|
| Message/Author |
|
|
|
|
| I have a large survey data set with a fairly high proportion of missing data for some important observed covariates. The data appear to be MAR. I was reading this: http://www.ats.ucla.edu/stat/mplus/faq/fiml_counts.htm and it seems to suggest that by including the variables in brackets after the model command, you can model them and thus use FIML estimates for the missing data. The page seems to suggest, however, that these data need to be normal. So, is this method inappropriate for binary or other categorical variables? |
|
|
| Jon Heron posted on Wednesday, August 12, 2015 - 5:45 am
|
|
|
Hi Rebecca
firstly you can't use your data to tell you if you have MAR or MNAR. Are you talking about ruling out MCAR there?
The method you mention renders the covariates dependent by making them each a single indicator for a continuous latent factor. So yes, it is a little harder to justify with non-continuous data - though not impossible. |
|
|
|
|
| Many use it for binary also as a rough approximation even though binary is not close to normal Alternatively, you can use multiple imputation which explicitly handles binary. |
|
|
| Back to top |
|
|
