Student 09 posted on Friday, November 19, 2010 - 12:43 am
I am not sure how to analyze my data with the following missing values. Here is an example.
There are two variables x1 and x2.
Variable x1 was asked for both males (sex=0) and females (sex =1), has 4 response categories (1 to 4) plus a missing value category “don’t know” (coded as 99).
Variable x2 has the same 4 response categories as x1 plus a missing value category “don’t know” (coded as 99) - but in the questionnaire there was a filter such that only females were asked x2.
I want to handle the missingness due to “don’t know” in x1 and x2 by fiml, but of course don’t want to estimate missing values from the male respondents for variable x2. I was experimenting with the “pattern” command, but did not succeed.
Can anyone give me a syntax example how to handle this missingness by design?
Sounds like you don't think the correlation between x1 and x2 observed for males is relevant for females. Then, why not analyze males and females separately, where for females you only have x1?
Missing due to don't know (99) is handled as usual by ML under MAR in Mplus by default.
JOEL WONG posted on Monday, July 29, 2013 - 9:11 pm
Is there a particular percentage of missingness for which the FIML method becomes a problem?
I'm testing an SEM model with 2 latent variables (each indicated by 5 manifest variables) predicting a bunch of manifest outcome variables. My total sample size is 476 but for one of my outcome variables (SIS), we intentionally discontinued data collection for SIS after having 79 participants. Hence, my level of missingness for SIS is 83%. All my variables are continuous. I have no covariates.
Is it appropriate to argue that the missingness for SIS is not NMAR since we intentionally discontinued data collection?
Can the FIML method in Mplus handle this type and level of missingness? I tried running the model in Mplus and it seems to work but I'm not sure if I'm doing anything problematic.
No, there is not a particular percentage of missingness that causes a problem with FIML. You would need to do a Monte Carlo study to determine if it is a problem in your case. It sounds like your missingness on SIS is MCAR. The only issue is whether 17% present is enough to estimate parameters for SIS.