Quick question. I'm using the mice R package to impute missing data. I go by the presumption that the missing data are MAR, but I wouldn't be surprised if a few binary variables were MNAR. I followed Van Buuren's recommendations (https://stefvanbuuren.name/fimd/ch-practice.html) by including variables potentially linked to missingness in my imputation model so as to make MAR more likely. I have 20 variables in my imputation model, 10 of which have missing data, and 3 of which are the binary variables I suspect might be MNAR.
I was checking how my original data was distributed vs my imputed data. I see that for the binary variables I thought might have been MNAR, there is a higher proportion of one category in the imputations (more 1s than 0s vs original data). Van Buuren suggests on page 194 of his book that this may mean that data are MNAR or that something else went wrong in the imputation process. I suspect it may be more the former scenario.
Might be a silly question but: are these imputations still usable? The variables that I initially suspected might be MNAR seem to be accounted for in the imputations (more 1s than 0s). Can I use these data and run some tipping point analyses (for example, https://cran.r-project.org/web/packages/smdi/vignettes/d_narfcs_sensitivity_analysis.html) and leave it at that?
