Kevin Smith posted on Thursday, September 28, 2006 - 12:54 pm
I have a question and I could use your input. I have a multipart survery what was given at 2 different time points. One part of the survey contains 8 questions about future mentoring, but at the second time point many of the participants were not going to have any future mentoring, and thus did not answer the questions. Some participants did however. Should I consider these items missing at random? It seems like there was a systematic reason why participants skipped the section. I would like to use MI to fill in the missing data, but I don't want to violate the MAR assumption.
Any insight that you could provide would be appreciated!
You might consider this MAR if you think that the 8 questions of the first time point are predictive of the time 2 missing and other important factors are not needed. If not, adding covariates that are predictive will make the MAR approximation better.
Eric Teman posted on Saturday, February 25, 2012 - 6:21 pm
I used the logistic regression feature in Mplus to create MAR data. However, I am not so sure the data created are actually MAR. When I determined the biasedness of using listwise, the results showed no real bias. Please examine my code to see where I went wrong in generating MAR data. [x5-x8@-4.59511985]; x5 ON x1*.10 x2*.10 x3*.10 x4*.10 x9*.10 x10*.10 x11*.10 x12*.10; x6 ON x1*.10 x2*.10 x3*.10 x4*.10 x9*.10 x10*.10 x11*.10 x12*.10; x7 ON x1*.10 x2*.10 x3*.10 x4*.10 x9*.10 x10*.10 x11*.10 x12*.10; x8 ON x1*.10 x2*.10 x3*.10 x4*.10 x9*.10 x10*.10 x11*.10 x12*.10;
I assume these are Model Missing statements in a Monte Carlo run. From your notation it looks like you are predicting missing from covariates. Slopes of regressions on covariates that predict missing will not be biased (see literature on selection and Pearson-Lawley formulas), so listwise will only suffer from less precision.
If you want to see biases, you need to have missing be a function of dependent variables in the model.
Eric Teman posted on Saturday, February 25, 2012 - 7:21 pm
The model is a CFA, so all variables are DVs, right?